Ugo Pagallo · Monica Palmirani

Pompeu Casanovas · Giovanni Sartor

Serena Villata (Eds.)

AI Approaches
LNAI 10791

to the Complexity
of Legal Systems
AICOL International Workshops 2015–2017:
AICOL-VI@JURIX 2015, AICOL-VII@EKAW 2016,
AICOL-VIII@JURIX 2016, AICOL-IX@ICAIL 2017,
and AICOL-X@JURIX 2017, Revised Selected Papers

123
Lecture Notes in Artificial Intelligence 10791

Subseries of Lecture Notes in Computer Science

LNAI Series Editors

Randy Goebel
University of Alberta, Edmonton, Canada
Yuzuru Tanaka
Hokkaido University, Sapporo, Japan
Wolfgang Wahlster
DFKI and Saarland University, Saarbrücken, Germany

LNAI Founding Series Editor

Joerg Siekmann
DFKI and Saarland University, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/1244
Ugo Pagallo Monica Palmirani
•

Pompeu Casanovas Giovanni Sartor

•

Serena Villata (Eds.)

AI Approaches
to the Complexity
of Legal Systems
AICOL International Workshops 2015–2017:
AICOL-VI@JURIX 2015, AICOL-VII@EKAW 2016,
AICOL-VIII@JURIX 2016, AICOL-IX@ICAIL 2017,
and AICOL-X@JURIX 2017
Revised Selected Papers

123
Editors
Ugo Pagallo Giovanni Sartor
University of Turin University of Bologna
Turin, Italy Bologna, Italy
Monica Palmirani Serena Villata
University of Bologna Inria - Sophia Antipolis-Méditerranée
Bologna, Italy Sophia Antipolis, France
Pompeu Casanovas
La Trobe University
Melbourne, VIC, Australia

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Artificial Intelligence
ISBN 978-3-030-00177-3 ISBN 978-3-030-00178-0 (eBook)
https://doi.org/10.1007/978-3-030-00178-0

Library of Congress Control Number: 2018953366

LNCS Sublibrary: SL7 – Artificial Intelligence

© Springer Nature Switzerland AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

AICOL stands for Artificial Intelligence Approaches to the Complexity of Legal

Systems. This volume presents the revised selected papers of the five different AICOL
Workshops during 2015–2017. The first took place as part of the JURIX 2015 con-
ference at the Universidade do Minho in Braga, Portugal, on December 9, 2015. The
second took place in Bologna, Italy, on November 19, 2016, in conjunction with the
20th International Conference on Knowledge Engineering and Knowledge Manage-
ment (EKAW 2016). The third was held at the JURIX 2016 conference at the Inria
Sophia Antipolis Mediterranée in Sophia Antipolis, France, on December 14, 2016, as
an AICOL Workshop. The fourth was the Workshop on MIning and REasoning with
Legal texts, held on June 16th, 2017, in London (UK) which was connected with the
MIREL (MIning and REasoning with Legal texts) project, H2020 Marie
Skłodowska-Curie grant agreement No 690974 (http://www.mirelproject.eu). It was
organized in conjunction with ICAIL 2017, the 16th International Conference on
Artificial Intelligence and Law. The fifth was held on December 13, 2017, at the
JURIX 2017 conference in Luxembourg. All these workshops had the common rational
to combine different disciplines for managing legal systems complexity, both in the
foundations and in the applications aspects. Two workshops in particular investigated
the Semantic Web techniques (EKAW 2016) and Natural Processing and Legal Rea-
soning models (MIREL 2016). The 37 selected papers, including the introduction,
represent a comprehensive picture of the state of the art in legal informatics.
The present volume follows the previous AICOL volumes: AICOL I-II, published in
2010, include papers from the first AICOL conference in Beijing (24th IVR Congress
in September 15–20, 2009, China) and the follow-up in Rotterdam (JURIX-09,
November 16–18, The Netherlands); AICOL III, published in 2012, resulting from the
third AICOL conference, held in Frankfurt a. M. (25th IVR, August 15–20, 2011,
Germany); AICOL-IV@IVR in Belo Horizonte Brazil, July 21–27, 2013, and
AICOL-V@SINTELNET-JURIX in Bologna, Italy, December 11, 2013.
Like its predecessors, this volume embodies the philosophy of the AICOL con-
ferences, which is to provide a meeting point for various researchers, such as legal
theorists, political scientists, linguists, logicians, and computational and cognitive
scientists, eager to discuss and share their findings and proposals. In this sense, the
keywords “complexity” and “complex systems” sum up the perspective chosen to
describe recent developments in AI and law, legal theory, argumentation, the Semantic
Web, and multi-agent systems.
AICOL incorporates in its VI edition the perspective of social intelligence, the
intertwined human–machine perspective on cognition, agency, and institutions. This
promising approach brings together the analytical and empirical perspectives of soci-
ety. Stemming from this starting point, the volume is divided into six main sections:
(i) Legal Philosophy, Conceptual Analysis, and Epistemic Approaches; (ii) Rules and
Norms Analysis and Representation; (iii) Rules and Norms Analysis and
VI Preface

Representation; (iv) Legal Ontologies and Semantic Annotation; (v) Legal Argumen-
tation; and (vi) Courts, Adjudication and Dispute Resolution. New entry topics need a
particular spotlight like Legal Design or Legal Data Analytics.
Finally, a special thanks is due to the excellent Program Committee for their hard
work in reviewing the submitted papers. Their criticism and very useful comments and
suggestions were instrumental in achieving a high quality of publication. We also thank
the authors for submitting good papers, responding to the reviewers’ comments, and
abiding by our production schedule.

July 2018 Ugo Pagallo

Monica Palmirani
Pompeu Casanovas
Giovanni Sartor
Serena Villata
 Organization

Organizing Committee
Danièle Bourcier Université de Paris II, France
Pompeu Casanovas Autonomous University of Barcelona, La Trobe
University
Monica Palmirani University of Bologna, Italy
Ugo Pagallo University of Turin, Italy
Giovanni Sartor European University Institute and University of Bologna,
Italy
Serena Villata Inria Sophia Antipolis, France

Program Committee
Laura Alonso Alemany Universidad Nacional de Córdoba
Michał Araszkiewicz Jagiellonian University
Guido Boella University of Torino
Daniele Bourcier Centre d’Etudes et de Recherches de Science
Administrative et Politique, Universite de Paris II
Pompeu Casanovas Autonomous University of Barcelona, La Trobe
University
Marcello Ceci GRCTC - Governance, Risk and Compliance Technology
Center
Pilar Dellunde Autonomous University of Barcelona
Luigi Di Caro University of Torino
Angelo Di Iorio University of Bologna
Enrico Francesconi ITTIG-CNR
Michael Genesereth Stanford University
Jorge Gonzalez-Conejero UAB Institute of Law and Technology
Guido Governatori Data61, CSIRO
Davide Grossi University of Groningen
John Hall Model Systems
Renato Iannella Queensland Health
Beishui Liao Zhejiang University
Arno R. Lodder Vrije Universiteit Amsterdam
Marco Manna University of Calabria
Martin Moguillansky Universidad Nacional del Sur
Paulo Novais University of Minho
Ugo Pagallo University of Torino
Marco Pagliani Senate of Italian Republic
Monica Palmirani University of Bologna
Adrian Paschke Freie Universität Berlin
VIII Organization

Silvio Peroni University of Bologna

Ginevra Peruginelli ITTIG-CNR
Enric Plaza IIIA-CSIC
Marta Poblet Autonomous University of Barcelona, RMIT University
Francesco Poggi University of Bologna
Martín Rezk Rakuten Inc.
Livio Robaldo University of Torino
Víctor Rodríguez Doncel Universidad Politécnica de Madrid
Antoni Roig Autonomous University of Barcelona
Piercarlo Rossi Università del Piemonte Orientale
Antonino Rotolo University of Bologna
Giovanni Sartor EUI, University of Bologna
Erich Schweighofer University of Vienna
Barry Smith SUNY Buffalo
Clara Smith Universidad Nacional de La Plata, Universidad Católica
de la Plata
Davide Sottara Arizona State University
Said Tabet EMC Corporation
Raimo Tuomela University of Helsinki
Leon van der Torre University of Luxembourg
Tom Van Engers University of Amsterdam
Anton Vedder Tilburg University
Serena Villata Inria Sophia Antipolis, Sophia Antipolis Cedex
Fabio Vitali University of Bologna
Radboud Winkels University of Amsterdam
Adam Wyner Swansea University
John Zeleznikow Victoria University
 Contents

Introduction: Legal and Ethical Dimensions of AI, NorMAS,

and the Web of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Ugo Pagallo, Monica Palmirani, Pompeu Casanovas,
Giovanni Sartor, and Serena Villata

Legal Philosophy, Conceptual Analysis, and Epistemic Approaches

RoboPrivacy and the Law as “Meta-Technology” . . . . . . . . . . . . . . . . . . . . 23

Ugo Pagallo

Revisiting Constitutive Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Giovanni Sileno, Alexander Boer, and Tom van Engers

The Truth in Law and Its Explication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Hajime Yoshino

From Words to Images Through Legal Visualization . . . . . . . . . . . . . . . . . . 72

Arianna Rossi and Monica Palmirani

Rules and Norms Analysis and Representation

A Petri Net-Based Notation for Normative Modeling: Evaluation

on Deontic Paradoxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Giovanni Sileno, Alexander Boer, and Tom van Engers

Legal Patterns for Different Constitutive Rules . . . . . . . . . . . . . . . . . . . . . . 105

Marcello Ceci, Tom Butler, Leona O’Brien, and Firas Al Khalil

An Architecture for Establishing Legal Semantic Workflows

in the Context of Integrated Law Enforcement . . . . . . . . . . . . . . . . . . . . . . 124
Markus Stumptner, Wolfgang Mayer, Georg Grossmann, Jixue Liu,
Wenhao Li, Pompeu Casanovas, Louis De Koker, Danuta Mendelson,
David Watts, and Bridget Bainbridge

Contributions to Modeling Patent Claims When Representing

Patent Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Simone R. N. Reis, Andre Reis, Jordi Carrabina,
and Pompeu Casanovas

Modeling, Execution and Analysis of Formalized Legal Norms

in Model Based Decision Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Bernhard Waltl, Thomas Reschenhofer, and Florian Matthes
X Contents

Causal Models of Legal Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

Ruta Liepina, Giovanni Sartor, and Adam Wyner

Developing Rule-Based Expert System for People with Disabilities –

The Case of Succession Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Michał Araszkiewicz and Maciej Kłodawski

Legal Vocabularies and Natural Language Processing

EuroVoc-Based Summarization of European Case Law . . . . . . . . . . . . . . . . 205

Florian Schmedding, Peter Klügl, David Baehrens, Christian Simon,
Kai Simon, and Katrin Tomanek

Towards Aligning legivoc Legal Vocabularies by Crowdsourcing . . . . . . . . . 220

Hughes-Jehan Vibert, Benoit Pin, and Pierre Jouvelot

Data Protection in Elderly Health Care Platforms . . . . . . . . . . . . . . . . . . . . 233

Angelo Costa, Aliaksandra Yelshyna, Teresa C. Moreira,
Francisco C. P. Andrade, Vicente Julian, and Paulo Novais

Assigning Creative Commons Licenses to Research Metadata:

Issues and Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Marta Poblet, Amir Aryani, Paolo Manghi, Kathryn Unsworth,
Jingbo Wang, Brigitte Hausstein, Sunje Dallmeier-Tiessen,
Claus-Peter Klas, Pompeu Casanovas, and Victor Rodriguez-Doncel

Dataset Alignment and Lexicalization to Support Multilingual Analysis

of Legal Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Armando Stellato, Manuel Fiorelli, Andrea Turbati, Tiziano Lorenzetti,
Peter Schmitz, Enrico Francesconi, Najeh Hajlaoui,
and Brahim Batouche

A Multilingual Access Module to Legal Texts . . . . . . . . . . . . . . . . . . . . . . 272

Kiril Simov, Petya Osenova, Iliana Simova, Hristo Konstantinov,
and Tenyo Tyankov

Combining Natural Language Processing Approaches for Rule Extraction

from Legal Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Mauro Dragoni, Serena Villata, Williams Rizzi, and Guido Governatori

Analysis of Legal References in an Emergency Legislative Setting . . . . . . . . 301

Monica Palmirani, Ilaria Bianchi, Luca Cervone,
and Francesco Draicchio

Legal Ontologies and Semantic Annotation

Using Legal Ontologies with Rules for Legal Textual Entailment . . . . . . . . . 317
Biralatei Fawei, Adam Wyner, Jeff Z. Pan, and Martin Kollingbaum
 Contents XI

KR4IPLaw Judgment Miner - Case-Law Mining for Legal

Norm Annotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
Shashishekar Ramakrishna, Łukasz Górski, and Adrian Paschke

Towards Annotation of Legal Documents with Ontology Concepts . . . . . . . . 337

Kolawole John Adebayo, Luigi Di Caro, and Guido Boella

Reuse and Reengineering of Non-ontological Resources

in the Legal Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Cristiana Santos, Pompeu Casanovas, Víctor Rodríguez-Doncel,
and Leendert van der Torre

Ontology Modeling for Criminal Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

Chiseung Soh, Seungtak Lim, Kihyun Hong, and Young-Yik Rhim

ContrattiPubblici.org, a Semantic Knowledge Graph on Public

Procurement Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
Giuseppe Futia, Federico Morando, Alessio Melandri, Lorenzo Canova,
and Francesco Ruggiero

Application of Ontology Modularization for Building a Criminal

Domain Ontology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
Mirna El Ghosh, Hala Naja, Habib Abdulrab, and Mohamad Khalil

A Linked Data Terminology for Copyright Based on Ontolex-Lemon . . . . . . 410

Víctor Rodriguez-Doncel, Cristiana Santos, Pompeu Casanovas,
Asunción Gómez-Pérez, and Jorge Gracia

Legal Argumentation

Anything You Say May Be Used Against You in a Court of Law:

Abstract Agent Argumentation (Triple-A). . . . . . . . . . . . . . . . . . . . . . . . . . 427
Ryuta Arisaka, Ken Satoh, and Leendert van der Torre

A Machine Learning Approach to Argument Mining in Legal Documents . . . 443

Prakash Poudyal

Answering Complex Queries on Legal Networks: A Direct

and a Structured IR Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Nada Mimouni, Adeline Nazarenko, and Sylvie Salotti

Inducing Predictive Models for Decision Support

in Administrative Adjudication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
L. Karl Branting, Alexander Yeh, Brandy Weiss, Elizabeth Merkhofer,
and Bradford Brown

Arguments on the Interpretation of Sources of Law . . . . . . . . . . . . . . . . . . . 478

Robert van Doesburg and Tom van Engers
XII Contents

Courts, Adjudication and Dispute Resolution

Dynamics of the Judicial Process by Defeater Activation . . . . . . . . . . . . . . . 495

Martín O. Moguillansky and Guillermo R. Simari

Claim Detection in Judgments of the EU Court of Justice . . . . . . . . . . . . . . 513

Marco Lippi, Francesca Lagioia, Giuseppe Contissa,
Giovanni Sartor, and Paolo Torroni

A Non-intrusive Approach to Measuring Trust in Opponents

in a Negotiation Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
Marco Gomes, John Zeleznikow, and Paulo Novais

Network, Visualization, Analytics. A Tool Allowing Legal Scholars

to Experimentally Investigate EU Case Law . . . . . . . . . . . . . . . . . . . . . . . . 543
Nicola Lettieri, Sebastiano Faro, Delfina Malandrino,
Armando Faggiano, and Margherita Vestoso

Electronic Evidence Semantic Structure: Exchanging Evidence Across

Europe in a Coherent and Consistent Way . . . . . . . . . . . . . . . . . . . . . . . . . 556
Maria Angela Biasiotti, Sara Conti, and Fabrizio Turchi

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575

 Introduction: Legal and Ethical Dimensions
of AI, NorMAS, and the Web of Data

Ugo Pagallo1(&) , Monica Palmirani2 , Pompeu Casanovas3,4 ,

Giovanni Sartor2 , and Serena Villata5
1
Torino Law School, University of Torino, Lungo Dora Siena 100, 10153
Turin, Italy
ugo.pagallo@unito.it
2
CIRSFID, University of Bologna, via Zamboni 33, 40126 Bologna, Italy
{monica.palmirani,giovanni.sartor}@unibo.it
3
CRC D2D, La Trobe Law School, Bundoora Campus, Melbourne, Australia
pompeu.casanovas@uab.cat
4
IDT, Autonomous University of Barcelona, Bellaterra, Spain
5
INRIA Sophia Antipolis, Sophia Antipolis Cedex, France
serena.villata@inria.fr

Abstract. AICOL workshops aim to bridge the multiple ways of understanding

legal systems and legal reasoning in the field of AI and Law. Moreover, they pay
special attention to the complexity of both legal systems and legal studies, on
one hand, and the expanding power of the internet and engineering applications,
on the other. Along with a fruitful interaction and exchange of methodologies
and knowledge between some of the most relevant contributions to AI work on
contemporary legal systems, the goal is to integrate such a discussion with legal
theory, political philosophy, and empirical legal approaches. More particularly,
we focus on four subjects, namely, (i) language and complex systems in law;
(ii) ontologies and the representation of legal knowledge; (iii) argumentation and
logics; (iv) dialogue and legal multimedia.

Keywords: AI & law
Legal theory
Complex systems
Semantic web

Legal ontologies
Legal semantic web services
Argumentation

1 Introduction

The first volume of this workshop series on Artificial Intelligence approaches to the
complexity of legal systems (AICOL) was released at the very beginning of this decade
(2010). In the meanwhile, the field of Artificial Intelligence (“AI”) has known a new
renaissance: for instance, according to the tally Google provided to MIT Technology
Review in March 2017, the company published 218 journal or conference papers on
machine learning in 2016 alone, nearly twice as many as it did two years before [1].
Google’s AI explosion illustrates a more general trend that has to do with the
improvement of more sophisticated statistical and probabilistic methods, the increasing
availability of large amount of data and of cheap, enormous computational power, up to
the transformation of places and spaces into AI-friendly environments, e.g. smart cities

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 1–20, 2018.
https://doi.org/10.1007/978-3-030-00178-0_1
2 U. Pagallo et al.

and domotics. All these factors have propelled a new ‘summer’ for AI. After the
military and business sectors, AI applications have entered into people’s lives. From
getting insurance to landing credit, from going to college to finding a job, even down to
the use of GPS for navigation and interaction with the voice recognition features on our
smartphones, AI is transforming and reshaping people’s daily interaction with others
and their environment. Whereas AI apps and systems often go hand-in-hand with the
breath-taking advancements in the field of robotics, the internet of things, and more, we
can grasp what is going on in this field in different ways [2]. Suffice it to mention in this
context two of them.
First, according to the Director of the Information Innovation Office (I2O) at the
Defense Advanced Research Projects Agency (DARPA) in the U.S. Department of
Defense, John Launchbury, there would have been so far two waves of research in AI.1
The first wave concerns systems based on “handcrafted knowledge,” such as programs
for logistics scheduling, programs that play chess, and in the legal domain, TurboTax.
Here, experts turn the complexity of the law into certain rules, and “the computer then
is able to work through these rules.” Although such AI systems excel at complex
reasoning, they were inadequate at perception and learning. In order to overcome these
limits, we thus had to wait for the second wave of AI, that is, systems based on
“statistical learning” and the “manifold hypothesis.” As shown by systems for voice
recognition and face recognition, the overall idea is that “natural data forms lower-
dimensional structures (manifolds) in the embedding space” and that the task for a
learning system is to separate these manifolds by “stretching” and “squashing” the
space.2
Along these lines, Richard and Daniel Susskind similarly propose to distinguish
between a first generation and a second generation of AI systems, namely, between
expert systems technologies and systems characterized by major progress in Big Data
and search [3]. What this new wave of AI entails has to do more with the impact of AI
on society and the law, than the law as a rich test bed and important application field for
logic-based AI research. Whether or not the first wave aimed to replicate knowledge
and reasoning processes that underpin human intelligence as a form of deductive logic,
the second wave of AI brings about “two possible futures for the professions,” that is,
either a more efficient version of the current state of affairs, or a profound transfor-
mation that will displace much of the work of traditional professionals. According to
this stance, we can thus expect that “in the short and medium terms, these two futures
will be realized in parallel.” Yet, the thesis of Richard and Daniel Susskind is that the
second future will prevail: in their phrasing, “we will find new and better ways to share
expertise in society, and our professions will steadily be dismantled” [3].
Against this highly problematical backdrop, three different levels of analysis should
be however differentiated. They concern the normative challenges of the second wave
of AI from a political, theoretical, and technical viewpoint, namely (i) the political
decisions that should be—or have already been—taken vis-à-vis current developments

1
See the video entitled “A DARPA Perspective on Artificial Intelligence,” available online at https://
youtube/-O01G3tSYpU.
2
Ibid.
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 3

of e.g. self-driving cars, or autonomous lethal weapons; (ii) the profound transforma-
tions that affect today’s legal systems vis-à-vis the employment of e.g. machine
learning techniques; and, (iii) the advancements in the state-of-the-art that regard such
areas, as semantic web applications and language knowledge management in the legal
domain, or ejustice advanced applications. Each one of these levels of analysis is
deepening in the following sections.

2 Architectural Challenges

The political stance on current developments of AI hinges on a basic fact: the more the
second wave of AI advances, the more AI impacts on current pillars of society and the
law, so that political decisions will have to be taken as regards some AI applications,
such as lethal autonomous weapons, or self-driving cars. Over the past years, scholars,
non-profit organizations, and institutions alike have increasingly stressed the ethical
concerns and normative challenges brought about by many autonomous and intelligent
system designs [4–6]. The aim of the law to govern this field of technological inno-
vation suggests that we should distinguish between two different levels of political
intervention, that is, either through the primary rules of the law, or through its sec-
ondary rules [7].
According to the primary rules of the law, the goal is to directly govern social and
individual behaviour through the menace of legal sanctions. Legislators have so far
aimed to attain this end through methods of accident control that either cut back on the
scale of the activity via, e.g., strict liability rules, or intend to prevent such activities
through bans, or the precautionary principle. Regulations can be divided into four
different categories, that is, (a) the regulation of human producers and designers of AI
systems through law, e.g. either through ISO standards or liability norms for users of
AI; (b) the regulation of user behaviour through the design of AI, that is, by designing
AI systems in such a way that unlawful actions of humans are not allowed; (c) the
regulation of the legal effects of AI behaviour through the norms set up by lawmakers,
e.g. the effects of contracts and negotiations through AI applications; and, (d) the
regulation of AI behaviour through design, that is, by embedding normative constraints
into the design of the AI system [8].
Current default norms of legal responsibility can entail however a vicious circle,
since e.g. strict liability rules—let aside bans, or the precautionary principle—may end
up hindering research and development in this field. The recent wave of extremely
detailed regulations on the use of drones by the Italian Civil Aviation Authority, i.e.
“ENAC,” illustrates this deadlock. The paradox stressed in the field of web security
decades ago, could indeed be extended with a pinch of salt to the Italian regulation on
the use of drones as well: the only legal drone would be “one that is powered off, cast in
a block of concrete and sealed in a lead-lined room with armed guards – and even then
I have my doubts.” [9] As a result, we often lack enough data on the probability of
events, their consequences and costs, to determine the levels of risk and thus, the
amount of insurance premiums and further mechanisms, on which new forms of
accountability for the behaviour of such systems may hinge. How, then, can we prevent
legislations that may hinder the research in AI? How should we deal with the peculiar
4 U. Pagallo et al.

unpredictability and risky behaviour of some AI systems? How should we legally

regulate the future?
A feasible way out can be given by the secondary rules of the law, namely, the rules
of the law that create, modify, or suppress the primary rules of the system. Among the
multiple legal techniques with which we can properly address the normative challenges
of the second wave of AI, suffice it to mention here three of them. First, focus should be
on Justice Brandeis’s doctrine of experimental federalism, as espoused in New State Ice
Co. v Leibmann (285 US 262 (1932)). The idea is to flesh out the content of the rules
that shall govern individual behaviour through a beneficial competition among legal
systems. This is what occurs nowadays in the field of self-driving cars in the US, where
several states have enacted their own laws for this kind of technology. At its best
possible light, the same policy will be at work with the EU regulation in the field of
data protection [10, 11].
Second, attention should be drawn to the principle of implementation neutrality,
according to which regulations are by definition specific to that technology and yet do
not favour one or more of its possible implementations. The 2016 Federal Automated
Vehicles Policy of the U.S. Department of Transportation illustrates this legal tech-
nique. Although regulations are by definition specific to that technology, e.g. auton-
omous vehicles, there is no favouritism for one or more of its possible implementations.
Even when the law sets up a particular attribute of that technology, lawmakers can draft
the legal requirement in such a way that non-compliant implementations can be
modified to become compliant.
Third, legislators can adopt forms of legal experimentation. For example, over the
past decade and a half, the Japanese government has worked out a way to address the
normative challenges of robotics through the creation of special zones for their
empirical testing and development, namely, a form of living lab, or Tokku [12].
Likewise, in the field of autonomous vehicles, several EU countries have endorsed this
kind of approach: Sweden has sponsored the world’s first large-scale autonomous
driving pilot project, in which self-driving cars use public roads in everyday driving
conditions; Germany has allowed a number of tests with various levels of automation
on highways, e.g. Audi’s tests with an autonomously driving car on highway A9
between Ingolstadt and Nuremberg.
In general terms, these forms of experimentation through lawfully de-regulated
special zones represent the legal basis on which to collect empirical data and sufficient
knowledge to make rational decisions for a number of critical issues. We can improve
our understanding of how AI systems may react in various contexts and satisfy human
needs. We can better appreciate risks and threats brought on by possible losses of
control of AI systems, so as to keep them in check. We can further develop theoretical
frameworks that allow us to better appreciate the space of potential systems that avoid
undesirable behaviours. In addition, we can rationally address the legal aspects of this
experimentation, covering many potential issues raised by the next-generation AI
systems and managing such requirements, which often represent a formidable obstacle
for this kind of research, as public authorizations for security reasons, formal consent
for the processing and use of personal data, mechanisms of distributing risks through
insurance models and authentication systems, and more. The different legal techniques
and types of rules that lawmakers may employ, on the one hand, should not overlook
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 5

the importance of the goals and values that are at stake with choices of technological
dependence, delegation and trust, in order to determine the good mix between legal
automation and public deliberation [13]. On the other hand, such choices of techno-
logical dependence, delegation and trust, through AI systems and procedures of legal
automation are affecting pillars and tenets of today’s law. As stressed above in this
introduction, AI technology profoundly affects both the requirements and functions of
the law, namely, what the law is supposed to be (requirements), and what it is called to
do (functions). This profound transformation has to be examined separately in the next
section.

3 Ethical and Legal Challenges: Device and Linked

Democracy

In one of the most celebrated 2014 John Klossner’s cartoons on the Internet of Things
the husband resignedly says to his wife: “We have to go out for dinner. The refrigerator
isn’t speaking to the stove.”3 This is not a joke anymore, and neither is the possibility
of connecting thousands of billions of devices that can literally speak to each other.
A world of smart objects shreds new challenges into the interconnected world of
humans and machines. The 2015 IBM Institute for Business Value Report [14] has
pointed out five major challenges: (i) the cost of connectivity (prohibitively high),
(ii) the Internet after trust (in the after-Snowden era “trust is over” and “IoT solutions
built as centralized systems with trusted partners is now something of a fantasy”),
(iii) not-future proof (many companies are quick to enter the market but it is very hard
to exit: the cost of software updates and fixes in products long obsolete and discon-
tinued), (iv) a lack of functional value (lack of meaningful value creation), (v) broken
business models (in information markets, the marginal cost of additional capacity—
advertising—or incremental supply—user data—is zero).
This is setting the conditions for “Device Democracy”, in which “devices are
empowered to autonomously execute digital contracts such as agreements, payments
and barters with peer devices by searching for their own software updates, verifying
trustworthiness with peers, and paying for and exchanging resources and services. This
allows them to function as self-maintaining, self-servicing devices” [14].
IBM suggests three new methodological trends for a scalable, secure, and efficient
IoT regarding: (i) architecture (private-by-design), (ii) business and economic insights
(key vectors of disruption), (iii) and product and user experience design (the trans-
formation of physical products into meaningful digital experiences).
The keyword here is the emergence of “meaningful experiences” in between
relationships, properties, and objects. Interestingly this has been also enhanced by 2016
and 2017 Gartner Hype Cycle of Emerging Technologies: (i) transparently immersive
experiences (such as human augmentation), (ii) perceptual smart machines (such as

3
https://www.computerworld.com/article/2858429/enterprise-applications/2014-the-tech-year-in-
cartoons.html#slide5.
6 U. Pagallo et al.

personal analytics), and (iii) digital platforms (including blockchain technologies).4

“AI everywhere” and blockchain are especially highlighted.
Our contention is that meaning is created, distributed, and framed through a
complex and dynamic world in which ethics and law cannot be set apart or let alone to
regulate processes, actions, and outcomes. Thus, regulations are entrenched and evolve
dynamically according to the network and the specific ecosystems they are contributing
to create. Meaningful experiences and smart devices entail smart regulations.
This is a feature already stressed for many years now by all attempts to frame new
developments in norMAS [15], non-standard deontic logic [16] and law and the
semantic web [17]. What is new is the attention brought to encompass innovation,
semantic developments and blockchain technologies with ethical, democratic (politi-
cal), and legal values alike.
In a recent Nexxus Whitepaper,5 Gavin Andresen, the leader of Blockchain
Foundation, equates crypto-currencies with empowering people: “an unstoppable
grass-roots movement that won’t be trampled on by any government or bank. It’s all
about the people’s freedom and reclaiming it” [18].
The first generation produced blockchain-based network protocols, such as Bitcoin,
Ethereum, Litecoin, and Zerocash. The second one can take this same idea of creating
encrypted blocks further: it uses distributed ledger databases along with user-
programmable smart contracts6 to create a number of social contracts in many fields—
universal basic income schemes, birth and death certificates, business licenses, property
titles, educational qualifications, marriage etc. This actually is the original dream of the
semantic web. MIT Digital Currency Initiative furnishes good examples in the public
domain, as they have launched the identity based services offered by BitNation, a
project aimed at decentralising governance at a global scale, e.g. a World Citizen-
ship ID based on blockchain, and a Refugee Emergency Response project. Ideally,
when applied to social or economic institutions, the result would be Decentralized
Autonomous Organizations (DAO): self-running organisations in which users can
become part of the chain performing things that computers cannot do [19].7
However, there is still a long road ahead before implementing DAO properly, as
complexity does not only lie on regulations, and transactions fuel a mixed, hybrid
social and economic reality that displays its own problems and conflicts. The recent
Ethereum crisis has shown security vulnerabilities [20]. Although preventive mecha-
nisms such as distributed consensus, cryptography, and anonymity are put in place,
iblockchain technologies remain vulnerable to many types of risks—mainly: the 51%
attack, account takeover, digital identity theft, money laundering, and hacking [21].

4
https://www.gartner.com/smarterwithgartner/.
5
Nexxus Partners was established in January 2016 in Texas, USA as a services company for the
bitcoin and cryptocurrency industry.
6
“Smart contracts are computer programs that can be correctly executed by a network of mutually
distrusting nodes, without the need of an external trusted authority” [20].
7
See the first Decentralized Autonomous Organization (DAO) code to automate organizational
governance and decision-making at [19].
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 7

There are problems regarding the legal dimension as well. Some criticisms have
already stated that “cryptocurrencies cannot solve the problem of incomplete [rela-
tional] contracts, and as long as contracts are incomplete, humans will need to resolve
ambiguities” [22]. The same diffidence has been shown from a public legal standpoint,
as crypto-currencies cannot build up by their own a new public space [23].
The other way around, there are economic interpretations that highlight its positive
aspects. Ethereum and blockchain platforms have been received “as a new type of
economy: a ‘spontaneous organization’, which is a self-governing organization with the
coordination properties of a market (Hayek), the governance properties of a commons
(Ostrom), and the constitutional properties of a nation state (Brenan and Buchanan)” [24].
Perhaps this syncretic view is too over-confident, equating different dimensions
(economic, social, and legal) but, be as it may, law and the definition of law—what it
counts for—are at stake: “the legal status of DAOs remains the subject of active and
vigorous debate and discussion. Not everyone shares the same definition. […] Ulti-
mately, how a DAO functions and its legal status will depend on many factors,
including how DAO code is used, where it is used, and who uses it.” [19]
It is our contention that the synergy between different kinds of complementary
technologies can help to solve these regulatory puzzles and tensions, i.e. Blockchain is
the result of assembling two software paradigms (peer-to-peer applications and dis-
tributed hash tables). They are not the only ones. Semantic technologies and linked data
can be used to ease the tensions and create the shared scenarios in which crypto-
currencies and smart contracts can be safely and effectively used in a personalised
manner by a vast plurality of users. “Similarly as block chain technology can facilitate
distributed currency, trust and contracts application, Linked Data facilitated distributed
data management without central authorities” [25].
But for this to happen, to cross jurisdictions and different types of legal obstacles,
smart regulations and values are essential and should be similarly linked and har-
monised. Traditional legal tools at national, European and international levels, are
important, but they still fall short to cope with the complexity of algorithm governance
to reach metadata regulatory dimensions and layers [26]. This is why law, governance,
and ethics are at the same time being embedded into design, and re-enacted again as
contextually-driven to shape sustainable regulatory ecosystems. Beyond epistemic and
deliberative democracy, one of the concepts that have recently coined to describe this
new situation is linked democracy, i.e. the endorsement of (embedded) democratic
values to preserve rights and protect people on the web of data [27]. It is worth noticing
that these common trends are related to the combination of political crowdsourcing,
legal and ethical argumentation, and expert knowledge [28]. Innovation is deemed to
be a crucial component of democracy [29]. Thus, what the law is supposed to be and
what it is called to do are related not only to its architecture and tools (e.g. normative
systems, laws and rights) but to the many ways of balancing citizens’ compliance and
participation.
8 U. Pagallo et al.

4 Web of Data and Legal Analytics

Over the last two decades we have witnessed a remarkable volume of legal documents
and legal big data being put out in open format (e.g., the legal XML movement). The
information was represented using specific technical standards capable of modelling
legal knowledge, norms, and concepts [17, 30] in machine-readable format.
NormInRete [31] is an XML standard the Italian government issued in 2001 as the
official XML vocabulary for the country’s legislative documents. MetaLex was created
in 2002 in the Netherlands, and it evolved into CEN-MetaLex as a general format for
the interoperability of legal documents across Europe, this thanks to the EU Project
ESTRELLA [32]. Another significant outcome of the ESTRELLA project was the Legal
Knowledge Interchange Format-LKIF, composed of two main pillars: (i) a core legal
ontology [33] and (ii) a legal-rule language [34]. Even if these outcomes are encour-
aging, they lack a common-framework technical design making it possible to easily
integrate all the Semantic Web layers (e.g., text, norms, ontology). For this reason, the
Akoma Ntoso project [35] (an UNDESA-based African initiative)8 took the best
practices from those experiences and in 2006 designed a unique XSD schema for all
legal documents (e.g., including caselaw and UN resolutions [39]) and lawmaking
traditions (e.g., common law and civil law). In 2012, LegalDocML TC,9 of OASIS,
expanded the Akoma Ntoso XML vocabulary to embrace an international vision of
legal-document annotation. OASIS’s LegalCiteM10 TC provides semantic representa-
tion of legal references so as to foster a convergence of many existing syntaxes for legal
and legislative identifiers, including ELI [37], ECLI [36], URN-LEX,11 and the Akoma
Ntoso Naming Convention [40], making sure that legal document collections can
unambiguously be referred to and are also connectable to Linked Data assertions.
OASIS’s LegalRuleML TC [38] provides a standard for modelling constitutive and
prescriptive norms using formal language for rules. LegalDocML, LegalRuleML, and
LegalCiteM provide a common framework for modelling legal documents and for
fostering contextual metadata. The CLOUD4EU [41, 42] project offers a rare example
of a platform where those standards can act in an integrated manner: it is designed for
the General Data Protection Regulation (GDPR), making it possible to provide com-
pliance reports for this regulation.
LegalRuleML also provides an RDFS meta-model for modelling the deontic and
defeasible logic operators applied in the legal domain in order to export metadata in
RDF format. LegalDocML makes it possible to extract legal metadata and to convert it
into RDF. In the web of data paradigm, RDF triples produce a distributed and net-
worked legal knowledge repository that can be useful in enhancing the searchability of
relevant legal concepts, the semantic classification of documents, a light legal-
reasoning approach, and the integration of metadata with other nonlegal sources (e.g.,

8
United Nations Department of Economic and Social Affairs https://www.un.org/development/desa/
en/.
9
https://www.oasis-open.org/committees/legaldocml/.
10
https://www.oasis-open.org/committees/tc_home.php?wg_abbrev=legalcitem.
11
https://datatracker.ietf.org/doc/draft-spinosa-urn-lex/.
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 9

DBpedia). However, the RDF technique is based on a collection of triples (assertions

composed of a subject, a predicate, and an object), often not validated, and the
underlying theory of inferential logic only approximates reality, like a map. In the
complex nonmonotonic legal system, RDF reasoning is quite risky, especially in situ-
ations where antecedents are rebutted (e.g., retroactivity, modifications of modifica-
tions, exceptions to exceptions, suspensions of the application of norms, annulment,
etc.). For this reason, the legal XML community is inclined to extract RDF assertions
as a subproduct of legal XML documents that are authentically, officially validated by
experts (e.g., principle of self-contained and self-explained assertions in a unique XML
file validated by the expert).
In the meantime, this large web of legal data connected with official digital legal
documents (often available from official-journal portals) in open format provides a rare
opportunity to apply artificial intelligence (AI) techniques in order to rethink the legal
theory and their institutions. In particular, legal analytics (LA)—combining data sci-
ence, artificial intelligence, machine learning (ML), natural language processing
techniques, and statistical methods—can reuse that vast and varied body of legal big
data, even if approximate, to infer new patterns and legal knowledge, so as to then
predict new models and bring out hidden correlations, often providing unexpected
insights into the relation between legal phenomena [43].
The automated application of learning methods to vast sets of examples makes it
possible to reproduce human behaviour and improve upon it, exploiting their com-
putational power to learn from successes and failures and thus improve performance.
Thanks to these new techniques and the sheer volume of the legal web of data now
available on the Internet, artificial intelligence has been able to advance from mock
examples to a host of real-life applications: conceptual retrieval and ranking, speech
and image recognition, question-answering, recommendations, translation, planning,
autonomous mobile robots, etc. Machine learning has been extensively applied to text
analytics, which refers to the use of multiple technologies—such as linguistic, statis-
tical, and machine learning techniques—to capture the information content of textual
sources. Considering that legal sources of all kinds are recorded in textual form (this is
true of statutes, regulations, judicial and other types of decisions and opinions, legal
doctrines, contracts), the application of text analytics to the law has huge potentials.
And indeed a number of applications are emerging, making it possible to automatically
classify and extract documents, identify principles in judicial decisions, and predict the
outcome of judicial cases.
Legal analytics (LA) has the potential to contribute to different aspects of legal
scholarship and practice. LA can advance legal informatics, making for a vast range of
successful new legal applications in the public and private sectors alike. In particular, it
supports the provision of AI applications, but overcoming the knowledge-acquisition
bottleneck. LA can support legal research, in particular, through its ability to unveil
hidden patterns in data and documents, revealing unseen features of the structure and
the functioning of legal institutions. The insights provided by LA, in combination with
computable models of the law, support the development of a new, empirically based
understanding of the law. LA can improve the efficiency and effectiveness of legal
institutions, while providing all legal actors with better knowledge of the law and of its
application. However, the use of big-data analytics has also raised a number of ethical
10 U. Pagallo et al.

issues that are already being discussed in several domains (targeted deceptive com-
mercial and noncommercial communication, discrimination, manipulation of public
opinion, etc.). Also already emerging in the law are some questionable practices,
particularly where law enforcement tries to predict illegal or otherwise unwanted
behaviour (e.g., a tendency to offend or reoffend). More generally, the knowledge
provided by analytics brings new ways to assess, influence, and control behaviour, and
these aspects have yet to be fully analysed. For instance, no study exists so far on how
the ability to predict court decisions, even those of specific judges, could influence
judicial decision-making.
The combination of LA and legal XML techniques improve our interpretation of
the AI inferences. XML nodes provide structural information and contextual metadata
that, in combination with the predictive assertions of LA, could be used to mitigate two
important negative side effects of LA techniques: (i) the introduction of bias from the
past experiences and mistakes (e.g., negative case-law, bad legal drafting practices in
legislation, influences due to socio-historical conditions) that can reinforce a tendency
to reiterate incorrect models and may impair the ability to creatively find brilliant new
solutions in the future (e.g., through filter-bubble effects); (ii) the fragmentation of legal
knowledge into separate sentences or isolated data without a logical connection making
it possible to achieve a consistent legal and logical narrative flow (e.g., contextless
prediction). XML nodes could provide the skeleton needed to reassemble the huge
amount of unexpected insights produced by the LA layer.
Finally, also crucial is the usability and the easy access to legal knowledge pro-
duced by LA. It is essential that the outcome of LA and legal XML sources in web
applications and new devices (e.g., augmented reality) be also understandable by
people who are not legal experts, without reframing the message; and, at any event, it is
also essential to provide clear mechanisms for explaining the algorithm decision-
making process and outcome. In this effort to achieve transparent communication we
can turn to human-computer interaction techniques, making it possible to create a fair
environment in which to better communicate the legal concepts and principles
extracted by LA. The legal design community is working to create new design patterns,
looking to provide better ways of displaying content, in such a way that the legal
community and end users (e.g., citizens) can place greater trust in LA and legal XML
[44–46].

5 On the Content of This Volume

This new AICOL volume is divided into six parts. They concern (i) legal philosophy,
conceptual analysis, and epistemic approaches; (ii) rules and norms analysis and rep-
resentation; (iii) legal vocabularies and natural language processing; (iv) legal
ontologies and semantic annotation; (v) legal argumentation; and, (vi) courts, adjudi-
cation and dispute resolution.
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 11

5.1 Legal Philosophy, Conceptual Analysis, and Epistemic Approaches

In the first part of this volume, four papers deal with matters of legal philosophy,
conceptual analysis, and epistemic approaches. In RoboPrivacy and the Law as “Meta-
technology”, Ugo Pagallo examines how a particular class of robotic applications, i.e.
service robots, or consumer robots, may affect current legal frameworks of privacy and
data protection. Instead of a one-way movement of social evolution from technology to
law, a key component of the analysis concerns the aim of the law to govern techno-
logical innovation as well as human and artificial behaviour through the regulatory
tools of technology. By distinguishing between the primary rules of the law and its
secondary rules, e.g. forms of legal experimentation, the chapter illustrates some of the
ways in which the secondary rules of the law may allow us to understand what kind of
primary rules we may want for our robots. In Revisiting Constitutive Rules, Giovanni
Sileno, Alexander Boer and Tom Van Engers investigate how behaviour relates to
norms, i.e. how a certain conduct acquires meaning in institutional terms. By
addressing the double function of the ‘count-as’ relation, generally associated to
constitutive rules and mostly accounted for its classificatory functions, the chapter
reconsiders the relation between constitutive rules and regulative rules, and introduces
a preliminary account on the ontological status of constitution. In The Truth in Law and
Its Explication, Hajime Yoshino discusses what types of truth play their role in law and
in what way such types of truth can be explicated. Whereas the concept of truth in law
is classified into three types of truth, i.e. truth as fact, truth as validity and truth as
justice, the chapter provides their formal semantic foundation, and analyses both the
ways to explicate truth as validity and truth as justice in terms of intensional and
extensional explication, and how to grasp the reasoning of justification and of creation.
In From Words to Images Through Legal Visualization, Arianna Rossi and Monica
Palmirani discuss the process of sense-making and interpretation of visual legal con-
cepts that have been introduced in legal documents to make their meaning clearer and
more intelligible. Whilst visualizations have also been automatically generated from
semantically-enriched legal data, the analysis of current approaches to this subject
represents the starting point to propose an empirical methodology that is inspired by the
interaction with design practices and that will be tested in the future stages of the
research.

5.2 Rules and Norms Analysis and Representation

The second part of the volume, which has to do with rules and norms analysis and
representation, comprises seven contributions. In A Petri Net-based Notation for
Normative Modelling: Evaluation on Deontic Paradoxes, Giovanni Sileno, Alexander
Boer and Tom Van Engers focus on some of the problems that derive from the
development of systems operating in alignment with norms, e.g. the continuous flow of
events that modifies the normative directives under scrutiny. The chapter presents an
alternative approach to some of these problems, by extending the Petri net notation to
Logic Programming Petri Nets, so that the resulting visual formalism represents in an
integrated, yet distinct fashion, procedural and declarative aspects of the system. In
Legal Patterns for Different Constitutive Rules, Marcello Ceci, Tom Butler, Leona
12 U. Pagallo et al.

O’Brien and Firas Al Khalil illustrate a heuristic approach for the representation of
alethic statements as part of a methodology aimed at ensuring effective translation of
the regulatory text into a machine-readable language. The methodology includes an
intermediate language, accompanied by an XML persistence model, and introduces a
set of “legal concept patterns” to specifically represent the different constitutive
statements that can be found in e.g. financial regulations. In An Architecture for
Establishing Legal Semantic Workflows in the Context of Integrated Law Enforcement,
Markus Stumpner, Wolfgang Mayer, Pompeu Casanovas and Louis de Koker develop
a federated data platform that aims to enable the execution of integrated analytics on
data accessed from different external and internal sources, and to enable effective
support of an investigator or analyst working to evaluate evidence and manage
investigation structure. By preventing the shortcomings of traditional approaches, e.g.
high costs and silos-effects, the chapter also aims to show how this integration can be
compliant. In Contributions to Modelling Patent Claims when Representing Patent
Knowledge, Simone Reis, Andre Reis, Jordi Carrabina and Pompeu Casanovas
examine the modelling of patent claims in ontology based representation of patent
information. They relate to the internal structure of the claims and the use of the all-
element rule for patent coverage, in order to offer the general template for the structure
of the claim, and provide the visualization of the claims, the storage of claim infor-
mation in a web semantics framework, and the evaluation of claim coverage using
Description Logic. In Execution and Analysis of Formalized Legal Norms in Model
Based Decision Structures, Bernhard Waltl, Thomas Reschenhofer and Florian Matthes
describe a decision support system to represent the semantics of legal norms, whereas a
model based expression language (MxL) has been developed to coherently support the
formalization of logical and arithmetical operations. Such legal expert system is built
upon model based decision structures and three different components, namely a model
store, a model execution component, and an interaction component, have been worked
out, so as to finally test the execution and analysis of such structured legal norms vis-à-
vis the German child benefit regulations. In Causal Models of Legal Cases, Ruta
Liepina, Giovanni Sartor and Adam Wyner draw the attention to the requirements for
establishing and reasoning with causal links. In light of a semi-formal framework for
reasoning with causation that uses strict and defeasible rules for modelling factual
causation in legal cases, the chapter takes into account the complex relation between
formal, common sense, norm and policy based considerations of causation in legal
decision making with particular focus on their role in comparing alternative causal
explanations. In Developing Rule-Based Expert System for People with Disabilities,
Michał Araszkiewicz and Maciej Klodawski present the features of a moderately
simple legal expert system devoted to solving the most frequent legal problems of
disabled persons in Poland. By casting light on the structure of the expert system and
its methodology, the succession law of Poland and its procedures delivers sufficient
material to reveal the most important issues concerning such a project on a rule-based
expert system.
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 13

5.3 Legal Vocabularies and Natural Language Processing

The third part of the volume regards legal vocabularies and natural language pro-
cessing. The eight contributions include EuroVoc-based Summarization of European
Case Law, in which Florian Schmedding illustrates the on-going development of a
multilingual pipeline for the summarization of European case law. By applying the
TextRank algorithm on concepts of the EuroVoc thesaurus, so as to extract summa-
rizing keywords and sentences, the intent is to demonstrate the feasibility and use-
fulness of the presented approach for five different languages and 18 document sources.
In Aligning Legivoc Legal Vocabularies by Crowdsourcing, Hughes-Jehan Vibert,
Benoit Pin and Pierre Jouvelot present the first Internet-based platform dedicated to the
diffusion, edition and alignment of legal vocabularies across countries. As a seamless
path for governments to disseminate their legal foundations and specify semantic
bridges between them, the chapter describes the general principles behind the legivoc
framework while providing some ideas about its implementation, e.g. crowdsourcing
the alignment of legal corpora together. In Data Protection in Elderly Health Care
Platforms, Ângelo Costa, Aliaksandra Yelshyna, Teresa C. Moreira, Francisco
Andrade, Vicente Julián and Paulo Novais deal with solutions to the increasing cog-
nitive problems that affect the elderly population in light of the iGenda project, which
aims to build safe environments that adapt themselves to one’s individual needs
through a Cognitive Assistant inserted in the Ambient Assisted Living area. Whereas
one of the main issues concerns the protection of the data flowing within the system
and the protection of the user’s fundamental rights, the chapter clarifies the principles
and legal guarantees of data protection, embracing appropriate solutions for techno-
logical features that may be a threat. In Assigning Creative Commons Licenses to
Research Metadata: Issues and Cases, Marta Poblet, Amir Aryani, Paolo Manghi,
Kathryn Unsworth, Jingbo Wang, Brigitte Hausstein, Sunje Dallmeier-Tiessen, Claus-
Peter Klas, Pompeu Casanovas and Víctor Rodriguez-Doncel tackle the problem of
lack of clear licensing and transparency of usage terms and conditions for research
metadata. Making research data connected, discoverable and reusable are the key
enablers of the new data revolution in research. Accordingly, the chapter does not only
discuss how the lack of transparency can hinder discovery of research data and make it
disconnected from publication and other trusted research outcomes. In addition, the
chapter suggests the application of Creative Commons licenses for research metadata,
and provides some examples of the applicability of this approach to internationally
known data infrastructures. In Dataset Alignment and Lexicalization to Support Mul-
tilingual Analysis of Legal Documents, Armando Stellato, Manuel Fiorelli, Andrea
Turbati, Tiziano Lorenzetti, Peter Schmitz, Enrico Francesconi, Najeh Hajlaoui and
Brahim Batouche tackle the complexity of the EU legal system, in which both the
linguistic and the conceptual aspects mutually interweave into a knowledge barrier that
is hard to break. In order to create a platform for multilingual cross-jurisdiction
accessibility to legal content in the EU, the chapter addresses the challenge of Semantic
Interoperability at both the conceptual and lexical level, by developing a coordinated
set of instruments for advanced lexicalization of RDF resources (be them ontologies,
thesauri and datasets in general) and for alignment of their content. In A Multilingual
Access Module to Legal Texts, Kiril Simov, Petya Osenova, Iliana Simova, Hristo
14 U. Pagallo et al.

Konstantinov and Tenyo Tyankov introduce a Multilingual Access Module, which

translates the user’s legislation query from its source language into the target language,
and retrieves the detected texts that match the query. More particularly, the unit con-
sists of two sub modules, i.e. an Ontology-based and a Statistical Machine Translation
units, which have their own drawbacks, so that both are used in an integrated archi-
tecture, in order to profiting from each other. In Combining Natural Language Pro-
cessing Approaches for Rule Extraction from Legal Documents, Mauro Dragoni,
Serena Villata, Williams Rizzi and Guido Governatori address the problem of moving
from a natural language legal text to the respective set of machine-readable conditions,
by combining the linguistic information provided by WordNet and a syntax-based
extraction of rules from legal texts, with a logic-based extraction of dependencies
between chunks of such texts. Such a combined approach leads to a powerful solution
towards the extraction of machine-readable rules from legal documents, which is
evaluated over the Australian “Telecommunications consumer protections code”. In
Analysis of Legal References in an Emergency Legislative Setting, Monica Palmirani,
Luca Cervone, Ilaria Bianchi and Francesco Draicchio provide for a taxonomy of legal
citations that set up an interesting apparatus for analysing a country’s legislative
approach. By investigating the references of a legal corpus of ordinances issued by the
Regional Commissioner for Emergency and Reconstruction over the first eighteen
months after the 2012 earthquake in Emilia-Romagna, the chapter scrutinizes the
critical issues arising in the regulative strategy for emergency situations. By distin-
guishing groupings based on lexical-textual analysis and groupings based on structural
element, the aim is to help lawmakers act better in future disasters, to extract infor-
mation concerning the number and the types of modifications produced, and to support
the debate on emergency national laws that deal with natural disasters.

5.4 Legal Ontologies and Semantic Annotation

As to the fourth part of this volume on legal ontologies and semantic annotation, it
comprises eight chapters. In Using Legal Ontologies with Rules for Legal Textual
Entailment, Biralatei James Fawei, Adam Wyner, Martin Kollingbaum and Jeff Z. Pan
describe an initial attempt to model and implement the automatic application of legal
knowledge using a rule-based approach. Whilst an NLP tool extracts information to
instantiate an ontology relative to concepts and relations, ontological elements are
associated with legal rules written in SWRL to draw inferences to an exam question.
Although further development of the methodology and identification of key issues
require future analysis, the preliminary results on a small sample are promising. In
KR4IPLaw Judgment Miner- Case-Law Mining for Legal Norm Annotation, Sha-
shishekar Ramakrishna, Łukasz Górski and Adrian Paschke offer a proof-of-concept
implementation for automatizing the process of identifying the most relevant judgments
pertaining to a legal field and further transforming them into a formal representation
format. On this basis, the annotated legal section and its related judgments can be
mapped into a decision model for further down the line processing. In Conceptual
Annotation of Legal Documents with Ontology Concepts, Kolawole John Adebayo,
Luigi Di Caro and Guido Boella illustrate a novel task of semantic labelling, which
exploits ontology in providing a fine-grained conceptual document segmentation and
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 15

annotation. By dividing documents into semantically-coherent blocks and performing

conceptual tagging for efficient information filtering, the chapter presents a promising
solution, since the proposed task has several applications, such as granular information
filtering of legal texts, text summarization and information extraction among others. In
Reuse and Reengineering of Non-ontological Resources in the Legal Domain, Cristiana
Santos, Pompeu Casanovas, Víctor Rodríguez Doncel and Leon van der Torre explain
the processes of reusing and reengineering available non-ontological resources in the
legal domain into ontologies, taking into account their specificities, as they are highly
heterogeneous in their data model and contents. The description of these processes is
further clarified by using a case-study in the consumer law domain. In Ontology
Modelling for Criminal Law, Chiseung Soh, Seungtak Lim, Kihyun Hong and Young-
Yik Rhim propose a general methodology for designing criminal law ontologies and
rules. The chapter illustrates the super-domain ontology that contains the common
characteristics of criminal law, in order to explain the rule design method of criminal
law and present the application of the anti-graft act in Korea as an example. In Con-
trattiPubblici.org, a Semantic Knowledge Graph on Public Procurement Information,
Giuseppe Futia, Federico Morando, Alessio Melandri, Lorenzo Canova and Francesco
Ruggiero present the ContrattiPubblici.org project, which aims to develop a semantic
knowledge graph based on linked data principles in order to overcome the fragmen-
tation of existent datasets, to allow easy analysis, and to enable the reuse of infor-
mation. The objectives are to increase public awareness about public spending, to
improve transparency on the public procurement chain, and to help companies
retrieving useful knowledge for their business activities. In Application of Ontology
Modularization for Building a Criminal Domain Ontology, Mirna El Ghosh and Habib
Abdulrab carry out a survey on ontology modularization and present a modular
approach to build a criminal modular domain ontology (CriMOnto) for modelling the
legal norms of the Lebanese criminal system. CriMOnto, which will also be used for a
legal reasoning system, is composed of four independent modules that will be com-
bined together to compose the whole ontology. In A Linked Data Terminology for
Copyright Based on Ontolex-Lemon, Víctor Rodriguez-Doncel, Cristiana Santos,
Pompeu Casanovas, Asunción Gómez Pérez and Jorge Gracia discuss Ontolex-lemon,
namely, the de facto standard to represent lexica relative to ontologies that can be used
to encode term banks as RDF. A multi-lingual, multi-jurisdictional term bank of
copyright-related concepts has been published as linked data based on the ontolex-
lemon model. The terminology links information from several sources and the terms
have been hierarchically arranged, spanning multiple languages and targeting different
jurisdictions. Whilst the term bank has been published as a TBX dump file and is
publicly accessible as linked data, it has been used to annotate common licenses in the
RDFLicense dataset.

5.5 Legal Argumentation

Five chapters compose the fifth part of this volume on legal argumentation. In Abstract
Agent Argumentation (Triple-A), Ryuta Arisaka, Ken Satoh and Leon van der Torre
introduce a Dung style theory of abstract argumentation, which they call triple-A, in
which each agent decides autonomously whether to accept or reject her own arguments.
16 U. Pagallo et al.

By distinguishing between trusted arguments, selfish agents, and social agents, the
extensions of globally accepted arguments are defined using a game theoretic equi-
librium definition. In A Machine Learning Approach to Argument Mining in Legal
Documents, Prakash Poudyal analyzes and evaluates the natural language arguments
present in the European Court of Human Right (ECHR) Corpus. By dividing the
research into four modules, work on argumentative sentences vs. non-argumentative
sentences in narrative legal texts, is accomplished, so as to flesh out the features of this
module and conduct an experiment in Sequential Optimization Algorithm and Random
Forest Algorithm, which can be used as the basis of a general argument mining
framework. In Answering Complex Queries on Legal Networks: a Direct and a
Structured IR Approaches, Nada Mimouni, Adeline Nazarenko and Sylvie Salotti
compare two methods of search in legal collection networks, so as to present new
functionalities of search and browsing. Relying on a structured representation of the
collection graph, the first approach allows for approximate answers and knowledge
discovery, whilst the second one supports richer semantics and scalability but offers
fewer search functionalities. As a result, the chapter indicates how those approaches
could be combined to get the best of both. In Inducing Predictive Models for Decision
Support in Administrative Adjudicati, Karl Branting, Alexander Yeh and Brandy Weiss
explore the hypothesis that predictive models induced from previous administrative
decisions can improve subsequent decision-making processes. In light of three different
datasets, three different approaches for prediction in their domains were tested, showing
that each approach was capable of predicting outcomes. By exploring several
approaches that use predictive models to identify salient phrases in the predictive texts,
the chapter proposes a design for incorporating this information into a decision-support
tool. In Arguments on the Correct Interpretation of Sources of Law, Robert van
Doesburg and Tom van Engers deal with the formalization of legal reasoning and the
representation of law through computational models of argumentation. Whereas most
examples presented in literature can be characterized as post-hoc theory construction,
the chapter aims to provide an instrument that can be used to inform legal experts on
relevant issues in the process of solving current cases, i.e. using the interpretations of
legal sources ex-ante. An actual case that is in discussion in the Dutch Tax Admin-
istration, in court as well as in Parliament, helps to further clarify this approach.

5.6 Courts, Adjudication and Dispute Resolution

The sixth part of the volume is devoted to courts, adjudication and dispute resolution,
which are the subject matter of five chapters. In Dynamics of the Judicial Process by
Defeater Activation, Martin Moguillansky and Guillermo Simari illustrate a novel
activating approach to Argument Theory Change (ATC) for the study of the dynamics
of the judicial process. By considering the sentences of two different real criminal
procedures, the aim is to contribute to the discussion of how to deal with circumstances
of the judicial process like hypothetical reasoning for conducting investigations of a
legal case, and for handling the dynamics of the judicial process. In Claim Detection in
Judgments of the EU Court of Justice, Marco Lippi, Francesca Lagioia, Giuseppe
Contissa, Giovanni Sartor and Paolo Torroni address recent approaches to argumen-
tation mining in juridical documents, so as to present two distinct contributions. The
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 17

first one is a novel annotated corpus for argumentation mining in the legal domain,
together with a set of annotation guidelines. The second one is the empirical evaluation
of a recent machine learning method for claim detection in judgments. Whereas the
latter method has been applied to context-independent claim detection in other genres
such as Wikipedia articles and essays, the chapter shows that this method also provides
a useful instrument in the legal domain, especially when used in combination with
domain-specific information. In A Non-intrusive Approach to Measuring Trust in
Opponents in a Negotiation Scenario, Marco Gomes, John Zeleznikow and Paulo
Novais propose a threefold approach to trust, that regards the possibility of measuring
trust based on quantifiable behaviour, the use of Ambient Intelligence techniques that
use a trust data model to collect and evaluate relevant information based on the
assumption that observable trust between two entities (parties) results in certain typical
behaviours and, finally, relational aspects of trust and parties’ conflict styles based on
cooperativeness and assertiveness. The main contribution of this chapter is the iden-
tification of situations in which trust relationships influence the negotiation perfor-
mance. In Network, Visualization, Analytics. A Tool Allowing Legal Scholars to
Experimentally Investigate EU Case Law, Nicola Lettieri, Sebastiano Faro, Delfina
Malandrino, Margherita Vestoso and Armando Faggiano dwell on the intersection
between Network Analysis (NA), visualization techniques and legal science research
questions. Their aim is to bring the network approach into “genuinely legal” research
questions, and to create tools that allow legal scholars with no technical skills to make
experiments with NA and push new ideas both in legal and NA science, so as to use
NA and visualization in their daily activities. In Electronic Evidence Semantic Struc-
ture: Exchanging Evidence across Europe in a Coherent and Consistent Way, Maria
Angela Biasiotti and Fabrizio Turchi provide for a seminal work on a common and
shared understanding of what Electronic Evidence is and how it should be treated in the
EU context and in the EU member states. The chapter develops a tailor-made cate-
gorization of relevant concepts which provides a starting analysis for the exchange of
Electronic Evidence and data between judicial actors and LEAs, with a specific focus
on issues of the criminal field and criminal procedures. This semantic structure might
represent a good starting point for the alignment of electronic evidence concepts all
over Europe in a cross border dimension.

Acknowledgments. Law and Policy Program of the Australian government funded Data to
Decisions Cooperative Research Centre (http://www.d2dcrc.com.au/); Meta-Rule of Law
DER2016-78108-P, Research of Excellence, Spain. This work was also partially supported by
the European Union’s Horizon 2020 research and innovation programme under the Marie
Skłodowska-Curie grant agreement No 690974 “MIREL: MIning and REasoning with Legal
texts”.
18 U. Pagallo et al.

References
1. Regalado, A.: Google’s AI explosion in one chart. MIT technology review, 25 March 2017.
https://www.technologyreview.com/s/603984/googles-ai-explosion-in-one-chart/
2. Pagallo, U., Massimo Durante, M., Monteleone, S.: What is new with the internet of things
in privacy and data protection? Four legal challenges on sharing and control in IoT. In:
Leenes, R., van Brakel, R., Gutwirth, S., de Hert, P. (eds.) Data Protection and Privacy: (In)
visibilities and Infrastructures. LGTS, pp. 59–78. Springer, Dordrecht (2017)
3. Susskind, R., Susskind, D.: The Future of the Professions. Oxford University Press, Oxford
(2015)
4. Abelson, H., et al.: Keys under the doormat: mandating insecurity by requiring government
access to all data and communications. MIT Computer Science and AI Laboratory Technical
report, 6 July 2015
5. Abelson, H., et al.: The future of life institute, an open letter: research priorities for robust
and beneficial artificial intelligence (2015). http://futureoflife.org/ai-open-letter/. Accessed
18 Oct 2016
6. IEEE Standards Association: The Global Initiative for Ethical Considerations in the Design
of Autonomous Systems, Forthcoming (2017)
7. Hart, H.L.A.: The Concept of Law. Clarendon, Oxford (1961)
8. Leenes, R., Lucivero, F.: Laws on robots, laws by robots, laws in robots: regulating robot
behaviour by design. Law Innov. Technol. 6(2), 193–220 (2014)
9. Garfinkel, S., Spafford, G.: Web Security and Commerce. O’Reilly, Sebastopol (1997)
10. Pagallo, U.: The legal challenges of big data: putting secondary rules first in the field of EU
data protection. Eur. Data Protect. Law Rev. 3(1), 34–46 (2017)
11. Pagallo, U.: LegalAIze: tackling the normative challenges of artificial intelligence and
robotics through the secondary rules of law. In: Corrales, M., Fenwick, M., Forgó, N. (eds.)
New Technology, Big Data and the Law. Perspectives in Law, Business and Innovation.
PLBI, pp. 281–300. Springer, Singapore (2017)
12. Weng, Y.H., Sugahara, Y., Hashimoto, K., Takanishi, A.: Intersection of “Tokku” special
zone, robots, and the law: a case study on legal impacts to humanoid robots. Int. J. Soc.
Robot. 7(5), 841–857 (2015)
13. Pagallo, U., Durante, M.: The pros and cons of legal automation and its governance. Eur.
J. Risk Regul. 7(2), 323–334 (2016)
14. Brody, P., Pureswaran, V.. Device democracy: saving the future of the internet of things.
IBM, September 2014
15. Boella, G., Van Der Torre, L., Verhagen, H.: Ten challenges for normative multiagent
systems. In: Dagstuhl Seminar Proceedings. Schloss Dagstuhl-Leibniz-Zentrum für Infor-
matik (2008)
16. Hansen, J., Pigozzi, G., Van Der Torre, L.: Ten philosophical problems in deontic logic. In:
Dagstuhl Seminar Proceedings. Schloss Dagstuhl-Leibniz-Zentrum für Informatik (2007)
17. Casanovas, P., Palmirani, M., Peroni, S., van Engers, T., Vitali, F.: Semantic web for the
legal domain: the next step. Semant. Web 7(3), 213–227 (2016)
18. Andresen, G.: Nexxus Whitepaper (2017). http://nexxusuniversity.com
19. Jentzsch, C.: Decentralized autonomous organization to automate governance (2016). https://
download.slock.it/public/DAO/WhitePaper.pdf. Asseced 23 June 2016
20. Atzei, N., Bartoletti, M., Cimoli, T.: A survey of attacks on ethereum smart contracts (SoK).
In: Maffei, M., Ryan, M. (eds.) POST 2017. LNCS, vol. 10204, pp. 164–186. Springer,
Heidelberg (2017). https://doi.org/10.1007/978-3-662-54455-6_8
21. Xu, J.J.: Are blockchains immune to all malicious attacks? Financ. Innov. 2(1), 25 (2016)
 Introduction: Legal and Ethical Dimensions of AI, NorMAS, and the Web of Data 19

22. Arruñada, B.: Property as sequential exchange: the forgotten limits of private contract.
J. Inst. Econ. 13(4), 753–783 (2017)
23. DuPont, Q., Maurer, B.: Ledgers and Law in the Blockchain. Kings Review, 23 June 2015.
http://kingsreview.co.uk/magazine/blog/2015/06/23/ledgers-and-law-in-the-blockchain
24. Davidson, S., De Filippi, P., Potts, J.: Economics of blockchain. In: Proceedings of Public
Choice Conference Public Choice, Conference, May 2016, Fort Lauderdale, United States
(2016). https://doi.org/10.2139/ssrn.2744751. HAL Id. hal-01382002
25. English, M., Auer, S., Domingue, J.: Block chain technologies and the semantic web: a
framework for symbiotic development. In: Lehmann, J., Thakkar, H., Halilaj, L., Asmat, R.
(eds.) Computer Science Conference for University of Bonn Students, pp. 47–61 (2016)
26. Rodríguez-Doncel, V., Santos, C., Casanovas, P., Gómez-Pérez, A.: Legal aspects of linked
data—the European framework. Comput. Law Secur. Rev. 32(6), 799–813 (2016)
27. Casanovas, P., Mendelson, D., Poblet, M.: A linked democracy approach for regulating
public health data. Health Technol. 7(4), 519–537 (2017)
28. Poblet, M., Casanovas, P., Plaza, E. (ed.) Linked Democracy: Artificial Intelligence for
Democratic Innovation. http://ceur-ws.org/Vol-1897/
29. Poblet, M., Casanovas, P., Rodríguez-Doncel, V.: Linked Democracy. Springer Briefs,
Cham: Springer Nature (2018, forthcoming)
30. Sartor, G., Palmirani, M., Francesconi, E., Biasiotti, M.A.: Legislative XML for the
Semantic Web: Principles, Models, standards for Document Management. Springer, New
York (2011)
31. Giovannini, M.P., Palmirani, M., Francesconi, E.: Linee guida per la marcatura dei
documenti normativi secondo gli standard Normainrete, p. 200. European Press Academic
Publishing, Firenze (2012)
32. Boer, A., de Maat, E., Francesconi, E., Lupo, C., Palmirani M.,, Winkels R.: General XML
Format(s) for Legal Sources, Deliverable 3.1, Estrella Project - European project for
Standardized Transparent Representations in order to Extend LegaL Accessibility,
Proposal/Contract no.: 027655, European Commission (2007)
33. Breuker, J.A.P.J., et al.: OWL ontology of basic legal concepts (LKIF-Core). Estrella:
Deliverable 1.4., AMSTERDAM, UVA 2007, pp. 138 (2007)
34. Gordon, Thomas F.: Constructing legal arguments with rules in the legal knowledge
interchange format (LKIF). In: Casanovas, P., Sartor, G., Casellas, N., Rubino, R. (eds.)
Computable Models of the Law. LNCS (LNAI), vol. 4884, pp. 162–184. Springer,
Heidelberg (2008). https://doi.org/10.1007/978-3-540-85569-9_11
35. Vitali F., Palmirani M.: Akoma Ntoso Release Notes. Accessed 5 April 2018
36. Opijnen, M., Palmirani, M., Vitali, F., Agnoloni, T.: Towards ECLI 2.0. In: 2017
International Conference for E-Democracy and Open Government, Los Alamitos, CA, IEEE,
2017, P6082, pp. 1–9 (2017)
37. ELI Task Force: Technical ELI implementation guide citations, Publications Office of the
European Union, Luxembourg (2015). https://doi.org/10.2830/74251
38. Athan, T., Governatori, G., Palmirani, M., Paschke, A., Wyner, A.: LegalRuleML: design
principles and foundations. In: Faber, W., Paschke, A. (eds.) Reasoning Web 2015. LNCS, vol.
9203, pp. 151–188. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-21768-0_6
39. Peroni, S., Palmirani, M., Vitali, F.: UNDO: the united nations system document ontology.
In: d’Amato, C., Fernandez, M., Tamma, V., Lecue, F., Cudré-Mauroux, P., Sequeda, J.,
Lange, C., Heflin, J. (eds.) ISWC 2017. LNCS, vol. 10588, pp. 175–183. Springer, Cham
(2017). https://doi.org/10.1007/978-3-319-68204-4_18
20 U. Pagallo et al.

40. Barabucci, G., Cervone, L., Palmirani, M., Peroni, S., Vitali, F.: Multi-layer markup and
ontological structures in Akoma Ntoso. In: Casanovas, P., Pagallo, U., Sartor, G., Ajani, G.
(eds.) AICOL - 2009. LNCS (LNAI), vol. 6237, pp. 133–149. Springer, Heidelberg (2010).
https://doi.org/10.1007/978-3-642-16524-5_9
41. Casalicchio, E., Cardellini, V., Interino, G., Palmirani, M.: Research challenges in legal-rule
and QoS-aware cloud service brokerage. Future Gen. Comput. Syst. 78, 211–223 (2016)
42. Governatori, G., Hashmi, M., Lam, H.-P., Villata, S., Palmirani, M.: Semantic business
process regulatory compliance checking using LegalRuleML. In: Blomqvist, E., Ciancarini,
P., Poggi, F., Vitali, F. (eds.) EKAW 2016. LNCS (LNAI), vol. 10024, pp. 746–761.
Springer, Cham (2016). https://doi.org/10.1007/978-3-319-49004-5_48
43. Ashley, K.D.: Artificial Intelligence and Legal Analytics: New Tools for Law Practice in the
Digital Age, p. 446. Cambridge University Press, Cambridge (2017)
44. Palmirani, M., Vitali, F.: Legislative drafting systems. In: Usability in Government Systems.
Morgan Kaufmann, New York, pp. 133–151 (2012)
45. Rossi, A., Palmirani, M.: A visualization approach for adaptive consent in the European data
protection framework. In: 2017 International Conference for E-Democracy and Open
Government, Los Alamitos, CA, IEEE, 2017, pp. 1–12 (2017)
46. Haapio, H., Hagan, M., Palmirani, M., Rossi, A.,: Legal design patterns for privacy. In: Data
Protection/LegalTech Proceedings of the 21st International Legal Informatics Sympo-
sium IRIS 2018, II, pp. 445–450. Editions Weblaw, Bern (2018)
Legal Philosophy, Conceptual Analysis,
and Epistemic Approaches
 RoboPrivacy and the Law
as “Meta-Technology”

Ugo Pagallo(&)

Law School, University of Torino, Lungo Dora Siena 100 A, 10153 Turin, Italy
ugo.pagallo@unito.it

Abstract. The paper examines how a particular class of robotic applications,

i.e. service robots, or consumer robots, may affect current legal frameworks of
privacy and data protection. More particularly, the focus is on (i) a new
expectation of privacy brought about by these robotic applications; (ii) the
realignment of the traditional distinction between data processors and data
controllers; and, (iii) a novel set of challenges to the principle of privacy by
design. Instead of a one-way movement of social evolution from technology to
law, however, a key component of the analysis concerns the aim of the law to
govern technological innovation as well as human and artificial behaviour
through the regulatory tools of technology. Since most domestic and service
robots are not a sort of “out-of-the-box” machine and moreover, their behaviour
and decisions can be unpredictable and risky, special attention is drawn to the
experiment of the Japanese government that has worked out a way to address
(some of) the legal issues, which are at stake in this paper, through the creation
of special zones for robotics empirical testing and development, namely a form
of living lab, or Tokku. Interestingly, some EU member states have already
followed suit.

Keywords: Data protection
Design
General theory of law

Level of abstraction
Philosophy of technology
Privacy
Robot
Special zone
Techno-regulation

1 Introduction

The level of abstraction (LoA) of this paper has to do with the impact of robotics
technology in the fields of privacy and data protection, vis-à-vis the representation of
the law as a “meta-technology.” From a methodological viewpoint, a LoA sets the
proper level of analysis as a sort of interface that defines the features representing the
observables and variables of the research, the result of which provides a model for the
field under exam [5, 19]. This methodological approach can be illustrated with a figure
on the interface of the model, its observables and variables (see Fig. 1).
The next step is to determine the LoA concerning the law as a meta-technology,
namely the aim of the law to govern the process of technological innovation, e.g., in the
field of robotics. This stance corresponds to a glorious philosophical tradition, at least
from Kant to Kelsen, according to which the law can conveniently be understood as a
technique. In the phrasing of the General Theory of the Law and the State [9, at 26],

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 23–38, 2018.
https://doi.org/10.1007/978-3-030-00178-0_2
24 U. Pagallo

Fig. 1. Levels of abstraction of a model.

“what distinguishes the legal order from all other social orders is the fact that it
regulates human behaviour by means of a specific technique,” that hinges on the threat
of physical coercion: “if A, then B.” Therefore, once such technique regulates other
techniques and moreover, the process of technological innovation, we may accordingly
conceive the law as a meta-technology.
From this LoA, however, it does not follow that we have to buy any of Kelsen’s
ontological commitments: the stance this paper adopts on the law as meta-technology
does not imply either that the law is merely a means of social control, or that no other
meta-technological mechanisms exist. Rather, by insisting on the intent of the law to
govern the process of technological innovation, we should recall that the latter, pace
Kelsen, is affecting pillars of current legal systems. This has been, after all, a fil rouge
of the AICOL series throughout the past years (2009–2017). Contrary to previous
human societies that have used information and communication technology (“ICT”),
but have been mainly dependent on technologies that revolve around energy and basic
resources, today’s societies are increasingly dependent on ICT and furthermore, on
information as a vital resource [6, 20]. The processing of well formed and meaningful
data is reshaping essential functions of current societies, such as governmental services,
transportation and communication systems, business processes, or energy production
and distribution networks, up to our understanding about the world and about our-
selves. In a nutshell, after the revolutions of Copernicus, Darwin, and Freud, we are
dealing with the “fourth revolution” [6].
The observables of the analysis set up by the LoA on the interplay between law and
technology, are illustrated with a new figure (Fig. 2). The latter sheds light on both the
impact of e.g. robotic technology on the law and the aforementioned intent the law has
to govern the race of technological innovation, by further distinguishing the purposes
that law-making can have and the aim of legal systems to regulate human and artificial
behaviour.
The first observable of Fig. 2 suggests that the focus should be on how the fourth
revolution is affecting the tenets of the law. In addition to transforming the approach of
experts to legal information, e.g. the development of such fields as AI and the law,
 RoboPrivacy and the Law as “Meta-Technology” 25

Fig. 2. Law as a meta-technology in context.

technology has induced new types of lawsuits, or modified existing ones. Consider new
offences such as computer crimes (e.g. identity theft) that would be unconceivable once
deprived of the technology upon which they depend. Moreover, reflect on traditional
rights such as copyright and privacy, both turned into a matter of access to, and control
and protection over, information in digital environments. By examining the legal
challenges of technology, we thus have to specify those concepts and principles of legal
reasoning that are at stake. Then, we can begin to determine whether the information
revolution: (a) affects such concepts and principles; (b) creates new principles and
concepts; or, (c) does not concern them at all, the latter being the view of traditional
legal scholars [16].
The second observable of Fig. 2 has to do with the old, Kelsenian account of the
law as a social technique of a coercive order enforced through the threat of physical
sanctions: “if A, then B.” As previously stressed, we can grasp the law as a form of
meta-technology without buying any of Kelsen’s ontological commitments. Rather, we
should pay attention to the impact of technology on the formalisms of the law, much as
how legal systems deal with the process of technological innovation, through such a
complex network of concepts, as agency, accountability, liability, burdens of proofs,
clauses of immunity, or unjust damages. In this latter case, the aim of the law to govern
the field of technological innovation comprises several different techniques, so as to
attain: (a) particular effects; (b) functional equivalence between online and offline
activities; (c) non-discrimination between technologies with equivalent effects; and,
(d) future-proofing of the law that should neither hinder the advance of technology, nor
require over-frequent revision to tackle such a progress [11].
The third observable of Fig. 2 regards both the traditional aim of the law to regulate
human behaviour, and the field of techno-regulation, or legal regulation by design.
Since current advancements of technology have increasingly obliged legislators and
policy makers to forge more sophisticated ways to think about legal enforcement, the
focus is here on how national and international lawmakers have complemented the
traditional hard tools of the law through the mechanisms of design, codes, and IT
architectures. Although some of these architectural measures are not necessarily digital,
e.g. the installation of speed bumps in roads as a means to reduce the velocity of cars,
many impasses of today’s legal and political systems are progressively tackled, by
embedding normative constraints and constitutional safeguards into ICTs [20].
Against this backdrop, the paper adopts a further stance: the legal impact of
domestic, or service, robots is illustrated in accordance with the different fields with
which we are dealing, i.e. privacy and data protection, vis-à-vis the intent of the law to
govern the process of robotic innovation. This perspective partially overlaps with the
previous observables of Fig. 2 and yet, it allows us to pinpoint the new observables and
26 U. Pagallo

variables of the analysis, i.e. what issues ought to be questioned, prioritized and made
relevant, so as to stress the legal impact of domestic and/or service robots. Admittedly,
we could have chosen a different interface for our analysis, such as the legal issues of
liability and security, compulsory insurance and copyright, consumer law and envi-
ronmental regulation, that robots brought about by in the fields of criminal law, civil
law, administrative law, etc. However, matters of roboprivacy and data protection
clarify how the traditional aim of the law to govern innovation and to regulate human
behaviour is crucially changing nowadays. A new figure introduces the LoA of this
paper, with its observables and variables (Fig. 3).
Next Section introduces the technology under scrutiny, namely robotics and the
next generation of consumer and service robot applications. Then, in Sects. 3 and 4
respectively, the analysis dwells on the first two variants of Fig. 3, namely a new
expectation of privacy triggered by this novel generation of robots, and the realignment
of the traditional distinction between data processors and data controllers. As to the
second observable of Fig. 3, Sect. 5 draws the attention to a novel set of challenges in
the field of techno-regulation and especially, in connection with the principle of privacy
by design. On this basis, the conclusions of the analysis will bring us back to the aim of
the law to govern the process of technological innovation and consequently, the dif-
ferent purposes that the law can have. Since robots are here to stay, it seems fair to
affirm that the aim of the law should be to wisely govern our mutual relationships.

2 The Charge of Consumer and Service Robots

Since the early 1960s and for more than three decades, the field of robotics mostly
appeared as an automobile industry-dependent sector. A crucial point in this process
occurred in the early 1980s, when Japanese industry first began to implement this
technology on a large scale in their factories, acquiring strategic competitiveness by
decreasing costs and increasing the quality of their products. Western car producers
learned a hard lesson and followed Japanese thinking, installing robots in their factories
a few years later. This trend expanded so much that, according to the UN World 2005
Robotics Report, the automotive industry in Europe still received around 60%–70% of
all new robot installations in the period 1997–2003 [24].
Yet, in the same years as covered by the UN World report, things began to rapidly
change. The traditional dependence of robotics on the automobile industry dramatically
opened up to diversification, first with water-surface and underwater unmanned vehi-
cles, or “UUVs,” used for remote exploration work and the repairs of pipelines, oil rigs
and so on, developing at an amazing pace since the mid-1990s. Ten years later,
unmanned aerial vehicles (“UAVs”), or systems (“UAS”), upset the military field [16].
Over the past decade, robots have spread in both the industrial and service fields.
Together with robots used in the manufacture of textiles and beverages, refining pet-
roleum products and nuclear fuel, producing electrical machinery and domestic
appliances, we also have a panoply of robot surgeons and robot servants, robot nannies
and robot scientists, and even divabots, e.g. the Japanese pop star robot singer HRP-4C.
The old idea of making machines (e.g. cars) through further machines (e.g. robots), has
thus been joined – and increasingly replaced – by the aim to build fully autonomous
 RoboPrivacy and the Law as “Meta-Technology” 27

Fig. 3. The interface of this paper.

robots, namely AI machines that “sense,” “think,” and “act,” in the engineering
meaning of these words [1, 25]. In a nutshell, this means that robots can respond to
stimuli by changing the values of their properties or inner states and, furthermore, they
can improve the rules through which those properties change without external stimuli.
As a result, we are progressively dealing with agents, rather than simple tools of human
interaction [19].
The more robots become interactive, self-sufficient, and adaptable, however, the
more attention should be drawn to the normative challenges of this technology, i.e. the
reasons why such AI machines should, or should not, be deployed at our homes, in the
market, or on the battlefield. Consider current debate on whether lethal force can be
fully automated, or whether the intent to create robots that people bond with is ethically
justifiable. The complexity of the subject-matter, i.e. the normative challenges of
robotics, can be grasped with a new figure (Fig. 4).
Leaving aside the aim of the moral, political and economic fields, in governing the
process of technological innovation, what matters here is the normative side of the law.
As mentioned above in the introduction, the challenges of robotics may concern pillars
of current international law, criminal law, civil law, both in contracts and tort law,
administrative law, and so forth. Think of tiny robotic helicopters employed in a
jewellery heist vis-à-vis other robotic applications trading in auction markets and how
the random-bidding strategy of these apps clarifies, or even has provoked, real life
bubbles and crisis, e.g. the financial troubles of late 2009 that may have been triggered
by the involvement of such artificial agents. Since the “bad nature” of robots vary in
accordance with the field under examination, we thus have to pinpoint the specific legal
challenges of this technology. A thematic, rather than field-dependent, LoA seems
fruitful, in order to flesh out some of the most urgent and trickier challenges of robotics.
The last variable of Fig. 4 brings us back to the observables and variables of Fig. 3: the
normative challenges of robotics are here grasped in connection with a particular class
of robotic applications, i.e. “service robots” [24], or “consumer robots” [4], in order to
determine whether, and to what extent, they may affect current legal frameworks of
privacy and data protection. In the wording of the EU Agenda, “these robots will be
bought or leased and used to provide services to individuals. They will be operated by,
or interact with, untrained, or minimally trained people in everyday environments.
Applications range from helping the elderly stay safely mobile in their own homes to
the automation of everyday household chores and the provision of remote monitoring
and security for home” [4, at 34].
In accordance with the LoA of this paper on the law as a meta-technology, the
analysis follows a twofold approach. On the one hand, the attention should be drawn to
the aim of the law to govern the process of technological innovation: this intent has to
28 U. Pagallo

Fig. 4. The complexity of the field.

do, in this context, with the regulation of producers and designers of robots through
specific sets of norms, or the regulation of user behaviour through the design of their
robots. On the other hand, over the past years, scholars have increasingly stressed the
many issues fated to remain open with the protection of people’s privacy and the
transparency with which service, or consumer, robots will increasingly collect, process,
and make use of personal data [13, 19, 21, 23]. Remarkably, in The Right to Privacy
(1890), Samuel Warren and Louis Brandeis claimed that “instantaneous photographs
and newspaper enterprise have invaded the sacred precincts of private and domestic
life; and numerous mechanical devices threaten to make good the prediction that ‘what
is whispered in the closet shall be proclaimed from the house-tops’” [26, at 195]. By
taking into account current trends of robotics, should we expect a new invasion of the
sacred precincts of our private life?

3 Private Expectations

Many readers of this paper may not have met a consumer robot so far and yet, they are
familiar with some privacy challenges that will be brought about by these applications.
Reflect on how a number of mobile devices, such as your smartphone, collect a myriad
of different data, e.g. images and video through cameras, motion and activities through
gyroscopes and accelerometers, fingerprints through biometric sensors, geo-location
data through GPS techniques, and so forth. Likewise, consider such fitness applica-
tions, as Nike+ or Adidas miCoach, that track route, pace and time activities of users
through GPS and sensors. Moreover, contemplate the real time facial recognition app
NameTage for Google Glasses. Risks for user’s informational privacy have been
stressed time and again: for instance, as to the threats raised by sensors, consider how
personal information can be inferred from such data, as occurs with information on
mobility patterns, activity and face recognition, health information, and so on. In light
of current risks for user’s privacy, what is new about consumer robots is that sensors,
cameras, GPS, facial recognition apps, Wi-Fi, microphones and more, will be
 RoboPrivacy and the Law as “Meta-Technology” 29

assembled in a single piece of high-tech that will likely affect what US legal scholars
dub as a “reasonable expectation of privacy.”
In a nutshell, the formula means that individuals have the right to be protected
against unreasonable searches and seizures under the Fourth Amendment. Pursuant to
the jurisprudence of the US Supreme Court from the 1967 Katz v. United States case
(389 U.S. 347), onwards, the overall idea is that the opinion of a person that a cer-
tain situation or location is private, must go hand in hand with the fact that society at
large would recognize this expectation, so as to protect the latter as a fundamental right.
Scholars and also justices of the Supreme Court, however, have emphasised that such
twofold dimension of this reasonable expectation, both social and individual, can entail
a vicious circle, much as “the chicken or the egg” causality dilemma. Moreover, the
right to a reasonable expectation of privacy rests on the assumption that both indi-
viduals and society have developed a stable set of privacy expectations, whereas
technology can dramatically change these very expectations. As Justice Alito empha-
sizes in his concurring opinion in United States v. Jones from 23 January 2012 (565 U.
S. __), “dramatic technological change may lead to periods in which popular expec-
tations are in flux and may ultimately produce significant changes in popular attitudes.”
The legal framework is different in Europe. According to the EU legal rules and
principles of privacy and data protection, the opinion of individuals does not play any
normative role, in order to determine the legitimacy of the acts and statutes laid down
by the public institutions. On the contrary, what individuals and society can reasonably
expect, is that public organizations, multinational corporations, and other private par-
ties, abide by the set of rules and principles established by the EU, or national, leg-
islators. Notwithstanding this approach, it does not follow that social and individual
expectations of privacy are totally irrelevant in Europe. Consider the proposal for a new
data protection regulation in the EU legal system, presented by the Commission in
January 2012. The same day in which the Parliament approved the new set of rules, the
Commission was keen to inform us with a press release on 12 March 2014, that the
intent to update and modernize the principles enshrined in the 1995 data protection
directive is strictly connected with “a clear need to close the growing rift between
individuals and the companies that process their data.”1 The source of this “clear need”
was provided by the Flash Eurobarometer 359 from June 2011, on the attitudes con-
cerning data protection and electronic identity in the EU. According to this source, 9
out of 10 Europeans (92%) said they are worried about mobile apps collecting their
data without their consent, 7 out of 10 are concerned about the potential use that
companies may make of the information disclosed, etc. Whether the new EU regulation
n. 679 from 2016, the so called GDPR, will close the rift between individual and
companies is, of course, an open issue and yet, it is highly likely that consumer and
service robots will add new worries about radars, sensors or laser scanners of artificial
agents collecting data of their human masters, much as companies that may infer
personal information from such data on mobility patterns, user’s preferences, lifestyles,
and the like.

1
See the press release at http://europa.eu/rapid/press-release_MEMO-14-186_it.htm.
30 U. Pagallo

A common expectation of privacy should thus be expected (not only, but also) in
Europe and US, in the basic sense that users of robots will likely assume that some
“degree of friction,” restraining the flow of personal information, should be respected.
Clearly, this is not to say that personal choices will have no role in determining
different levels of access to, and control over, information. Rather, from a legal point of
view, there are two ways in which we can appreciate the role of these personal choices
in keeping firm distinctions between individuals and society, agents and the system. On
the one hand, the different types of information which robots may properly reveal,
share, or transfer, will often hinge on personal preferences of the human master on
whether it is appropriate to trace back information to an individual, and how infor-
mation should be distributed according to different standards in different contexts.
Depending on how humans have taken care of their artificial agents, specimens of the
same model of robot will accordingly behave in different ways. On the other hand,
personal choices on both norms of appropriateness and flow will further hinge on the
type of robot under scrutiny. The type of information that makes sense to communicate
and share with an artificial personal assistant, would be irrelevant or unnecessary to
impart to a robot toy. It is thus likely that individuals will modulate different levels of
access to, and control over, information, depending on the kind of the artificial inter-
locutor, the context of their interaction, and the circumstances of the case.
The impact of domestic and service robots on current expectations of privacy,
however, not only regards problems of reliability, traceability, identifiability, trustful-
ness and generally speaking, how the interaction with such robots and their presence in
“the sacred precincts of private and domestic life” may realign both norms of appro-
priateness and of informational flow [15]. In addition, we should expect psychological
problems related to the interaction with robots as matters of attachment and feelings of
subordination, deviations in human emotions, etc. [25]. This scenario suggests that we
should go back to the general intent of the law to govern the process of technological
innovation through the four different categories stressed by Ronald Leenes and Fed-
erica Lucivero in Laws on Robots, Laws by Robots, Laws in Robots (2014). Accord-
ingly, the focus should be on (a) the regulation of human producers and designers of
robots through law, e.g. either through ISO standards or liability norms for users of
robots; (b) the regulation of user behaviour through the design of robots, that is, by
designing robots in such a way that unlawful actions of humans are not allowed; (c) the
regulation of the legal effects of robot behaviour through the norms set up by law-
makers, e.g. the effects of robotic contracts and negotiations; and, (d) the regulation of
robot behaviour through design, that is, by embedding normative constraints into the
design of the artificial agent. This differentiation can be complemented with further
work on the regulation of the environment of the human-robot interaction and the legal
challenges of “ambient law” [7, 8]. We should thus take into account the set of values,
principles, and norms that constitute the normative context in which the consequences
of such regulations have to be evaluated. In addition, we have to consider issues of data
protection that mostly revolve around the transparency with which personal data are
collected, processed, and used.
In the EU legal system, for example, individuals have the right to know the pur-
poses for which their data are processed, much as the right to access that data and to
have it rectified. In the wording of Article 8(2) of the EU Charter of fundamental rights,
 RoboPrivacy and the Law as “Meta-Technology” 31

“such data must be processed fairly… and on the basis of the consent of the person
concerned or some other legitimate basis laid down by law.” This type of protection
through the principles of minimization and quality of the data, its controllability and
confidentiality, may of course overlap with the protection of the individual privacy. In
such cases, the aim is to constraint the flow of information, and keep firm distinctions
between individuals and society, in order to protect what the German Constitutional
Court has framed in terms of “informational self-determination” since its Volkszäh-
lungs-Urteil (“census decision”) from 15 December 1983. Yet, there are several cases
in which the norms of data protection do not entail the safeguard of any privacy.
Together with the mechanism of “notice and consent,” laid down by Article 7 of the
EU directive 46 from 1995, reflect on how the processing of personal data can – and at
times should – go hand in hand with the strengthening of further rights and interests of
individuals, such as freedom of information and the right to knowledge, freedom of
expression and access to public documents, up to participatory democracy and the
functioning of the internal market with the free circulation of services and information
pursuant to the EU directive on the reuse of public sector information, i.e. D-
37/2013/EC [22].
As a result of this differentiation between privacy and data protection, how should
we strike a fair balance in the case of domestic and service robots?

4 Holders of Personal Data

The reference point for today’s state-of-art in roboprivacy is given by the guidelines
that a EU-sponsored project, namely “RoboLaw,” presented in September 2014.
According to this document, the principle of privacy by design can play a key role in
making and keeping robots data protection-compliant [23, at 19]. For example, some
legal safeguards, such as data security through encryption and data access control, can
be embedded into the software and interface of the robot. Likewise, “requirements such
as informed consent can be implemented in system design, for example through
interaction with users displays and input devices” (ibid). After all, this is what already
occurs with some operating systems, such as Android, that require user’s consent
whenever an application intends to access personal data. Furthermore, robots could be
designed in a privacy-friendly way, so that the amount of data to be collected and
processed is reduced to a minimum and in compliance with the finality principle. This
means that, pursuant to Article 6(1)(b) of the EU data protection directive 46 from 1995
and now, Article 5(1)(b) of the GDPR, robots should collect data only insofar as it is
necessary to achieve a specified and legitimate purpose.
In addition, this set of legal safeguards on data minimization, finality principle,
informed consent, etc., shall be pre-emptively checked through control mechanisms
and data protection impact assessments, so as to ensure that privacy safeguards are at
work even before a single bit of information has been collected. More particularly, in
the words of the RoboLaw Guidelines, “as a corollary of a privacy impact assessment,
a control mechanism should be established that checks whether technologies are
constructed in the most privacy-friendly way compatible with other requirements (such
as information needs and security)” [23, at 190]. Leaving aside specific security
32 U. Pagallo

measures for particular classes of service robots, such as health robots, personal care
robots, or automated cars examined by the EU project, the latter suggests that “the
adoption of updated security measures should not be considered only as a user’s
choice, but also as a specific legal duty. It is clear that the illicit treatment of the data is
unlikely to be considered a responsibility of the manufacturer of the robot, but rather a
liability of its user, who is the ‘holder’ of the personal data” [23, at 190].
Whether the end-user, or “human master,” of the robot should be deemed as the
data controller and hence, liable for any illicit treatment of personal data, is however
debatable. As stressed above in the previous sections, we may admit cases in which the
role of personal choices suggests that end-users should be conceived as data controllers
and thus, liable for how their artificial agents collect, process, and make use of personal
data. But, as occurs today with issues of internet connectivity, or sensors and mobile
computing applications, several other cases indicate that the illicit treatment of personal
data may depend on designers and manufacturers of robots, internet providers, appli-
cations developers, and so forth. After all, the illicit treatment of personal data may be
traced back to the malfunctioning of the robot, or to HTTP headers in packets of
network traffic data that can be used to determine interests and other personal infor-
mation about the master of the robot, along with applications that leak identifiable data,
such as device ID, GPS, and more. What all these cases make clear is not only
hypotheses of illicit treatment of data that do not depend on end-users or masters of
robots as data controllers. Additionally, the liability of designers and manufacturers of
robots, internet providers, etc., can be problematic in connection with different inter-
pretations of current rules and principles of the data protection legal framework, e.g. the
EU 2016 norms on the protection of individuals with regard to the processing of
personal data and on the free movement of such data. As stressed by Art. 29 Working
Party in the opinion 1/2010 (WP 169), “the concept of controller is a functional
concept, intended to allocate responsibility where the factual influence is, and thus
based on a factual, rather than a formal analysis,” which “may sometimes require an in-
depth and lengthy investigation” (op. cit., 9).
However, even admitting the conclusions of the Working Party, so that liability of
data controllers “can be easily and clearly identified in most situations” (ibid.), we still
have to face a major problem. Although normative safeguards can be embedded into
the software and interface of domestic robots, significant differences between multiple
data protection jurisdictions, e.g. between US and EU, remain. Whereas, in the US,
privacy policies of the industry and the agreement between parties mostly regulate
matters of data protection in the private sector, we already stressed that the EU has
adopted a comprehensive legislation since its 1995 data protection directive, up to the
current provisions of the GDPR. Principles and rules of this legal framework on data
minimization, finality principle, informed consent, etc., set limits to the contractual
power of individuals and companies. This divergence between US and EU will likely
increase with the GDPR. Even the RoboLaw Guidelines concede that these “significant
differences… could make it difficult for manufacturers catering for the international
market to design in specific data protection rules” [23, at 19]. As a matter of legal fact,
which norms and rules should designers and manufactures of domestic robots embed
into their products? Should such norms and rules vary according to the specific market
 RoboPrivacy and the Law as “Meta-Technology” 33

(and jurisdiction)? Would this latter option be technically and economically

sustainable?
A feasible way-out is pragmatic. Following Anu Bradford’s thesis on “the Brussels
effect” and how Europe’s regulatory model wields unilateral influence across such legal
fields, as data protection, antitrust, or health and environmental legislation [2], we may
envisage a similar effect in the case of domestic robots. The non-divisibility of data and
the compliance costs of multinational corporations dealing with multiple regulatory
regimes, may prompt most robot manufacturers to adopt and adapt themselves to the
strictest international standards across the board, that is, the EU data protection
framework, much as occurred in the case of internet companies vis-à-vis data pro-
tection issues. However, compared with traditional privacy regulation, we should not
overlook some peculiarities of domestic and service robots. By affecting what US
lawyers dub as a reasonable expectation of privacy, as explored above in the previous
section, it is highly likely that such expectation, both individual and social, will be “in
flux” for a while. Some insist on this flux to stress that lawmakers, rather than judges or
data protection authorities, are in the best position to determine the rules of the game
and guide social and individual behaviour [10]. But, even in light of the strictest
international standards of the EU legislation, it is still vague how we should interpret
some of its key assumptions, e.g. the principle of privacy by design. Neither the
Commission’s proposal for a new data protection regulation in January 2012, nor the
amendments of the EU Parliament in March 2014, nor the final 2016 text of the general
data protection regulation (GDPR), clarify how to design robots that abide by the law.
All in all, we lack a regulatory model that may represent a reference point for inter-
national standards on the design, production and commercialization of domestic robots.
The aim of next section is thus to deepen current uncertainties on the principle of
privacy by design, by fleshing out how such uncertainties are intertwined with the aim
of the law to govern the process of technological innovation.

5 How to Design Robots that Abide by the Law

Legal design has different and even opposite aims. Think about the latter according to a
spectrum: at one end, the purpose is to determine and control both social and individual
behaviour through the use of self-enforcing technologies and such automatic tech-
niques, as filtering systems and digital rights management (DRM)-tools, that intend to
restrict any form of access, use, copy, replacement, reproduction, etc., of informational
resources in the environment. At the other end of the spectrum, design may aim to
encourage the change of people’s behaviour by widening the range of the choices
through incentives based on trust (e.g. reputation mechanisms), or trade (e.g. services
in return). In between the ends of the spectrum, design may aim to decrease the impact
of harm-generating behaviour through security measures, default settings, user friendly
interfaces, and the like. Notwithstanding these different ends, it is noteworthy that
legislators and scholars alike often refer to the aim to embed legal constraints into
technology, e.g. privacy by design, in a neutral manner, that is, as if the intent of this
legal embedding could be impartial and value-free. Consider articles 23 and 30 of the
EU Commission’s proposal for a new data protection regulation, much as § 3.4.4.1 of
34 U. Pagallo

the document with which the Commission illustrated the proposal. Here, the formula of
“privacy by design” is so broad, or vague, that it can include whatever end design may
have. Although, in the amendment 118 of the EU Parliament, the latter refers to
“comprehensive procedural safeguards regarding the accuracy, confidentiality, integ-
rity, physical security and deletion of personal data,” it is still unclear vis-à-vis Article
25 of the GDPR whether the aim should be to decrease the impact of harm-generating
conducts or rather, to widen the range of individual options, or both. In light of these
uncertainties, how about the design of consumer and service robots and the environ-
ment of human-robot interaction through sensors, GPS, facial recognition apps, Wi-Fi,
RFID, NFC, or QC code-based environment interaction?
First of all, the principle of privacy by design and the EU Parliament’s “compre-
hensive procedural safeguards” can be grasped in terms of security measures, e.g. data
access control and encryption, much as user-friendly default configurations of robotic
interfaces. Robots can indeed be designed in such a way that values of design are
appropriate even for novice users and still, the robot improves efficiency. Furthermore,
the intent can be to seamlessly integrate robots into domestic workflows and IT systems
of smart houses via compliant motion control systems and situation awareness tech-
nologies, much as flexible and modular systems for the measurement of physical,
physiological and electro-physiological variables, that should make the user experience
an integral and even natural part of the process. In addition, we should take into
account the set of legal safeguards on data minimization, finality principle, or informed
consent, that were mentioned in the previous section, so as to tackle the convergence of
robotic data processing and the internet (of things, of everything, etc.).
However, a number of further cases suggest that domestic robots could alternatively
be designed with the aim to prevent any harm-generating behaviour from occurring.
This is not only a popular stance among Western lawmakers in such fields as intel-
lectual property (“IP”) protection, data retention, or online security [19]. Moreover, in
the field of robotics, two further reasons may reinforce this design policy. On the one
hand, in the phrasing of the EU Parliament, “the accuracy, confidentiality, integrity,
physical security and deletion of personal data,” processed by domestic robots, will
more often concern data of third parties. On the other hand, we must reflect on both the
psychological problems related to the very interactions with robots, and the case of
human masters that do not properly fulfill their role of caretakers. Lawmakers, data
protection authorities and courts may thus adopt a stricter version of the principle of
privacy by design, in order to preclude any data protection infringement through the
use of self-enforcing technologies, e.g. filtering systems, in the name of security rea-
sons. This scenario is not only compatible with the new EU regulation, but has been
endorsed by some popular versions of the principle. In Ann Cavoukian’s account of
privacy by design, for example, personal data should be automatically protected in
every IT system as its default position, so that a cradle-to-grave, start-to-finish, or end-
to-end lifecycle protection ensures that privacy safeguards are automatically at work
even before a single bit of information has been collected [3]. But, is this automatic
version of privacy by design technically feasible and even desirable?
There are several ethical, legal, and technical reasons why we should resist the aim
of some lawmakers to protect citizens even against themselves. First, the use of self-
enforcing technologies risks to curtail freedom and individual autonomy severely,
 RoboPrivacy and the Law as “Meta-Technology” 35

because people’s behaviour and their interaction with robots would be determined on
the basis of design rather than by individual choices [14, 28]. Once the normative side
of the law is transferred from the traditional “ought to” of rules and norms to what
actually is in automatic terms, a modeling of individual conduct follows as a result,
namely, that which Kant used to stigmatize as “paternalism” [17].
Second, specific design choices (not only, but also) in robotics may result in
conflicts between values and furthermore, conflicts between values may impact on the
features of design. Since both privacy and data protection may be conceived in terms of
human dignity or property rights, of contextual integrity or total control, it follows that
privacy by design acquires many different features. In the case of self-enforcing
technologies, their use would make conflicts between values even worse, due to
specific design choices, e.g. the opt-in vs. opt-out diatribe over the setting of infor-
mation systems [17].
Third, attention should be drawn to the technical difficulty of applying to a robot
concepts traditionally employed by lawyers, through the formalization of norms, rights,
or duties. As stressed by Bert-Jaap Koops and Ronald Leenes, “the idea of encoding
legal norms at the start of information processing systems is at odds with the dynamic
and fluid nature of many legal norms, which need a breathing space that is typically not
something that can be embedded in software” [12, at 7]. All in all, informational
protection safeguards present highly context-dependent notions that raise several rel-
evant problems when reducing the complexity of a legal system where concepts and
relations are subject to evolution [18].
At the end of the day, it should be clear that the use of self-enforcing technologies
would not only prevent robotic behaviour from occurring. By unilaterally determining
how the artificial agent should act when collecting, for example, the information they
need for human-robot interaction and task completion from networked repositories,
such design policies do impinge on individual rights and freedom. If there is no need to
humanize our robotic applications, we should not robotize human life either. The time
is ripe for the conclusions of this paper.

6 Conclusions

Most service and consumer robots are not a mere “out of the box” machine. Rather, as a
sort of prolonged epigenetic developmental process, robots increasingly gain knowl-
edge or skills from their own interaction with the living beings inhabiting the sur-
rounding environment, so that more complex cognitive structures emerge in the state-
transition system of the artificial agent. Simply put, specimens of the same model will
behave in quite different ways, according to how humans train, treat, or manage their
robots. Correspondingly, both the behaviour and decisions of these artificial agents can
be unpredictable and risky, thus affecting traditional tenets of the law, such as a
“reasonable expectation” of privacy, which was mentioned above in Sect. 3, together
with matters of data protection (Sect. 4), and the troubles with legal design (Sect. 5).
How, then, should legal systems proceed? What purposes should the law have? What
lessons did the LoA of this paper on the law as a meta-technology learn?
36 U. Pagallo

All in all, the conclusion is “pragmatic.” Over the past 15 years, the Japanese
government has worked out a way to address most of the issues examined in this paper,
through the creation of special zones for robotics empirical testing and development,
namely, a form of living lab, or Tokku. Whereas the world’s first special zone was
approved by the Cabinet Office in November 2003, covering the prefecture of Fukuoka
and the city of Kitakyushu, further special zones have been established in Osaka and
Gifu, Kanagawa and Tsukuba. The overall aim of these special zones is to set up a sort
of interface for robots and society, in which scientists and common people can test
whether robots fulfil their task specifications in ways that are acceptable and com-
fortable to humans, vis-à-vis the uncertainty of machine safety and legal liabilities that
concern, e.g., the protection for the processing of personal data through sensors, GPS,
facial recognition apps, Wi-Fi, RFID, NFC, or QC code-based environment interaction.
Although the Japanese typically are perceived as conservative and inclined to a for-
malistic and at times, pedantic interpretation of the law, it is remarkable that such
special zones are highly deregulated from a legal point of view. “Without deregulation,
the current overruled Japanese legal system will be a major obstacle to the realization of
its RT [Robot Tokku] business competitiveness as well as the new safety for human-
robot co-existence” [27]. Furthermore, the intent is “to cover many potential legal
disputes derived from the next-generation robots when they are deployed in the real
world” (ibid.).
So far, the legal issues addressed in the RT special zones regard road traffic laws (at
Fukuoka in 2003), radio law (Kansai 2005), privacy protection (Kyoto 2008), safety
governance and tax regulation (Tsukuba 2011), up to road traffic law in highways
(Sagami 2013). These experiments could obviously be extended, so as to strengthen
our understanding of how the future of the human-robot interaction could turn out.
Consider again some of the problems mentioned above in the previous sections, such as
the realignment of the traditional distinction between data processors and data con-
trollers, or the aim of the law to design robots that abide by the law. By testing these
scenarios in open, unstructured environments, the Japanese approach shows a prag-
matic way to tackle the challenges brought about by possible losses of control of AI
systems. Significantly, in the field of autonomous vehicles, several EU countries have
endorsed this kind of approach: Sweden has sponsored the world’s first large-scale
autonomous driving pilot project, in which self-driving cars use public roads in
everyday driving conditions; Germany has allowed a number of tests with various
levels of automation on highways, e.g. Audi’s tests with an autonomously driving car
on highway A9 between Ingolstadt and Nuremberg. Whereas the Japanese automotive
sector acquired a strategic competitiveness in the early 1980s through the use of robots,
the aim of Western producers and some of its lawmakers is to follow suit and prevent
on this basis another hard lesson.

References
1. Bekey, G.A.: Autonomous Robots: From Biological Inspiration to Implementation and
Control. The MIT Press, Cambridge (2005)
2. Bradford, A.: The Brussels effect. Northwest. Univ. Law Rev. 107(1), 1–68 (2012)
 RoboPrivacy and the Law as “Meta-Technology” 37

3. Cavoukian, A.: Privacy by design: the definitive workshop. Identity Inf. Soc. 3(2), 247–251
(2010)
4. EU Robotics: Robotics 2020 Strategic Research Agenda for Robotics in Europe, draft 0v42,
11 October 2013
5. Floridi, L.: The method of levels of abstraction. Mind. Mach. 18(3), 303–329 (2008)
6. Floridi, L.: The Fourth Revolution. Oxford University Press, Oxford (2014)
7. Hildebrandt, M.: Legal protection by design: objections and refutations. Legisprudence 5(2),
223–248 (2011)
8. Hildebrandt, M., Koops, B.-J.: The challenges of ambient law and legal protection in the
profiling era. Mod. Law Rev. 73(3), 428–460 (2010)
9. Kelsen, H.: General Theory of the Law and the State (trans: A. Wedberg). Harvard
University Press, Cambridge, Mass (1945/1949)
10. Kerr, O.: The fourth amendment and new technologies: constitutional myths and the case for
caution. Mich. Law Rev. 102, 801–888 (2004)
11. Koops, B.J.: Should ICT regulation be technology-neutral? In: Koops, B.J., et al. (eds.)
Starting Points for ICT Regulation: Deconstructing Prevalent Policy One-Liners, pp. 77–
108. TMC Asser, The Hague (2006)
12. Koops, B.-J., Leenes, R.: Privacy regulation cannot be hardcoded: a critical comment on the
“privacy by design” provision in data protection law. Int. Rev. Law Comput. Technol. 28,
159–171 (2014)
13. Leenes, R., Lucivero, F.: Laws on Robots, laws by robots, laws in robots: regulating robot
behaviour by design. Law Innov. Technol. 6(2), 193–220 (2014)
14. Lessig, L.: Free Culture: The Nature and Future of Creativity. Penguin Press, New York
(2004)
15. Nissenbaum, H.: Privacy as contextual integrity. Wash. Law Rev. 79(1), 119–158 (2004)
16. Pagallo, U.: Robots of just war: a legal perspective. Philos. Technol. 24(3), 307–323 (2011)
17. Pagallo, U.: On the principle of privacy by design and its limits: technology, ethics, and the
rule of law. In: Gutwirth, S., Leenes, R., De Hert, P., Poullet, Y. (eds.) European Data
Protection: In Good Health?, pp. 331–346. Springer, Dordrecht (2012). https://doi.org/10.
1007/978-94-007-2903-2_16
18. Pagallo, U.: Three roads to complexity, AI and the law of robots: on crimes, contracts, and
torts. In: Palmirani, M., Pagallo, U., Casanovas, P., Sartor, G. (eds.) AICOL 2011. LNCS
(LNAI), vol. 7639, pp. 48–60. Springer, Heidelberg (2012). https://doi.org/10.1007/978-3-
642-35731-2_3
19. Pagallo, U.: The Laws of Robots: Crimes, Contracts, and Torts. Springer, Dordrecht (2013).
https://doi.org/10.1007/978-94-007-6564-1
20. Pagallo, U.: Good onlife governance: on law, spontaneous orders, and design. In: Floridi, L.
(ed.) The Onlife Manifesto: Being Human in a Hyperconnected Era, pp. 161–177. Springer,
Dordrecht (2015). https://doi.org/10.1007/978-3-319-04093-6_18
21. Pagallo, U.: Teaching “consumer robots” respect for informational privacy: a legal stance on
HRI. In: Coleman, D. (ed.) Human-Robot Interactions. Principles, Technologies and
Challenges, pp. 35–55. Nova, New York (2015)
22. Pagallo, U., Bassi, E.: Open data protection: challenges, perspectives, and tools for the reuse
of PSI. In: Hildebrand, M., O’Hara, K., Waidner, M. (eds.) Digital Enlightenment Yearbook
2013, pp. 179–189. IOS Press, Amsterdam (2013)
23. RoboLaw: Guidelines on Regulating Robotics. EU Project on Regulating Emerging Robotic
Technologies in Europe: Robotics facing Law and Ethics, 22 September 2014
24. UN World Robotics: Statistics, Market Analysis, Forecasts, Case Studies and Profitability of
Robot Investment, edited by the UN Economic Commission for Europe and co-authored by
the International Federation of Robotics, UN Publication, Geneva (Switzerland) (2005)
38 U. Pagallo

25. Veruggio, G.: Euron roboethics roadmap. In: Proceedings Euron Roboethics Atelier, 27th
February–3rd March, Genoa, Italy (2006)
26. Warren, S., Brandeis, L.: The right to privacy. Harv. Law Rev. 14, 193–220 (1890)
27. Weng, Y.-H., Sugahara, Y., Hashimoto, K., Takanishi, A.: Intersection of “Tokku” special
zone, robots, and the law: a case study on legal impacts to humanoid robots. Int. J. Soc.
Robot. 7(5), 841–857 (2015)
28. Zittrain, J.: Perfect enforcement on tomorrow’s internet. In: Brownsword, R., Yeung, K.
(eds.) Regulating Technologies: Legal Futures, Regulatory Frames and Technological Fixes,
pp. 125–156. Hart, London (2007)
 Revisiting Constitutive Rules

Giovanni Sileno(B) , Alexander Boer, and Tom van Engers

Leibniz Center for Law, University of Amsterdam, Amsterdam, Netherlands

{g.sileno,aboer,vanengers}@uva.nl

Abstract. The paper is an investigation on how behaviour relates to

norms, i.e. how a certain conduct acquires meaning in institutional terms.
The simplest mechanism determining this phenomenon is given by the
‘count-as’ relation, generally associated with constitutive rules, through
which an agent has the legal capacity, via performing a certain action,
to create, modify or destroy a certain institutional fact. In the analytic
literature, however, the ‘count-as’ relation is mostly approached for its
classificatory functions, mapping entities to categories whose members
carry institutional properties. Besides making explicit this double func-
tion, the paper reconsiders the relation between constitutive rules and
regulative rules, and introduces a proposal on the ontological status of
constitution.

Keywords: Constitutive rules · Institutional rules · Regulative rules

Connotation · Import · Institutional power · Behaviour · Norms
Supervenience

1 Introduction
An important question, still unresolved in legal theory and in analytic literature,
concerns the nature (and for certain authors, the very existence) of constitutive
rules, and their distinction from regulative rules. The best known (and discussed)
account is the one developed by Searle [1–3]. As their name suggests, regulative
rules regulate pre-existing forms of behaviour. For example, eating is an activity
introduced well before that any rule of polite table behaviour was introduced.
On the contrary, the rules of playing chess are constitutive: actions in accordance
with them constitute the very activity of playing chess. Searle then argues that
institutions like marriage, money or promise are not different from games such
as baseball or chess, in the sense that they are all systems of constitutive rules.
Despite this simple and intuitive presentation, however, many authors have
attempted to better define the two types of rules, without reaching a definitive
agreement. Understanding institutional constitution is in effect a crucial part of
the study of social ontology, and for this reason it is addressed in linguistics,
social sciences, developmental psychology, economics, and information science,
as well as in philosophy.
While ontology is the general philosophical study about existence, social
ontology focuses on the social reality (distinguished from the physical reality,
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 39–55, 2018.
https://doi.org/10.1007/978-3-030-00178-0_3
40 G. Sileno et al.

and from the individual mental reality), normally by tracking the understanding
of properties and functions of institutions. As convincingly observed by Roversi
in [4] this type of investigation usually takes a rule-realist view : “rules constitu-
tive of an institution can exist only as part of the causal (mental or behavioural)
process through which the institutional activity they constitute is practiced”.
This is the most natural perspective that we could take by reflecting on our
experience as social participants: if mankind disappeared from the world, so
would its institutions. At the same time, Roversi observes that social ontology
is not (yet) a major field of interest for contemporary legal philosophy. Most
legal scholars embrace with more ease a rule-positivist view : “rules constitu-
tive of an institution can exist before and independently of the causal process
through which the institutional activity they constitute is practiced”. This pref-
erence can be explained: the rule-realist view undermines a general tenet of
legal positivism, i.e. the independence of the treatment of elements belonging to
the legal-institutional domain from considerations about their effectiveness (in
economic, social or psychological terms) in the actual world.
Are the rule-positivist and the rule-realist views irredeemably incompatible?
Works on legal institutions as those of [5,6] attempt this quest from a legal philo-
sophical standpoint. From a knowledge engineering point of view, the problem
can be put differently: can a system of norms be aligned—representation-wise—
with a system of practices guided by norms? The investigation of constitutive
rules is a necessary requirement to answer to this question. In the present paper,
for reasons of space, we will overlook technical details, preferring to give a more
exhaustive presentation of the problems at stake and of the solutions presented
in the literature (Sect. 2). Exploiting this analysis, we will introduce an inte-
grated account on constitution (Sect. 3), and utilize this to dissect institutional
power (Sect. 4). Additionally, we will set up the basis for an investigation of the
ontological status of constitution (Sect. 5), preparatory to check the alignment
of representational models.

2 Relevant Literature
Searle: Constitutive and Regulative Rules. Searle’s account on constitutive
and regulative rules can be plausibly taken as the starting reference on this topic
today. Elaborating on considerations by Anscombe and Rawls, he proposes (e.g.
in [1, p. 34]) that the underlying structure of constitutive rules is in the form of:

X counts as Y in context C. (1)

where X and Y are acts. Instead, regulative rules can be paraphrased as:

Do X. (2)

or in a conditional form:

If Y do X. (3)
 Revisiting Constitutive Rules 41

Acts of type X are ‘brute’, i.e. they may occur independently of the rules reg-
ulating them, whereas acts of type Y are institutional: they cannot occur if no
definite constitutive rule is applicable.
Conte: Ludus vs Lusus. Revisiting Wittgenstein, Conte [7] starts by observing
that there is an ontological difference between the rules eidetic-constitutive of a
‘game’ (ludus) and the rules perceived from the ‘play’ (lusus). The former are
necessary for the game to occur.1 He then identifies different and incongruous
uses of the term constitutive rules in Searle’s work:

– X-type of rule: e.g. “to make a promise is to undertake an obligation”, which

can be rewritten as “a promise counts as the undertaking of an obligation”,
with ‘promise’ occupying the position X according to the template (1);
– Y-type of rule: e.g. “a checkmate is made when the king is attacked in such
a way that no move will leave it unattacked ”, which can be rewritten to
“checks in which the king cannot meet the attack counts as checkmate”, with
‘checkmate’ occupying the position Y;
– rules as “one ought not to steal ”, which seem to fall more under the definition
of regulative rules;2
– rules related to (linguistic) performance: e.g. promises should be about future
behaviour.

According to Conte, Y-type rules are the only proper eidetic-constitutive rules.
The issues with the third and fourth case are evident. The argument against
the X-type is that the rule given in the example is not necessary to make a
promise, either ontologically (i.e. it is not necessary for the conception, the actual
possibility or the perception of the promise) or semantically, as it makes only
explicit an intension already present in the speech act of promising.
Jones and Sergot: “Count-As” as Conditional. According to Jones and
Sergot [8], a ‘count-as’ relation establishes that a certain state of affairs or an
action of an agent “is a sufficient condition to guarantee that the institution
creates some (usually normative) state of affairs”. They start by characterizing
this connection as a logic conditional calibrated to avoid unsound effects. Con-
sider, for example, a case in which x’s declaration ‘I pronounce you man and
wife’ “counts in the institution s as a means of guaranteeing that s sees to it
that a and b are married.” In classic propositional logic, the introduction of an
inclusive or in the consequent does not change the validity of the rule: if a → b
holds, then a → b ∨ c also holds. However, Jones and Sergot correctly observe
that it would “be bizarre to conclude that x’s utterance act would also count in

1
We may read the perspective of the legal scholar in this claim. In an actual social
setting, this is often not the case: players may play even without knowing any rule,
just mirroring what others are doing (mimesis) or, more rationally, fabricating their
own models of the rules in place.
2
In Searle’s words, the prohibition of stealing is “a constitutive rule of the institution
of private property”, [1, p. 168].
42 G. Sileno et al.

s as a means of guaranteeing that either Nixon is impeached or s sees to it that

a and b are married”.
Going further, they acknowledge that there “will surely be conditionals which
describe relations of logical consequence, of causal consequence and of deontic
consequence”. Rather than further defining the different types, they propose to
translate the conditional underlying the count-as relation as a constraint ‘if A
then B’ operative in the institution, or, via the material implication 3 , as the
incompatibility with the constraints operative in the institution such that ‘A
and not B’.
Boella and Van der Torre: Normative Goals and Belief Rules. Trying to
analyze the relation between regulative and constitutive norms, Boella and Van
der Torre [9] interpret the normative system via an agent metaphor, applying
the intentional stance [10]. A normative system promotes interests as goals or
values shared by some, most or all agents. These normative goals, delegated by
the individual agents at the collective level, are expressed by regulative rules
(obligation, prohibitions, etc.). What, then, are the ‘beliefs’ of the normative
system? Boella and Van der Torre identify them as ‘brute’ and institutional
facts. The creation of institutional facts (and therefore constitution) is obtained
through belief rules, which introduce institutional categories abstracting actual
situations or other institutional categories.
Grossi: Constitutive, Classificatory, Proper Classificatory Rules. Grossi
starts by observing how several authors in the analytic literature have high-
lighted the classificatory character of non-regulative elements of norms, calling
these determinative rules [11], conceptual rules [12], qualification norms [13] and
definitional norms [14]. This aligns with Searle’s argument about the definitional
nature of constitutive rules.4 Thus, acknowledging that ‘counts-as’ statements
function in practice as classifications, [15] concludes that they could ultimately
be modeled as subsumption relations.5 Constitutive rules would then define an
internal ontology, a conceptualization of the domain under regulation, crucial for
the operationalization of the regulative components, as in the famous example:
“vehicles are not admitted in public parks” (general norm), “bicycles are vehi-
cles” (classification rule), therefore “bicycles are not admitted in public parks”
(specific norm). Grossi proposes to discriminate three different components:

– constitutive rules: making explicit the extra-institutional conditions under

which an institutional term applies, e.g. “In normative system N, conveyances
transporting people or goods count as vehicles”

3
The material implication allows to convert a logic conditional into a composition
of disjunction and negation: (a → b) ↔ (¬a ∨ b). It makes explicit the ‘constraint’
nature of the operator of implication, rather than (epistemic) ‘production’ aspects.
4
“The rules for checkmate or touchdown must ‘define’ checkmate in chess or touch-
down in American Football [...]”, [1, p. 34].
5
Informally, given two concepts X and Y, ‘X subsumes Y’, or ‘Y is subsumed by X’,
means that X (e.g. animal) is an abstraction of Y (e.g. whale).
 Revisiting Constitutive Rules 43

– classificatory rules: making explicit the extra-institutional conditions that

specifies an extra-institutional term, e.g. “It is always the case that bikes
count as conveyances transporting people or goods”
– proper classificatory rules: connecting an extra-institutional term with an
institutional term, e.g. “In normative system N, bikes count as vehicles”
The classificatory rule is completely extra-institutional and can be seen as given,
while the others follow the XYC pattern proposed by Searle: the constitutive rule
is at a more abstract level and the proper classificatory rule contextualizes the
general constitutive rule in more specific terms, but they both refer to a ‘middle
term’ [16] or ‘intermediate concept’ [17]—vehicle, in our example.
Additionally, Grossi observes that, beyond rules constituting institutional
facts (i.e. new classificatory rules), there are rules which “constitute” in the sense
that they “define the normative system, or institution, to which they pertain”.
These rules can be connected to the third type identified by Conte.
Hindriks: Connotation and Import. Following [18], Hindriks [19] distin-
guishes two aspects of constitutive rules:
– connotation defines the conditions which have to be satisfied in order to apply
a certain institutional term: it is a descriptive component;
– import specifies the institutional consequences which occur once those condi-
tions are satisfied.
He proposes therefore to refine constitutive rules under a XYZ scheme. The
first part (XY) corresponds to connotation, which, including context (CXY),
takes the form proposed by Searle. Such constitutive rules (in a strict sense)
link the satisfaction of certain conditions to the applicability of an institutional
term. For instance, “In the United States, bills issued by the Federal Reserve
(X) count as money (Y)”. The second part (YZ) is a status rule, specifying
the practical significance of the institutional status constituted by the first. The
status rule defines the function of the institutional concept. For instance, “one
of the functions of money is that it can be used as a means of exchange, which
means that it facilitates or enables actions, in particular exchange of goods and
services without the use of barter. However, the same idea can be expressed
using the term ‘power’: money can be said to give people the power to perform
the action mentioned.” Hindriks’s convincing argument for this extension is that
without the import, the constitutive rule would not have any concrete role in
the institution.
Boer: Institutional Rules, Constituting, Constitutive Acts. All consti-
tutive rules require at least a ‘brute’ fact to create institutional facts. Boer
[20, p. 93] proposes that we also consider institutional rules: rules that operate
on institutional facts, on the basis of other institutional facts. Status rules can
be therefore seen as a sub-set of institutional rules. Furthermore, he correctly
highlights that ‘brute’ and ‘institutional’ respectively correspond not to physical
and social referents, but to extra-institutional and intra-institutional entities.
A ‘brute’ fact may be a fact that belongs to another institution.
44 G. Sileno et al.

Additionally, he suggests distinguishing constitutive acts, i.e. the acts

intended to constitute an institutional act, within the more general class of
constituting acts. This serves as a reminder that “the operative principle behind
constitutive rules and institutional facts is that people to a large extent have
control over what institutional facts they bring about”. An example of a consti-
tuting act is theft: thieves have no intent to be qualified as such. Interestingly, a
similar intentional/non-intentional characterization may specified distinguishing
regulative from regulating rules. The second would refer to side-effects that were
not intended by the legislator.
Hage: Regulative Rules are Constitutive Rules. A recent article presented
by Hage [21] contends that regulative rules are constitutive rules. Hage first
identifies three types of constitutive rules:
– dynamic rules, which create, modify, or remove facts as the consequence of
an event, e.g. “making a promise generates an obligation for the promisor”;
they may be conditional, e.g. “if it is dark, the occurrence of a car accident
obligates the drivers to place a light on the road next to the cars”.
– fact-to-fact rules, which (defeasibly) attach a fact to another fact in a timeless
fashion (not accounting change); e.g. “if P owns O, P is competent to alienate
O”; they may also be conditional, e.g. “in case of emergency, the mayor of a
city is competent to invoke the state of emergency”.
– count-as rules, rules of the type “individuals of type 1 count-as individuals
of type 2”, where individuals may be persons (e.g. “the parents of a minor
count-as the minor’s legal representatives”), or events (e.g. “under suitable
circumstances, causing a car accident counts-as committing a tort”).
Building upon on these categories, he argues that constitutive rules consist in
a more general class than count-as rules, and their general characterization is
that of rules that eventually affect facts of the world, but that also exhibit some
correspondence between their propositional content and what is in the world. If
we say, e.g. “criminals are liable to enforcement”, this means that any criminal
is liable to enforcement. For completeness, Hage includes a categorization of
regulative rules:
– prescriptive rules, which make a specific conduct obligatory, e.g. “car drivers
must drive on the right hand side of the road”;
– proscriptive rules, prohibiting a specific conduct, e.g. “it is forbidden to tor-
ture sentient beings”;
– rules that specify what should be done, e.g. “if the king is in chess, the threat
should immediately be removed”;
– rules about “how something should be done, without imposing a duty or an
obligation to do so”, as e.g. rules of etiquette.
but argues that they belong to the constitutive category as well, if, in addition
to (descriptive) facts, we take into account deontic facts. In the traditional sense,
a fact is associated with an objective, mind-independent description of what is
the case. However, social reality is a domain for which the ontological realist
 Revisiting Constitutive Rules 45

stance is not (directly) appropriate, as it mostly depends on what people accept

or recognize about it. Nevertheless, we do say “He is the owner of . . . ”, just
as “It is raining”. Thus, as these (descriptive) institutional facts depend on
standards, and are produced by rules, nothing forbids us from considering facts
that describe normative directives (e.g. “He has the duty to . . . ”) as deontic
facts, also produced by rules.

3 An Integrated Model for Constitutive Rules

3.1 Distinguishing Constitutive-Of and Constitutive-For
We acknowledge two meanings for constitutive elements:
(a) as characteristic regulative drivers (constitutive-of the institution),
(b) as part of an interpretative system (constitutive-for the institution).
The former category deals with what constitutes the institution, considered as a
‘subject’ acting in the world. As agents are primarily defined—in terms of the
impact they are disposed to produce on the world—by their desires, institutions
can thus be primarily defined by the requirements they put on the social system.6
The latter deals with how meaning is constituted for the institution, that is, with
the selection of what makes sense for the institution of what occurs in the social
environment, and with the processing of such selection. Interestingly, the two
components are inseparable, although for different reasons.
Let us imagine an institution consisting only of regulative rules. The opera-
tional minimal structure of an obligation consists of two recognition rules, one
for violation and one for satisfaction. Therefore, each regulative rule implicitly
brings at least two constitutive (more precisely, constituting) rules, namely those
defining what generates ‘violation’ and ‘satisfaction’ institutional facts related
to the given prescription. Thus, extending this observation, we unveil the first
function of constitutive rules: they serve to explicitly specify the operating terms
of regulative rules, defining not only satisfaction or violation conditions, but also
the classes of beneficiaries, of addressees, the initiating conditions, etc. In this
sense, they are participatory to the commitment-related structure implemented
by the institution.
On the other hand, games like chess for instance do not have standardized
regulative rules7 : they are in practice mere systems of constitutive rules used
to interpret what counts-as a valid move. Therefore, the second function of
constitutive rules is at the level of competences or abilities of the social partici-
pants. Interestingly, some of these rules may be expressed by referring to deontic
6
Considering that regulative norms can be interpreted as goals associated to the
normative system [9], they are constitutive in the same sense in which maintenance
goals are the policy or, in cybernetic terms, the identity of an autonomous system,
cf. the viable system model (VSM) [22].
7
Bulygin [12] suggests the following: “a player must make a given number of moves
in a given period of time on pain of losing the game”, where losing can be seen as a
sort of punishment, considering the pragmatics around games.
46 G. Sileno et al.

notions, e.g. “one must play with the piece which has been touched”, or “if the
king is in check, the threat should immediately be removed”, but despite what
is observed in the literature (e.g. [12,21]) these rules are not regulative as in
the previous sense. Invalidity entails nullity of the move, but not ‘breach’, nor
‘violation’, nor ‘offense’ (on these lines, see [23, p. 28]). Interpreting the game as
a system of conditional abilities, players follow the rules to acquire new abilities
with the purpose of being able to approaching the winning state, also defined
within the rules of the game. The ‘must’ made manifest in these rules is a deriva-
tion from this individual interest: if you want to win (or at least to play), you
need to make valid moves, and to make valid moves, you must follow the proce-
dures.8 The regulation of behaviour of two persons playing chess is a consequence
of this practical reasoning mechanism and not of regulative rules. Interestingly,
this ability-related structure can be interpreted as a soft form of control, because
it is constructed without any reference to coercion.
Thus, if we include the creation, modification, and destruction of potestative
positions as a form of regulation (just as Hohfeld brought forward the second
potestative square of fundamental legal concepts), we have completed the circle:
Regulative rules always consist of constitutive rules. Constitutive rules always
contribute to regulation. This circularity may explain the analytical difficulty
encountered in the literature to come up with consistent definitions of regulative
and constitutive rules.

3.2 Constitutive Elements

All constitutive elements play a role in the interplay between institutional and
extra-institutional domains. Elements constitutive-for the institution map facts,
actions, or events from the extra-institutional to the institutional domain. Ele-
ments constitutive-of the institution are obligative and potestative dispositions
that (supposedly) influence the behaviour of the agents, occurring in the extra-
institutional domain. Reusing part of the terminology used in the literature we
recognize the following elements:

constitutive rule rule mapping extra-institutional facts to institutional facts;

constitutive fact fact captured by the antecedent of a constitutive rule;
institutional rule rule relating institutional facts to other institutional facts;
status rule institutional rule mapping institutional facts (e.g. about roles) to
normative positions;

8
To reiterate, the ‘must’ that is used in certain normative statements does not refer to
a (conditional) duty, but to an institutional power. Consider for instance “in order to
perform a real estate transaction, buyers and sellers must sign a written contract”.
In this sort of cases, ‘must’ is derived from practical necessity (“to be obliged to”),
more than normative aspects (“to have the obligation to”): e.g. if buyer and seller
want to perform a sale, they don’t have any other way but signing a contract.
 Revisiting Constitutive Rules 47

Their interaction is visualized in Fig. 1. The regulation is the effect of the nor-
mative positions currently holding. Note that all -ive elements (explicit, inten-
tional) can be replaced by the wider -ing class (including implicit and non-
intentional mechanisms).

status
institutional fact rule
institutional
rule

institutional
domain
constitutive
rules

extra-institutional regulation
domain

extra-institutional fact constitutive facts

Fig. 1. Coupling between institutional and extra-institutional domains

Sources of Facts. The initial definitions of constitutive and regulative rules

given by Searle (1, 2, 3) are centered around acts, but later authors soon extended
them to events, to states of affairs, and then to facts. Unfortunately, ‘facts’ can
be quite different things depending on the tradition upon which the author
builds, consciously or not. From an ontological perspective, facts are arrange-
ments of entities, objects, events/actions, or processes. From an epistemic per-
spective, facts are justified true beliefs about such arrangements or occurrences,
and therefore transport propositional content. Which of these perspectives are
we referring to when dealing with constitutive rules?
Considering an agentive perspective, the associated philosophical problem
would be to settle whether these facts are directly perceived facts or represen-
tationally mediated facts. The distinction between presentation and representa-
tion is a traditional argument in phenomenology, e.g. [24, pp. 144–145], but it
has recently returned in analytic philosophy with the discussion about the ‘bad
argument’ [25]. Entering into the details of this debate is out of scope here; how-
ever, observing the reconstruction we have developed so far, illustrated in Fig. 1,
we can argue that the whole mechanism of constitution can be seen as a pro-
totypical mechanism of re-presentation. In effect, ‘counts-as’ can be interpreted
also as ‘stands-for’. Institutional facts are prototypical mediators and therefore
represented facts. On the other hand, extra-institutional facts may be perceived
(i.e. non-mediated) facts, or representations, if their meaning is built upon other
institutions (the use of language nicely fits with the second case). We do not
need to specify them further.
48 G. Sileno et al.

3.3 Separating Static and Dynamic Aspects

In general, systems can be divided into two categories (cf. [26]):
– transformational systems, characterized mostly by static, timeless, steady
aspects, which can be easily represented in functional terms;
– reactive systems, characterized by dynamic, temporal, asynchronous aspects,
which cannot be easily represented in functional terms;
A similar distinction can be applied to the sub-components of an institutional
interpretative system.
Static, Conditional Aspects. In agreement with the literature, conditional
classification or subsumption is plausibly the most effective relation to capture
static extra-institutional aspects of reality charged with institutional meaning.
For instance, “bikes counts as vehicles”. The related rule would be in the form
of a classificatory constitutive rule:

In context C, an entity of type X counts as an entity of type Y. (4)

Within the institutional system, we can also consider rules that are not
grounded on extra-institutional facts, but operate only at the institutional level.
These may be definitional, for instance “a check in which the king cannot meet
the attack counts as checkmate”, or “a formal charge which addresses a public
officer counts as an impeachment”. In these specific cases, constitution is rather
an is-a relation and the associated definitional institutional rule would be:

An entity of type Y1 is an entity of type Y2. (5)

However, most of institutional rules are status rules, mapping institutional

notions (Y) to normative aspects (Z), i.e. deontic and potestative character-
izations. Related examples are “a promise counts as an obligation”, “in case
of emergency, a mayor has the competence to declare the state of emergency”
(considering both promise and emergency as institutional facts, cf. Sect. 5):

An entity of type Y implies the existence of an entity of type Z. (6)

In this case it is not a matter of definition: the two entities are different, a
promise is not an obligation, and an emergency is not a competence. From a
logical point of view, these rules function as remapping of the parametric content
specifying one entity into the other, e.g. the promise of doing A implies the duty
of doing A.
Dynamic, Procedural Aspects. Generally speaking, the term act refers both
to a performance and to its outcome. However, from the outcome, we can always
refer back to the action. For instance, “a promise counts as an obligation” can
be rephrased as “positing a promise counts as undertaking an obligation”, i.e.
in terms of an initiating event. The result is an institutional event rule:

An event of type Y1 implies the occurrence of an event of type Y2. (7)

 Revisiting Constitutive Rules 49

To consider the relation at the production level (with the creation of the promise)
rather than at the outcome level (the settled promise) is, in this example, only
a matter of taste. If the promise is removed, so is the obligation. This example
does not support the introduction of a new modeling dimension. Let us consider
then another example: “raising a hand during an auction counts as making a
bid”. This is a constitutive event rule:

In context C, an event of type X counts as an event of type Y. (8)

In this case, there is a decoupling from the ‘brute’ result of the hand-raising
action and its institutional counterpart: we may let the hand go down, but
our bid would remain. These dynamic aspects of reality are not reducible at the
level of outcome, and the procedural/event component of the constitution plays a
crucial role. For those, the traditional logic notation is problematic, because logic
conditionals require an adequate machinery to deal with incremental change.9
Similar problems have been studied in contrary-to-duty (CTD) obligations [27].

4 Constitutive Dimensions of Institutional Power

Raising a hand to make a bid is an example of action conducted in the physical
reality to obtain a result in the institutional domain. If we turn our attention
from the action to the agent, we have already observed that what enables the
social participant to produce the intended institutional outcome is being dis-
posed, besides the practical ability, with the relevant institutional power (also
ability, capacity or competence, depending on the tradition). Without this power,
the agent would be not able to constitute the outcome. What, then, is the rela-
tion between institutional power and constitutive rules? Our proposal elaborates
on this notion in terms of dispositions.
In general, a disposition is a precondition necessary to reach, at the occur-
rence of an adequate stimulus, a now only potential state. This transformation,
and the resulting outcome, count as the manifestation of the disposition. Typi-
cal examples are being fragile or being soluble.10 Dispositions are requirements
for change (e.g. an element can be dissolved in a solution only if the element is
soluble). On the other hand, they provide also behavioural expectations about
the referent entities (a soluble element is expected to dissolve in a solution).
Applying this notion to our domain, we can define institutional power as a
disposition whose manifestation is the creation, destruction or modification of
institutional entities. This definition is wider than the one usually encountered
9
On the other hand, when a relation can be represented between the outcomes, the
procedural model requires the introduction of adequate revision mechanism for oper-
ational closure, and therefore, it becomes less efficient from a representational per-
spective.
10
Disposition is a long-debated notion in philosophy, especially in metaphysics. Lewis
provides in [28] a famous critique to the classic account based on logic condition-
als, and a reformulation in causation terms, which is compatible with the present
proposal.
50 G. Sileno et al.

in legal scholarship. For instance, offering, or infringing the law, are actions usu-
ally not considered associated to legal capabilities. The first because, differently
from accepting an offer, it does not create any obligation. The second because
it is not a type of action promoted by the legal system. However, from a formal
point of view, they do entail consequences at institutional level.11
Evidently, physical actions performed in a specific context become vectors to
constitute institutional facts through constitutive event rules. This concerns the
performance component of institutional power. Other orthogonal components
used in specifying institutional power concern the minimal requirements for the
qualification of the performer to the role he is enacting and the delimitation
of the institutional subject-matter on which the power may be exercised. Con-
sidering these three dimensions, we organize in Table 1 the examples of legal
specifications of power reported by Hart in [23, p. 28]. The case of judicial officer
could be extended similarly to other public officers. In dispositional terms, with
some approximation, qualification defines the disposition, performance defines
the stimulus and delimitation provides ingredients to specify the manifestation.
In terms of constitutive rules, the first component can be related to classifica-
tory rules (4), the second to constitutive event rules (8), and the third defines
or constrains the codomain of status rules (6).

Table 1. Specifications of institutional power defined by law, examples.

Private persons Judicial officers Legislative authority

Qualification Minimum requirements of Manner of appointment, Qualifications of identity
personal qualification qualifications for and of the members of the
(capacity) tenure of judicial officer legislative body
Performance Manner and form in which Procedure to be followed in Manner and form of
the power is exercised the court legislation, procedure to
(execution, attestation) be followed
Subject-matter Variety of rights and duties Jurisdiction Domain over which the
which may be created power may be exercised

5 On the Ontological Status of Constitution

The previous sections clarify how constitution functions, but we haven’t yet
investigated what type of relation constitution is. One way to approach this
topic is to start by addressing the domains of its terms.
Ontological Stratification of Institutions. Amongst the authors reviewed
in Sect. 2, only Hindriks [19] and Boer [20] explicitly elaborate and argue for an
ontological distinction between institutional and extra-institutional (including
‘brute’) realms. It is plausible that also the others share, implicitly or tacitly,
a similar perspective. In contrast, Searle rejects in several points of his works
11
In a similar spirit, Sartor extends in [29] action-power with generic-power, that can
be associated to natural events as well (e.g. death, timeouts, etc.).
 Revisiting Constitutive Rules 51

the idea that there are different levels in reality (e.g. [3, p. 1]). However, as
connotation is contextual, the same extra-institutional facts may yield different
institutional outcomes depending on the context, and, therefore, this argument
is difficult to maintain: at least from a formal point of view, Searle seems to
conflate constitution and identity relations.12 Secondly, this argument overlooks
the existence of a plurality of institutions, and of institutional interpretations,
and thus the intrinsic possibility of conflicting institutional outcomes.
Informal and Formal. Interestingly, the ontological distinction between intra-
and extra-institutional domains results in a framework affine with the legal
abstract model proposed by Breuker [30], advancing the idea that institutional
layers are built upon a common-sense knowledge layer. Consider the analysis of
promise given by Conte for the X-type of rule: “a promise counts as the under-
taking of an obligation”. His interpretation insists on the fact that the meaning
of promise lies already in linguistic practice as a fundamental speech act, and
consequently, the proposed rule is merely descriptive. In Hindriks’s terms, how-
ever, the rule can be interpreted as an import rule, which, in a legal context,
would instantiate a legal obligation (thus protected by law). For this reason, it
would be a different rule than the one followed in social practice. The nature of
the ‘promise’ term is not settled, however. When there is not a definite consti-
tutive rule that specifies the criteria for which a promise can be accepted as a
valid promise, the institutional system can be seen as relying on the meaning
constituted at extra-institutional level. The resulting mechanism can be modeled
in two ways:

– by introducing an implicit constitutive rule that remaps the extra-

institutional fact in a cloned institutional reference for institutional import;
– by considering connotation and import collapsing into the same link, directly
associating the constitutive fact to the normative fact.

The second is evidently simpler, and avoids the introduction of unnecessary

links. However, the first solution is interesting, as it prepares for consequent
developments. In effect, it is reasonable to expect the enactment of an explicit
constitutive rule in all cases in which the original extra-institutional term is
acknowledged to introduce non-predictability in the functioning of institutional
mechanisms. For instance, in certain contexts, promises are considered valid
(the speech act of promising counts effectively as an institutional promise) only
when they are in written form, plausibly because in oral form they turned out
to be insufficiently reliable. Thus, we may conclude that explicit constitutive
machinery ultimately responds to the requirement of reducing the frictions caused
by different interpretations of what moulds the institutional matter.
Emergence and Supervenience. Strangely enough, the recognition of differ-
ent ontological strata, i.e. a division of reality in domains to be treated for the
12
The canonic form of constitutive rules (1) implies that when we are not in C, X may
not count as Y. This shows that it is impossible that constitution corresponds to
identity, as X would be equal to Y in certain cases, and not equal to Y in others.
52 G. Sileno et al.

most part separately, would be, in principle, compatible with Searle’s attempt
to provide a naturalistic account of language [3, p. 61]. In effect, natural sciences
approach reality depending on various factors, such as the dimensional scale in
focus (e.g. particle physics vs astrophysics). Theories and accounts associated to
these approaches are often so incompatible, that they may be seen as targeting
different realities. Maintaining this distinction furnishes a framework compatible
with the analysis and treatment of emergent properties or emergent phenomena,
i.e. those arising out of more fundamental ones, but not reducible to them.
In philosophy, several authors have attempted to capture the relation
amongst different ontological strata working with the notion of supervenience. In
the simplest form, “we have supervenience when there could be no difference of
one sort without differences of another sort” [31, p. 14]. Considering for instance
the physical reality, we may say that the macroscopic level supervenes the micro-
scopic level because any difference observed at the macroscopic level necessarily
implies a difference at the microscopic level. But the notion is applied in other
domains as well, e.g. in support of the recognition of “the existence of mental
phenomena, and their non-identity with physical phenomena, while maintain-
ing an authentically physicalist world view” [32]. In other words, supervenience
makes explicit an intrinsic ontological asymmetry: e.g. mental or institutional
states cannot change without having a change occurring at the physical level.
What is Constitution? Why supervenience is relevant for constitutive rules?
Even without referring to supervenience, Hindriks [19] expresses a similar intu-
ition, citing Baker’s analogy with aesthetic relations. A painting does not directly
‘define’ its own beauty (determination), nor ‘cause’ it (material production), but
it ‘constitutes’ it. The connection of a painting with its beauty is a classical
example of the use of supervenience (although more debated than the macro-
micro scenario).13 The notion of supervenience is compatible with the idea of
constitution advanced by this work: constitutive (classificatory or event) rules
can be seen as reifying the interactions between extra-institutional and insti-
tutional domains, with the latter supervening the former.14 Informally stated,
many events (conditions) may occur (hold) in the world which are irrelevant
from an institutional point of view. However, if in a certain moment the institu-
tional domain was found to be different, this means that something necessarily
changed in the extra-institutional (e.g. ‘brute’) domain as well: i.e. a part of the
constitutive base must have triggered such a change at institutional level.
Towards the Operationalization of Alignment. The previous analysis sug-
gests an alternative approach in testing whether two representations are aligned.
In the literature, due to the prominent focus on their classificatory function,
constitutive rules are usually specified via a subsumption relation. Subsumption

13
If supervenience holds, it is impossible that there are two paintings that are the
same from a physical point of view (e.g. for their distribution of colours), but they
are different in respect of how beautiful they are (to respond to relativist critics, we
should add for the same observer and in the same mental state).
14
This idea was briefly presented in [33] as well, but it remains underspecified.
 Revisiting Constitutive Rules 53

between two prototypical entities is verified when all the properties of one entity
match a sub-set of the properties of the other. However, in the previous sections
we showed that the classificatory view is not sufficient to capture all the types of
constitution. In this context, supervenience offers a better frame than subsump-
tion: we do not target the verification of an equal (sub-set of) properties, but of a
fit alignment of differences after change. Intuitively, given two behavioural mod-
els, when the execution of the supposedly supervenient model exhibits a change,
we should verify that some aligned change occurs in the supposedly base model.
A preliminary operationalization following this idea has been presented in [34].

6 Conclusion and Further Developments

The paper revisits the notion of constitutive rules, attempting an integration and
synthesis of previous contributions. The intuition to carefully distinguish declara-
tive components from reactive components came after the examples of conflation
in both cognitive and computational domains remarked by Kowalski and Sadri
[35]. Our analysis confirms that the nature of constitutive rules is complex, and
suggests that this complexity is due to the integration of the different types of
interactions that may occur between ‘brute’ (or better, extra-institutional) and
institutional domains.
The study is functional to a more general research objective: the alignment
of representations of law (norms), of implementations of law (e.g. services), and
of intentional characterizations of behaviour (cases) [36]. With respect to rep-
resentation of law, we presented in [37] a revisitation of Hohfeld’s framework
in interactional terms; in [27] we investigated the contrary-to-duty (CTD) con-
structs studied in deontic logic. With respect to representation of behaviour, we
introduced in [38] an agent architecture based on the notions of commitment,
affordance, expectation and susceptibility, interpreted in analogy with Hohfel-
dian notions. This paper focuses on the theoretical aspects about the connection
between extra-institutional and institutional components, but, as the other refer-
ences show, our current efforts are also directed on establishing a unifying formal
visual notation (based on Petri Nets), in support to our theoretical proposal.
Evidently, the constitution of institutional meaning follows the sense of con-
stitutive rules (from behaviour to institutional domain) but it also implements
a feedback on behaviour through regulation. Furthermore, social systems adapt
to institutional mechanisms—a phenomenon observable through the emergence
of “nomotropic” behaviours, i.e. of “acting in light of rules” (which is different
from “in conformity with rules”) [39]—to which social systems respond again
by modifying their own institutional mechanisms. In the full picture, constitu-
tive rules establish a structural coupling between the two domains. However,
because adaptation mechanisms are much slower than operational mechanisms,
on shorter temporal scales the coupling is asymmetrical. This assumption allows
to associate constitution to the notion of supervenience, thus enabling the veri-
fication of alignment, but the analysis of the institutional dynamics accounting
for the change of norms remains to be investigated.
54 G. Sileno et al.

References
1. Searle, J.R.: Speech Acts: An Essay in the Philosophy of Language. Cambridge
University Press, Cambridge (1969)
2. Searle, J.R.: Intentionality: An Essay in the Philosophy of Mind. Cambridge Uni-
versity Press, Cambridge (1983)
3. Searle, J.R.: Making the Social World: The Structure of Human Civilization.
Oxford University Press, Oxford (2010)
4. Roversi, C.: Acceptance is not enough, but texts alone achieve nothing. A critique
of two conceptions in institutional ontology. Rechtstheorie 43(2), 177–206 (2012)
5. MacCormick, N.: Norms, institutions, and institutional facts. Law Philos. 17(3),
301–345 (1998)
6. Ruiter, D.W.P.: Structuring legal institutions. Law Philos. 17(3), 215–232 (1998)
7. Conte, A.G.: L’enjeu des règles. Droit et Société 17–18, 125–146 (1991)
8. Jones, A., Sergot, M.: A formal characterisation of institutionalised power. J. IGPL
4, 427–443 (1996)
9. Boella, G., van der Torre, L.: Constitutive norms in the design of normative multia-
gent systems. In: Toni, F., Torroni, P. (eds.) CLIMA 2005. LNCS (LNAI), vol. 3900,
pp. 303–319. Springer, Heidelberg (2006). https://doi.org/10.1007/11750734 17
10. Dennett, D.C.: The Intentional Stance, 7th edn. MIT Press, Cambridge (1987)
11. von Wright, G.H.: Norm and Action: A Logical Enquiry. Routledge & K. Paul,
London (1963)
12. Bulygin, E.: On norms of competence. Law Philos. 11(3), 201–216 (1992)
13. Peczenik, A.: On Law and Reason. Kluwer, Dordrecht (1989)
14. Jones, A.J., Sergot, M.: Deontic logic in the representation of law: towards a
methodology. Artif. Intell. Law 1(1), 45–64 (1992)
15. Grossi, D.: Designing invisible handcuffs, formal investigations in institutions and
organizations for multi-agent systems. Ph.D. thesis, University of Utrecht (2007)
16. Atkinson, K., Bench-Capon, T.J.M.: Levels of reasoning with legal cases. In: Pro-
ceedings of the ICAIL 2005 Workshop on Argumentation in AI and Law (2005)
17. Lindahl, L., Odelstad, J.: Intermediate concepts in normative systems. In: Goble,
L., Meyer, J.-J.C. (eds.) DEON 2006. LNCS (LNAI), vol. 4048, pp. 187–200.
Springer, Heidelberg (2006). https://doi.org/10.1007/11786849 16
18. Ransdell, J.: Constitutive rules and speech-act analysis. J. Philos. 68(13), 385–399
(1971)
19. Hindriks, F.: Constitutive rules, language, and ontology. Erkenntnis 71(2), 253–275
(2009)
20. Boer, A.: Legal theory, sources of law and the semantic web. Ph.D. thesis, Univer-
sity of Amsterdam (2009)
21. Hage, J.: Separating rules from normativity. In: Araszkiewicz, M., Banas, P.,
Gizbert-Studnicki, T., Pleszka, K. (eds.) Problems of Normativity, Rules and Rule-
Following. LAPS, vol. 111, pp. 13–22. Springer, Cham (2015). https://doi.org/10.
1007/978-3-319-09375-8 2
22. Beer, S.: Brain of the Firm. Wiley, New York (1995)
23. Hart, H.L.A.: The Concept of Law, 2nd edn. Clarendon Press, Oxford (1994)
24. Husserl, E.: The Shorter Logical Investigations. Taylor & Francis, Abingdon (2002)
25. Searle, J.R.: Perceptual intentionality. Organon F 19, 9–22 (2012)
26. Harel, D., Pnueli, A.: On the development of reactive systems. In: Apt, K.R. (ed.)
Logics and Models of Concurrent Systems. NATO ASI Series (Series F: Computer
and Systems Sciences), vol. 13, pp. 477–498. Springer, Heidelberg (1985). https://
doi.org/10.1007/978-3-642-82453-1 17
 Revisiting Constitutive Rules 55

27. Sileno, G., Boer, A., van Engers, T.: A Petri net-based notation for normative mod-
eling: evaluation on deontic paradoxes. In: Proceedings of MIREL 2017: Workshop
on MIning and REasoning with Legal texts, in conjunction with ICAIL 2017 (2017)
28. Lewis, D.: Finkish dispositions. Philos. Q. 47, 143–158 (1997)
29. Sartor, G.: Fundamental legal concepts: a formal and teleological characterisation.
Artif. Intell. Law 14(1), 101–142 (2006)
30. Breuker, J., den Haan, N.: Separating world and regulation knowledge: where is the
logic? In: Proceedings of ICAIL 1991: 3rd International Conference on Artificial
Intelligence and Law, pp. 92–97 (1991)
31. Lewis, D.K.: On the Plurality of Worlds. B. Blackwell, Oxford (1986)
32. Brown, R., Ladyman, J.: Physicalism, supervenience and the fundamental level.
Philos. Q. 59(234), 20–38 (2009)
33. Hage, J., Verheij, B.: The law as a dynamic interconnected system of states of
affairs: a legal top ontology. Int. J. Hum.-Comput. Stud. 51(6), 1043–1077 (1999)
34. Sileno, G., Boer, A., van Engers, T.: Bridging representations of laws, of implemen-
tations and of behaviours. In: Proceedings of the 28th International Conference on
Legal Knowledge and Information Systems (JURIX 2015). FAIA, vol. 279 (2015)
35. Kowalski, R., Sadri, F.: Integrating logic programming and production systems in
abductive logic programming agents. In: Polleres, A., Swift, T. (eds.) RR 2009.
LNCS, vol. 5837, pp. 1–23. Springer, Heidelberg (2009). https://doi.org/10.1007/
978-3-642-05082-4 1
36. Sileno, G.: Aligning law and action. Ph.D. thesis, University of Amsterdam (2016)
37. Sileno, G., Boer, A., van Engers, T.: On the interactional meaning of fundamen-
tal legal concepts. In: Proceedings of the 27th International Conference on Legal
Knowledge and Information Systems (JURIX 2014). FAIA, vol. 271, pp. 39–48
(2014)
38. Sileno, G., Boer, A., van Engers, T.: Commitments, expectations, affordances and
susceptibilities: towards positional agent programming. In: Chen, Q., Torroni, P.,
Villata, S., Hsu, J., Omicini, A. (eds.) PRIMA 2015. LNCS (LNAI), vol. 9387, pp.
687–696. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-25524-8 52
39. Conte, A.G.: Nomotropismo: agire in-funzione-di regole. Sociologia del diritto
27(1), 1–27 (2000)
 The Truth in Law and Its Explication

Hajime Yoshino(&)

Meiji Gakuin University, Tokyo, Japan

hyoshino@ls.meijigakuin.ac.jp

Abstract. This paper discusses what types of truth play their role in law and in
what way such types of truth can be explicated. For this purpose, this paper
applies the logical viewpoint and method of predicate logic to clarify the logical
structure of legal sentences and legal reasoning through their application. This
paper presents its central argument that the concept of truth in law is to be
classified into three types of truth, i.e., truth as fact, truth as validity and truth as
justice, and provides their formal semantic foundation. This paper analyzes the
way to explicate truth as validity and truth as justice in the ways of intensional
and extensional explications on the one hand and in the way of the reasoning of
justification and the reasoning of creation on the other hand.

Keywords: Truth
Validity
Justice
Rule
Fact
Logic

Predicate logic
Legal object sentence
Legal meta-sentence
Reasoning of justification
Reasoning of creation
Explication of truth

1 Introduction

The concept of truth has played a great role in law. This paper at first aims to clarify in
what dimensions the concept of truth plays its role in law. In other words, it aims to
clarify what types of truth play their role in law. Secondly, it aims to clarify in what
way, to be precise, through what inference, such truths in law can be determined.
In order to realize these purposes, this paper applies the classical mathematical
logic, i.e., predicate logic to discuss the topic of truth in law. Through the application of
this method, this paper shortly clarifies the logical structure of legal sentences (Sect. 2)
and legal reasoning (Sect. 3), both of which will be invoked to the discussion of the
concepts of truth in law. This paper then presents its central argument that the concept
of truth in law is to be classified into three types of truth, i.e., (1) truth as fact, (2) truth
as validity, and (3) truth as justice (Sect. 4). It provides the formal semantic foundation
of these three types of truth in law (Sect. 5). It discusses further the way to explicate the
concepts of truth as validity and truth as justice in law. In doing so, it analyzes not the
abstract concept of truth as validity or justice itself but the legal sentences which are
conceived as valid or just. It also discusses the way of clarification in terms of the
intension and extension of the concepts on the one hand and of the reasoning of legal
justification and the reasoning of legal creation on the other hand (Sect. 6). This paper
concludes by summarizing its achievements and listing further tasks which are still to
be solved (Sect. 7).

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 56–71, 2018.
https://doi.org/10.1007/978-3-030-00178-0_4
 The Truth in Law and Its Explication 57

2 The Logical Structure of Legal Sentences

and Its Representation

The concept of truth plays its role concerning sentences. In other words, sentences can
be said that they are true or false. The concept of truth also plays its role in the
reasoning, which operates with sentences. Therefore, to discuss the concept of truth in
law in terms of logic, it is necessary, as a preparing exercise, to clarify the logical
structure of legal sentences and its representation.
To clarify the logical structure of legal sentences, we apply the logical method to
them. The method of logic, which is applied to this paper to analyze legal sentences and
reasoning, is the method of classical mathematical logic, especially the first order
predicate logic. One may ask: why is a special logic of norms like Deontic Logic1
developed for norms not applied to legal sentences? The reason is this: the predicate
logic is effectively applicable to legal sentences on the one hand and it is questionable
whether real laws use the deontic conceptions of “obligation”, “prohibition” and
“permission” presented by Deontic Logic, and therefore, we do not need the calculation
between them defined by such logic of norms on the other hand. Such logics that have
been presented so far are so weak that they cannot adequately represent real legal
sentences describing complex legal states of affairs. Also, sometimes, paradoxes have
appeared in their application to legal arguments. Those paradoxes could not emerge if
one applied first order predicate logic to legal arguments.2
To logically formalize law, we should analyze law into the minimal units of legal
sentences and construct law as their logical connection. The author presents three fun-
damental types of legal sentences as the minimal units which have been found through
the author’s study on the construction of legal knowledge base systems [11, 16, 17].
These three fundamental types of legal sentences which are broken down into two
further sub-types are:
(1) Legal rule sentences and fact sentences
(2) Legal element sentences and complex sentences
(3) Legal object sentences and meta-sentences

2.1 Legal Rule Sentences and Fact Sentences

Legal rule sentences are sentences that have the following syntactic structure:
8Xfað X Þ bð X Þg.3

1
Georg Henrik von Wright has developed Deontic Logic [8].
2
The author shows that Ross’s paradox could not emerge if one applied the classical propositional
logic correctly to normative arguments [9]. The author also demonstrates that the paradox of
“Contrary-to-Duty Imperatives” presented by von R. M. Chisholm does not occur if one formalizes
the imperatives by means of predicate logic [10].
3
This is a predicate logical formula in which a string which begins with the upper cases is used for
variables and a string with the lower cases for constants.
58 H. Yoshino

Legal fact sentences on the other hand have the following syntactic structure:

bðx1Þ:

This structure of legal sentences, i.e., dividing them into rule sentences and fact
sentences, enables jurists to apply legal rule sentences to fact sentences of the problems
to deduce a decision as a conclusion of a logical inference based on the logical
inference rule of Modus Ponens. If a legal rule sentence like 8Xfað X Þ bð X Þg and a
legal fact sentence like bðx1Þ are set as premises, the conclusion a(x1) is logically
deduced based on Modus Tollens. The inference is represented as follows: 8Xfað X Þ
bð X Þg&bðx1Þ ) aðx1Þ.4 Here, the structure of “sentence” and the “inference” are
closely connected.

2.2 Legal Element Sentences and Complex Sentences

Legal element sentences mark the smallest units of legal sentences. Legal complex
sentences are the combination of legal element sentences. For example, a contract is a
legal complex sentence and an article of the contract is a legal element sentence. The
“United Nations Convention on Contracts for the International Sale of Goods”
(“CISG”) is a legal complex sentence, whereas CISG “Article 23 A contract is con-
cluded at the moment when an acceptance of an offer becomes effective” is a legal
element sentence. Each legal sentence has a unique name. Through the introduction of
these conceptual devices, the author breaks law into its smallest elements while also
reconstructing it logically from those elements on the way it exists in fact.5

2.3 Legal Object Sentences and Meta-sentences

A legal object sentence regulates an obligation for people to perform a certain action.
Therefore, a legal object fact sentence is formalized in the following predicate logical
formula:
is_obligatory(Person, Action).6
A concrete example of a legal object fact sentence and its predicate logical
formula is:
ofs1: Bernard is obligated to pay the price $5000 to Anzai.
ofs1: obligated(‘Bernard’,pay(‘Bernard’, $5000, ‘Anzai’,T))
If this sentence is used or asserted in an inference, it is meant that “Bernard is
obligated to pay the price $5000 to Anzai” is legally true. Jurists have represented this

4
The author feels that the structure of law, in the form of rule and fact sentences, might be one of the
great subconscious inventions by mankind and/or a miracle made by God.
5
The advantage of the concept of the legal complex sentences is to be able to deal with the validity of
legal element sentences which belong to a legal complex sentence all at once by determining the
validity of the relevant complex sentence. This relationship is regulated by implicit fundamental legal
rules which later will be logically formalized as mr01 and mr3aa1 in the Sect. 6.5.
6
“Person” and “Action” are used for the valuables for persons and for actions.
 The Truth in Law and Its Explication 59

kind of situation that a legal sentence which represents an obligatory state of affairs is
legally true by means of legal meta-sentences using the predicate “valid”.
A legal meta-sentence7 regulates the validity of another legal sentence.8 Therefore,
they should have the predicate representing “valid,”9 and the terms for legal sentences,
whose validity is in question, and the terms for the scope of validity (in which scope of
time, places, people and matters the sentence is valid). An example of a legal meta-fact
sentence on the legal object fact sentence ofs1 explained above (where the scope of the
validity is restricted only to time) and its predicate logical formulae is:
mfs1: ofs1 is valid on February 23rd, 2016:
mfs1: is valid(ofs1,23_02_2016).
The author has introduced the sentences describing the “justness” of other legal
sentences as a subclass of the concept of legal meta-sentences. Thus, the following
sentences are also legal meta-sentences:
mfs2: Sentence ofs1 is just on October 15, 2016.
Its predicate logical representation is:
mfs2: just(ofs1, 2016_10_15).
In the examples above of ofs1 and mfs1, when it is proven in the legal inference that
the legal meta-sentence mfs1 is true, it is conceived that the legal object sentence ors1
is valid in fact and that the state of affairs of the obligation represented by the object
sentence exists in the legal world. How these types of legal meta-sentences can be
inferred will be discussed in the foregoing chapter (Sect. 6).
These conceptual devices of legal sentences can logically represent real legal
sentences used in law as they exist and logically formalize the inference to decide the
validity of legal sentences as they are done.

3 The Logical Structure of Legal Reasoning

Legal sentences are developed through legal reasoning. The author clarifies the
structure of the legal reasoning in terms of the reasoning of legal justification and the
reasoning of legal creation.

3.1 The Logical Structure of the Reasoning of Legal Justification

The reasoning of legal justification is the reasoning to justify legal sentences as the
consequences of logical deductions from “valid” (legally true) legal rule sentences
together with “true” (factually true) fact sentences. The logical deduction is based on
the inference rule of “Modus Ponens”:

7
The terminology “meta” originates from Tarski’s “meta-language” [7].
8
The author got the idea of “legal meta sentence” from H.L.A. Hart’s “secondary rules.” Cf. Hart, H.
L.A., The concept of law, Oxford 1961, p. 79; 2nd edition, Oxford University Press, NY 1994, p. 80.
9
The predicate is not restricted to the noun “validity.” Other predicate which represent the conception
of the validity are available, e.g., “is valid,” “become valid,” “become null,” “is terminated,” and so
on.
60 H. Yoshino

ððA ) BÞ&AÞ ) B

The rule of Modus Ponens is a fundamental inference scheme of the reasoning of

justification in law, especially in the reasoning of the application of law to given cases.

3.2 The Logical Structure of the Reasoning of Legal Creation

The reasoning of legal creation (or discovery)10 is the reasoning to newly create legal
sentences which are necessary to constitute the reasoning of justification. The logical
structure of reasoning of legal creation is clarified in terms of two related types of
reasoning: (1) generating hypothetical legal sentences and (2) testing these hypothetical
legal sentences.
(a) The Logical Structure of Reasoning for Generating Legal Sentences. The
reasoning behind generating hypothetical legal sentences is not the deduction but the
abduction or induction. The abduction takes place when a fact sentence is generated
and the induction, on the contrary to that, takes place when a rule sentence is generated
according to the author’s definition.
An example of abduction is as follows:

(1) A goal q(a) is to be proven as true.

(2) There is a rule qð X Þ pð X Þ which has the goal qð X Þ as a head of the rule.
(3) To prove goal qðaÞ by the application of the rule, the necessary fact sentence is
generated through the abduction.
The abduction is mainly performed to confirm the fact of the given event, which
may fulfill the requirement of the relevant legal rule sentence that is applied to the case.
An example of an induction is as follows:

10
It may be common in traditional legal theories to use the terminology of “discovery”, not of
“creation,” to indicate such non-deductive legal reasoning. However, the word “discovered” should
be used for the case that one finds an object or a rule which already exists. Legal sentences which
are necessary to constitute the reasoning of justification do not beforehand exist but are actually
“created” by the applicators of law, which will be discussed later. Therefore, it is better not to use
“discovery” but to use “creation”.
 The Truth in Law and Its Explication 61

The induction is mainly performed to create a new rule sentence which is necessary
to solve the problem adequately in correspondence to the situation of society. 11
(b) The logical structure of the reasoning for testing the generated legal rule
sentences. In relation to Karl Popper’s falsification theory [4], the author thinks that the
logical structure of the reasoning for testing the generated hypothetical legal rule
sentences is “Modus Tollens”, which is represented in propositional logical formulas as
follows:

ðP ! QÞ&:Q ) :P

This shows: If Q is followed from P, and Q is negatively evaluated, i.e., falsified,

then the negative evaluation of P is followed, i.e., P is falsified.

4 The Concepts of Truth in Law – Truth as Fact, as Validity

and as Justice

The author thinks that three types of truth are used in law: (1) truth as fact, (2) truth as
validity and (3) truth as justice.

4.1 The Truth as Fact

“Truth as fact” concerns whether sentences describing the facts of cases are factually
true or false. If sentences describing the state of affairs of the case are factually true,
then the relevant state of affairs is regarded as existing in the factual world.

4.2 The Truth as Validity

“Truth as validity” concerns whether legal (rule) sentences are legally true or false. If it
is proven that a legal meta-sentence describing the validity of another legal (rule)
sentence is true, then the state of affairs described by the latter legal (rule) sentence is
regarded as the case in the legal world. Therefore, it is applicable. The concept of truth
as validity plays its role in the application of legal rules to solve problems.
It is remarkable that jurists have represented this concept of legal truth using the
predicate “valid” or “validity” and produced the system of legal rule sentences
according to which the (legal) truth of given legal sentences, i.e. whether they are
(legally) true or false, is determined.
For example, if a legal meta-fact sentence which describes the validity of a legal
object sentence representing an obligation of a person to do a certain action is proven as
true, the obligation of the person to take the action is regarded as existing in the legal
world.

11
The reasoning of generating a more abstract general rule sentence from many individual
concrete/specific rule sentences is also called “induction”. This reasoning has basically the same
inference structure as this.
62 H. Yoshino

If a legal meta-fact sentence which describes the validity of a legal rule sentence is
proven as true, the meaning of the rule sentence regulating a certain matter is regarded
as being the case. The legal rule sentence is applied for activating logical inferences
using the inference rule of Modus Ponens, as explained above in Sects. 2.1 and 3.1.
Below, the author tries to make a visual expression about this relationship between
the proof of the validity of a legal rule sentence and the activation of the logical
inference, applying the relevant legal rule sentences, in Fig. 1.

Fig. 1. The relationship between the inference to prove the validity of a legal rule sentence and
the inference in which the relevant valid legal rule sentence is applied.

(1) It is proven as true that CISG12 Article 23 is valid on December 14, 2016. The
Article expresses: “A contract is concluded when an acceptance of an offer becomes
effective”. (2) CISG 23 is applied to solve the problem of a contract on the day. (3) As
the legal rule sentence of Article 23 is proven as valid (legally true) in the legal world, a
logical inference is activated through the application of the rule sentence of Article 23
based on the inference rule of Modus Ponens. As a result, it is proven that the contract
is concluded on December 14, 2016.

4.3 The Truth as Justice

“Truth as justice” concerns whether legal (rule) sentences are just or unjust. If a meta-
sentence describing that a legal (rule) sentence is just is proven as true, then the latter
legal (rule) sentence is just and therefore the state of affairs described by this legal
(rule) sentence is conceived as just in the legal world.

12
CISG is the abbreviation of “United Nations Convention on Contracts for the International Sale of
Goods”.
 The Truth in Law and Its Explication 63

How can the above interpretation “truth as fact”, “truth as validity” and “truth as
justice” be semantically founded? The author will provide their formal semantic
foundation below by applying the scheme of the definition of truth in logic by Tarski.

5 The Formal Semantic Foundation of the Concepts

of Truth in Law

The concept of truth in classical logic, especially in the predicate-calculus, is formally

defined by Tarski [6, 7]. In the following paragraphs, the outline of Tarski’s inter-
pretation of the concept of truth will be presented in a concise manner. The following
symbols will be used: U: a single term predicate; i: an interpretation-functor; a1 ; . . .; an :
an individual constant or variable.

ð AÞ : Uða1 ; . . .; an Þ is true under i if \iða1 Þ; . . .; iðan Þ [ 2 iðUÞ; and

ðBÞ : Uða1 ; . . .; an Þ is false under i if \iða1 Þ; . . .; iðan Þ [ 62 iðUÞ

(A) and (B) are equivalent to (A) with “if and only if” instead of “if”.
Accordingly, when a one-term predicate applies to an individual constant or
variable being part of the set which is the extension of the interpreted predicate, then
the respective statement-formula is true and, if not, then it is false. For a better
understanding of this principle, an illustration will be given with a one-term predicate
below in Fig. 2.

Fig. 2. The definition of the concept of formal truth in logic.

Based on the foregoing demonstrations, one should point out that the definition by
Tarski of the concept of truth of logic is constructed purely formally. It does not matter
by what criteria the fulfillment must be decided. According to the definition by Tarski,
the logical calculus needs, as a presupposition, nothing but the purely formal principle
of bivalence, namely, that a value of two possible values “true” (“1”) or “false” (“0”) is
64 H. Yoshino

allocated uniformly to every sentence [10]. Legal sentences which law and legal rea-
soning consist of can be evaluated as valid or invalid and as just or unjust. Here, the
bivalence principle is valid so that legal sentences can be evaluated as true or false in
the sense of classical logic. There is no difficulty for the predicate logic to be applied
to law.

5.1 The Formal Semantic Definition of the Concept of Truth as Fact

The Tarski-type of the definition of the concept of truth described above can be applied
to define the concept of truth as fact:

ðA1 Þ : Uða1 :. . .; an Þ is factually true if \iða1 Þ; . . .; iðan ÞÞ [ 2 iðUÞ

ðB1 Þ : Uða1; . . .; anÞ is factually false if \iða1 Þ; . . .; iðan ÞÞ [ 62 iðUÞ

When an interpreted constant or variable falls under the class of the interpreted
predicate, then the predicate formula is factually true (A1) and otherwise it is factually
false (B1).

5.2 The Formal Semantic Definition of the Concept of Truth as Validity

The Tarski-type of the definition of the truth concept described above can be applied to
interpret the concept of truth as validity as follows:

ðA2 Þ : Uða1 :. . .; an Þ is valid if \i aq ; . . .; iðan ÞÞ [ 2 iðUÞ

ðB2 Þ : Uðða1 :. . .; an Þ is invalid if \iða1 Þ; . . .; iðan ÞÞ [ 62 iðUÞ

When a constant or variable comes under a predicate, i.e., when an interpreted

individual constant or –variable falls under the class of the interpreted predicate, then
the predicate formula is valid (A2) and otherwise it is not valid, namely invalid (B2).

5.3 The Formal Semantic Definition of the Concept of Truth as Justice

The Tarski-type of the definition of the concept of truth described above can be applied
to interpret the concept of truth as the concept of justice:

ðA3 Þ : Uða1 :. . .; an Þ is just if \iða1 Þ; . . .; iðan ÞÞ [ 2 iðUÞ

 The Truth in Law and Its Explication 65

ðB3 Þ : Uðða1 :. . .; an Þ is unjust if \iða1 Þ; . . .; iðan ÞÞ [ 62 iðUÞ

When an individual constant or – variable comes under a predicate, i.e. when an

interpreted individual constant or – variable falls under the class of the interpreted
predicate, then the predicate formula is just (A3) and otherwise it is unjust (B3).
These definitions show that the concepts of validity and justice can be treated as the
concept of truth in logic. But they do not intend to provide any criterion on validity or
on justice, namely any criterion to determine whether a given legal rule sentence is
valid or just. However, as it is proven that validity and justice can be conceived as truth
in logic, we can now proceed to discuss how the concept of truth as validity and the
concept of truth as justice could be explicated when we conceive validity and justice as
truth in logic.

6 The Ways to Explicate the Concepts of Truth in Law

How can we explicate the concepts of truth in law? The way to explicate the concept of
existing objects is conceivable on the following two ways: the intensional explication
and the extensional explication.

6.1 The Intensional Explication of an Object

The intensional explication is to describe the intension, i.e., the nature of the object. Let
us have a look at the explication of “dog”: A dog is “a highly variable domestic
mammal (Canis familiaris) closely related to the gray wolf” (by Merriam-Webster).

6.2 The Extensional Explication of an Object

The extensional explication is to give examples of the objects which are included in the
set represented by the term, for example, in the case of the extensional explication of
“dog”, to designate mammals which have the nature of a dog.

6.3 The Mutual Relationship Between the Intensional

and the Extensional Explication
The intensional explication of the objects is related to the extensional explication, in the
sense that the extension of the objects is decided when the intension of the objects is
decided.

6.4 The Explication of the Concept of Truth Itself or the Explication

of True Legal Sentences?
How can we intensionally or extensionally explicate the concept of truth in law?
According to the author’s opinion, the abstract concept of truth itself can neither be
explicated well intensionally nor extensionally as an object like a dog.
66 H. Yoshino

However, sentences which are to be confirmed as true, i.e., true sentences, can be
explicated intensionally and extensionally. It would be useful for the science of law and
legal practices if the methods to decide legally true sentences were provided. In the
following, the author will discuss the way to explicate the concepts of truth as validity
and truth as justice in the way to determine valid legal sentences and just legal sen-
tences. (The concept of truth as fact will be discussed in the authors future studies)

6.5 The Way to Determine Valid Legal Sentences

Valid legal sentences are decided through legal reasoning. In the reasoning, legal meta-
rule sentences, which regulate the validity of legal sentences, are applied. Therefore,
this reasoning can be called “legal meta-inference” [11].
The Way to Intensionally Explicate Valid Legal Sentences. Based on the relation
between the intension and the extension of a concept, as explained above, the extension
of a concept is determined when the intension of the concept is determined. The author
thinks that the way to intensionally identify legally valid sentences is to identify legal
meta-rule sentences which regulate the validity of legal sentences. To raise legal meta-
rule sentences can be a kind of an intensional description of valid legal sentences
criteria. To find and raise legal meta-rule sentences, which determine valid legal rule
sentences under the given circumstance, is an extensional explication of the relevant
legal meta-rule sentences themselves. However, it is an intensional explication of the
valid legal rule sentences, because the extension of valid legal rule sentences under the
given circumstance is determined through the application of the raised legal meta-rule
sentences to the circumstance. The author will raise several legal meta-rule sentences
below.
In real positive laws, there is a great amount of legal rule sentences which regulate
the validity of legal sentences. There are also several implicit fundamental legal meta-
rules as legal common sense. Those legal meta-rule sentences are united under the
following most fundamental meta-rule sentence (MFMR):

ðmr0Þ : 8S 8T 8T1fis valid ðS; T Þ

become valid ðS; T1Þ & beforeðT1; T Þ &
notððbecome nullðS; T2Þ & beforeðT2; T ÞÞÞg:

Positive legal meta-rule sentences and implicit fundamental legal meta-rule sen-
tences regulate the determination of the first part of the requirement or of the second
part of the requirement of this rule sentence. To determine the first part of the
requirement, the following fundamental legal meta-rule sentence must be implicitly
valid (From here on, the Universal Quantifier “8” is eliminated.):

ðmr01Þ : become valid ðS; T Þ

element sentenceðS; CSÞ&complex sentenceðCSÞ&become valid ðCS; T Þ:
 The Truth in Law and Its Explication 67

This rule regulates: an element sentence S becomes valid at time T if it is an element

of the complex-sentence CS and the complex sentence becomes valid at time T.
To determine whether the complex sentence becomes valid at time T, the following
fundamental legal meta-rule sentence must be implicitly valid:

ðmr3aa1Þ : becomes validðCS; TÞ !

complex sentenceðCSÞ&formed ðCS; T1Þ&beforeðT1; T Þ&
notðinvalid ðCS; T ÞÞ&
ðððstarting pointðT2; CSÞ&has comeðT; T2ÞÞv
ðconditionðCo; CSÞ&fulfilled ðCo; T ÞÞÞv
T ¼ T1Þ:

This rule regulates: a complex sentence CS becomes valid a time T, if it is formed at

time T1 being before T and it is not invalid, and while the complex sentence CS
includes a starting time point of validity and the time point has come at T or while it
includes the condition of the validity and the condition is fulfilled at time T or T1 is T.
As a contract is a complex sentence, CISG13 Article 23 “A contract is concluded
when an acceptance of an offer becomes effective.” is applicable to determine the first
part of the requirement of the legal fundamental meta-rule mr3aa1. In this way, further
positive legal meta-rule sentences (CISG articles 14 - 24) are applicable to determine
whether a contract is formed.
Positive legal sentences which describe “rights” are to be considered as legal meta-
sentences [13]. They mainly regulate the fulfillment of the first part of the MFMR
(mr01) in the way that the executions of rights make relevant sentences become valid.
In addition, they partially regulate the second part, for example in the way that the
execution of the right of cancelling a contract as a legal complex sentence makes the
contract become null.
To raise legal meta-rule sentences in this way, is a way to intentionally explicate
valid legal sentences.
The Way to Extensionally Explicate Valid Legal Sentences. To infer all legally
valid sentences is the way to extensionally explicate legal valid sentences. Valid legal
sentences can be determined and therefore raised through the legal meta-inference,
where legal meta-rule sentences, which are an intensional description of legally valid
sentences, are applied.
If one could install the necessary and sufficient amount of legal meta-rule sentences
in a legal knowledge base, one could theoretically let computers infer the whole legal
sentences which are valid at a certain time point, through the application of legal meta-
rule sentences. In this way, the extension of valid legal sentences can theoretically be
determined through a computer aided legal reasoning system.

13
CISG is an abbreviation of “United Nations Convention on Contracts for the International Sale of
Goods.”
68 H. Yoshino

Whereby it is to be noted that the validity of legal sentences is relative with respect
not only to “time” but also to “place”, “person” and “matter” to which the relevant legal
sentences are applied. In practicing legal reasoning, one should individually determine
whether a candidate of a legal rule sentence being applied to solve a legal problem is
legally valid at a certain time, at a certain place, for a certain person and regarding a
certain matter.

6.6 The Way to Determine just or Unjust Legal Sentences

Justice of law is one of the central topics of law, especially for the philosophy of law.
But it has been considered that it is difficult for a science to clarify the concept of
justice. Here, the author does expect that this paper’s approach to the concept of justice,
which considers the justice of law as a kind of truth of law and does not deal with the
abstract concept of justice itself but with “just or unjust legal sentences”, makes a
scientific approach to justice possible.
The way to determine just legal sentences. As justice in law is a type of truth as
clarified above, the justness of a legal sentence is proven when it is logically deduced
from other just legal sentences. However, the problem is how other just legal sentences
can be confirmed. In positive laws, there are many legal meta-rule sentences that
regulate the determination of valid legal sentences (Cf. [12, 13, 15]), whereas there is
no positive legal rule sentence which regulates the determination of just legal sen-
tences. This suggests that the reasoning of justification does not play a big role in
determining just legal sentences, but mainly the reasoning of creation or discovery will
play its role in determining just legal rule sentences.
The inference scheme of the reasoning of creation as a falsification represented as
Modus Tollens, explained in Sect. 3.2 (b), can be analyzed more precisely as follows
(by identifying the system of the already existing legal rule sentences (R) and the added
hypothetical legal rule sentence (r1) and by representing the event to which the relevant
rule sentences are applied as (E) and the consequence resulting from the application of
r1 together with R as (C)).

fðR [ r1Þ&E ! Cg&:C ) :r1

This scheme can be read as follows: If legal rule sentences which are merged of the
already existing legal rule sentences R with a hypothetical legal rule sentence r1 are
applied to an event of the case E, then the consequence C will result from their
application. However, the consequence is to be negatively evaluated as :C. Therefore,
the hypothetical legal rule sentence is to be negatively evaluated as :r1 [14].
This type of falsification reasoning should be performed for a sufficient number of
times to make clear that the relevant hypothetical legal sentence will bring no serious
unjust result and that the legal sentence can, therefore, be confirmed as tenable or
relatively just. The whole inference scheme of the falsification reasoning which leads to
a just legal sentence, being confirmed as not falsified, is represented in Scheme 1:
 The Truth in Law and Its Explication 69

Scheme 1. Reasoning of falsification to test hypothetical legal rules

We should analyze past arguments over justice to find out in which way, inten-
sionally or extensionally, they have tried to identify just or unjust legal sentences. We
should analyze them further from the viewpoint of the reasoning of legal creation or
discovery.
The Way to Intensionally Explicate Just or Unjust Legal Sentences. To the
question of “What is justice?”, Aristotle answered: “Justice is equality” [1]. This is an
intensional explication of the concept of justice. Hereby, the next question arises
immediately: How can “equality” play its role to decide whether a legal sentence is
just?
Here, the author only tries to answer to the second question. Legal rule sentences
are to be evaluated, in terms of justice, in possible results of the application of the
relevant legal rule sentences to the problems which are to be solved. The results can be
evaluated in terms of the criteria of “equality”. If the application of the relevant legal
rules causes such results which would be unequal, then the legal rules are to be
evaluated as unjust. Here, the inference of the legal creation based on Modus Tollens
(explained above) will play its role.
It is further necessary to discuss how the inequality of the results of the application
of the relevant legal rule sentence is evaluated. The author cannot avoid leaving the
discussion of this problem open until, in the future, the analysis of real legal arguments
regarding “equality” will be done.
In this sense, the theories of justice that try to provide the nature of justice, like
Aristotle, who did with “equality”, stand on the way to intensionally explicate just legal
sentences.
70 H. Yoshino

The Way to Extensionally Explicate Just Legal Sentences. Natural law theories in
modern times like Pufendorf’s seem to have provided a system of just legal rule
sentences [5]. This is an example of an extensional approach to explicate just legal rule
sentences.
How can the extension of such just legal rule sentences be acquired? This should be
precisely analyzed from a logical point of view. In contrast to the inference to deter-
mine the extension of valid legal rule sentences, the extension of just legal sentences
cannot even be theoretically determined by means of a computer inference. This is the
case because the former is determined based on the reasoning of justification, in which
the logical deduction plays its role, as far as necessary and sufficient legal meta-rule
sentences are presupposed as the premise of the deduction. However, the latter is
related to the legal reasoning of creation where, directly, only unjust legal rule sen-
tences can be identified through the falsification inference as it was explained in
Sect. 6.6.
The extension of just legal rule sentences in the sense of legal rule sentences which
passed enough falsification tests and therefore are confirmed as tenable or relatively just
can only be extended step by step at the moment. This still requires an evaluator’s
hardworking brain and debates between people because no practicable computer pro-
gram has yet been developed for this reasoning. The algorithm for such a computer
program should be researched intensively. The very computer simulation of these
falsification tests – in which the prediction of the result of the application of hypo-
thetical legal rule sentences plays an important role – is needed to be developed for a
genuine science of law.

7 Conclusion

The author believes that this paper has provided an overview of the concepts of truth in
law from the logical point of view.
The first achievement of this paper is to have presented three sorts of the concept of
truth in law, i.e., (1) truth as fact, (2) truth as validity and (3) truth as justice. It has also
formally and semantically founded such a classification of the concepts of truth in law.
The second achievement is that this paper has analyzed the inference to determine
the concepts of truth as validity and truth as justice in terms of an intensional or
extensional way on the one hand and in terms of the reasoning of justification and
creation of legal sentences on the other hand.
One of the future tasks related to this paper is to discuss the inference, to determine
truth as fact. It is necessary to clarify the inference which determines that legal fact
sentences are factually true or false.
Another future task is to clarify the logical structure of the inference to determine
“just” legal sentences more precisely. It is especially necessary to analyze real argu-
ments in positive law theories and legal philosophies regarding just legal sentences.

Acknowledgements. This paper was written by the author during his studies as a visiting
professor at the Christian-Albrechts-Universität zu Kiel (CAU), Faculty of Law, and during his
studies as a visiting scholar at Northeastern University School of Law (NUSL). The author is
 The Truth in Law and Its Explication 71

grateful to both schools for having provided a good research environment. The author expresses
his deep appreciation, especially to the host Prof. Robert Alexy together with his colleague Prof.
Ino Augsberg, Prof. Rudolf Meyer-Pritzl, Prof. Joachim Jickeli, and Prof. Michael Stöber at CAU
as well as the host Prof. Sonia Elise Rolland together with her colleague Dean Prof. Jeremy R.
Paul, Prof. Karl E. Klare, and Prof. Patrick Cassidy at the NUSL. Finally, the author would also
like to express his gratitude to his student assistants Sven Petersen, Regina Kardel, Dennis
Hardtke, and Amanda Dennis for their devoted assistance.

References
1. Aristotle: Politics, 2nd edn. University of Chicago Press, Chicago (2013). Edited and
Translated by Carnes Lord. 1282b 22
2. Chisholm, R.M.: Contrary-to duty imperatives and deontic logic. Analysis 24, 33–36 (1963)
3. J¢rgensen, J.: Imperative and logic. Erkenntnis 7, 88–296 (1937, 1938)
4. Popper, K.: The Logic of Scientific Discovery, p. 30. Hutchinson, London/New York (1959)
5. Pufendorf, S.: De Jure Naturae et gentium libri octo. Amsterdam edition (1688), 143 ff. The
translation by Oldfather, 208 ff
6. Tarski, A.: The semantic conception of truth and the foundation of semantics. J. Philos.
Phenomenol. Res. 4, 341–375 (1944)
7. Tarski, A.: The concept of truth in formalized language. In: Tarski, A. (ed.). Logic,
Semantics, Metamathematics, pp. 152–278, 167–168. Oxford University Press (1933, 1936)
8. von Wright, G.H.: Deontic logic. Mind LX, 1 (1951)
9. Yoshino, H.: Zu Ansätzen der Juristischen Logik. In: Tammelo (ed.) Strukturierungen und
Entscheidungen im Rechtsdenken, pp. 279–282. Wien, New York (1978)
10. Yoshino, H.: Über die Notwendigkeit einer besonderen Normenlogik als Methode der
juristischen Logik. In: Klug, U., Ramm, T., Rittner, F., Schmiedel, B. (eds.) Gesetzge-
bungstheorie, Juristische Logik, Zivil- und Prozeßrecht, pp. 140–161. Springer, Heidelberg
(1978). https://doi.org/10.1007/978-3-642-95317-0_13
11. Yoshino, H.: The systematization of legal meta-inference. In: Proceedings of the Fifth
International Conference on Artificial Intelligence and Law, pp. 266–275 (1995)
12. Yoshino, H.: The systematization of law in terms of the validity. In: Proceedings of the
Thirteenth International Conference on Artificial Intelligence and Law, Danvers MA,
pp. 121–125 (2011)
13. Yoshino, H.: The logical analysis of the concept of a right in terms of legal meta-sentence.
In: Proceedings of Internationales Rechtsinformatik Symposion (IRIS), pp. 305–312 (2012)
14. Yoshino, H.: Justice and Logic, Jusletter IT 11, September 2014
15. Yoshino, H.: The concept of truth in law as the validity. In: Yoshino, H., et al. (ed.) Truth
and Objectivity in Law and Morals, Stuttgart, pp. 13–31 (2016)
16. Yoshino, H., et al.: Legal expert system — LES-2. In: Wada, E. (ed.) LP 1986. LNCS, vol.
264, pp. 34–45. Springer, Heidelberg (1987). https://doi.org/10.1007/3-540-18024-9_20
17. Yoshino, H.: Legal expert project. J. Adv. Comput. Intell. Tokyo 1(2), 83–85 (1997)
 From Words to Images Through Legal
Visualization

Arianna Rossi(B) and Monica Palmirani

Università di Bologna, CIRSFID, via Galliera 3, 40121 Bologna, Italy

{arianna.rossi15,monica.palmirani}@unibo.it

Abstract. One of the common characteristics of legal documents is

the absolute preponderance of text and their specific domain language,
whose complexity can result in impenetrability for those that have no
legal expertise. In some experiments, visual communication has been
introduced in legal documents to make their meaning clearer and more
intelligible, whilst visualizations have also been automatically generated
from semantically-enriched legal data. As part of an ongoing research
that aims to create user-friendly privacy terms by integrating graphical
elements and Semantic Web technologies, the process of creation and
interpretation of visual legal concepts will be discussed. The analysis of
current approaches to this subject represents the point of departure to
propose an empirical methodology that is inspired by interaction and
human-centered design practices.

Keywords: Privacy · Legal design · Visualization · Legal XML

Legal semantic web · Interpretation · Interaction design
Human-centered design

1 Introduction
It is a common experience that legal terms, licenses, consent requests and in
general any legal notice overload web applications. At the same time, they are
ignored by most users, especially by digital natives. This is a paradox: on the
one hand, overregulation. On the other hand, individuals’ disregard. For these
reasons, interest towards the visualization of legal clauses is growing with the
aim of capturing and retaining individuals’ attention, while providing intelli-
gible and effective communication. In this light, the current research aims to
model a theory for the visual representation of legal documents, with a concrete
application to privacy terms.
To create visualizations in our research, we intend to leverage the different
layers through which legal documents can be represented in the Semantic Web:
text, structure, legal metadata, legal ontology and legal rules [32]. After having
offered a complete and correct representation of a privacy policy on all these
levels, we plan to build an additional layer on top of them: the visualization.
However, it can be argued that it is indispensable to address the topic of visual
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 72–85, 2018.
https://doi.org/10.1007/978-3-030-00178-0_5
 From Words to Images Through Legal Visualization 73

representation of legal knowledge and its interpretation, specifically by answering

the following questions:

1. What are the benefits and the risks of visualized legal information?
2. How can legal visualizations be generated?
3. How can machine-readable legal data be leveraged to create visualizations
and what are the advantages?
4. Is it possible to ensure a correct interpretation of legal visualizations?

In the present position paper, we intend to provide a preliminary answer

to these questions. Firstly, after having outlined the research scenario, some
successful cases of legal visualizations will be presented. Next, the possibil-
ity to build semi-automatic visualizations on semantically-enriched legal data
will be discussed. Then, the connection among communication theories, inter-
action, design and legal hermeneutics will be briefly introduced, alongside the
interpretative process carried out when legal information is transformed into a
machine-readable format. Current approaches for the generation and interpre-
tation of visualizations adopted by legal scholars and legal designers will also
be examined. Finally, some empirical design-oriented suggestions to address this
topic will be made. Our intention is to implement and test them in the ongoing
research described below.

2 Research Scenario

During an ongoing research, we propose semi-automatic visualizations of pri-

vacy policies and consent agreements (see also [31]). The debate around data
protection is extremely topical: concerns about the practices of collection and
processing of personal data are spreading, while regulations to protect data sub-
jects are enforced. In the European Union, the principle of transparency laid
down in Article 12 of the General Data Protection Regulation [13] (hereafter,
GDPR) mandates the provision of intelligible and easily accessible information
on data practices. The aim is that of empowering individuals to be knowledgeable
about how their data is used and, as a consequence, to make informed decisions
[2], for instance when they choose a certain service over another or when they
consent to certain processing operations.
However, privacy disclosures are typically not read or not understood [37].
The use of visual cues has been proposed as a possible solution [17,33], since it
has been demonstrated that they can enhance the effectiveness of legal commu-
nication (see next Section). Furthermore, it is the GDPR itself that suggests the
use of icons “to give in an easily visible, intelligible, and clearly legible manner
a meaningful overview of the intended processing” [13, Article 12(7)]. For these
reasons, a privacy icon set is under development [23]. But its creation has raised
some questions about the possibility of misrepresenting the legal terms during
their conversion into visual elements, as well as the possibility of misinterpreting
the visualizations (as it has emerged from some studies, e.g. [19,30]).
74 A. Rossi and M. Palmirani

3 Background
3.1 Legal Visualizations and Legal Design
The discussion on the comprehensibility of legal sources must be understood as
a part of the emerging research area of Legal Design, which is “the application
of human-centered design to the world of law, to make legal systems and services
more human-centered, usable, and satisfying” [18]. Thanks to the online envi-
ronment, the legal message has exited the exclusive realm of lawyers. This means
that new methods of communication must be considered to allow any individual,
even a layperson, to access and understand legal information. In some contexts,
as pointed out earlier, this is mandated by the law. We have entered a new era
where design, communication and information technology must produce novel,
user-friendly interfaces to the law [10].
Although the total absence of graphics is typical of modern legal texts, with
exceptions such as the highway code and patents [6], this tendency is chang-
ing. For instance, principles of information design and graphic design have been
applied to contracts [26–28], in order to produce user-friendly legal documents
that are able to elicit information effectively, easily and quickly. As for what
concerns the privacy ecosystem, innovative ways of communication and presen-
tation are arising, although these attempts are rare and scattered [16]. In these
experiments, visualization is crucial. Indeed, the support of visual elements helps
unburden the cognitive load that derives from reading, navigating and under-
standing cumbersome documents, such as legal texts. There exist several different
visual representation techniques, depending on the type of information, on the
addressee, on the context, on the goal etc. For instance, flowcharts (Fig. 1) can
express complex conditional structures that are typical of legal texts better than
prose, whereas swimlane tables (Fig. 2) can highlight vis-a-vis the roles, rights,
and responsibilities of different stakeholders [27]. Graphical symbols, such as
icons (Fig. 3), can also be used in legal texts to foster understanding, memoriza-
tion, and quick information retrieval. The present research around the generation
and interpretation of visual elements focuses on this latter type of visualizations.

3.2 Legal Visualizations and Legal Informatics

Legal visualizations can be integrated into legal informatics. In fact, both
research areas are concerned with the management of the complexity of legal
knowledge [15]. Moreover, both disciplines deal with the representation of legal
information: one in a visual format, the other in a machine-readable format.
Semantic Web technologies allow the automated processing of semantically-
enriched, machine-interpretable information, that can even be rendered graphi-
cally. For instance, visualizations were generated from XML marked-up legisla-
tive data to display the complexity of legal order overtime1 [22] and of the Italian
legislative procedure2 .
1
http://lodpiemonte.cirsfid.unibo.it.
2
http://code4italy.cirsfid.unibo.it.
 From Words to Images Through Legal Visualization 75

Fig. 1. Example of flowchart used to elicit payment procedures and consequences of

delayed payments in the visual guide for the Finnish terms of public procurement [27].
2013
c Aalto University & Kuntaliitto ry. Licensed under CC-BY-ND 3.0.

Fig. 2. Example of swimlane table used to illustrate the parties’ rights and responsi-
bilities in the visual guide for the correct with Finnish terms of public procurement
[27]. 2013
c Aalto University & Kuntaliitto ry. Licensed under CC-BY-ND 3.0.
76 A. Rossi and M. Palmirani

Fig. 3. The tabular format proposed in Annex 1 of the draft report on the proposal for
the GDPR [14] for standardised information policies. The first column contains privacy
icons, the second column contains the conditions represented by the icons, while the
third column must be filled by the data controller with either one of the graphical
symbols of Fig. 3b, depending on whether the condition is fulfilled.

The legal XML standard Akoma Ntoso [25] offers unique opportunities to
model the structural and semantic content of legal documents, so that it can be
processed by software applications. Furthermore, the metadata layer of Akoma
Ntoso allows great flexibility and, therefore, adaptation of any legal document
to any ontological representation of concepts [4]. The machine-readable infor-
mation that is captured by the mark-up enriches and is in turn enriched by the
resources available on the (legal) Semantic Web [32], thus creating a complex
network of sources and information. Legal ontologies enrich the Akoma Ntoso
XML representation with the necessary semantic level that permits the connec-
tion between text and legal rules. Another legal XML language, LegalRuleML
[3], can integrate Akoma Ntoso for what concerns the mark-up of the logical
 From Words to Images Through Legal Visualization 77

structure of legal rules. For instance, it can model deontic norms (obligations,
permissions, prohibitions, rights) and can manage negations.
The structural, semantic, logical and ontological layers of a legal document
can, thus, provide the information needed to propose a semi-automatic visu-
alization of its content [31]. Furthermore, the encoding of legal content in a
machine-readable format provides the opportunity to interact with it in order to
customize its presentation for an intended audience and in a certain context. For
instance, automated tools have been proposed to build interactive visualizations
of contractual terms according to the input provided by users [29].

4 Legal Visualizations: From Representation

to Interpretation
4.1 Communication, Design, Hermeneutics, and Visualizations
Visualization provides an alternative, supplementary manner of conveying legal
content (concepts, norms, etc.) to the traditional text-based legal communica-
tion. The nature of communicative processes must be briefly introduced3 in order
to analyze the role of interpretation in legal visualizations. Communication is
basically constituted by an encoder (the addresser) who sends a message to a
decoder (the addressee) who receives and makes sense of that message (and
reacts according to her sense-making) [20]. However, communication does not
merely correspond to the exact match between the encoders’ intended meaning
of the message and the decoders’ sense-making result: it is rather a continuous
process of meaning negotiation between them. Design can be considered as a
sort of communication: the designer gives a certain meaning to an artifact (e.g.
a certain element of a graphical user interface) and the user interprets how,
where, and when the artifact can be used [35]. However, no matter how care-
fully an artifact is designed, its meaning is not objective nor static: users unfold
and decode the message (embedded by designers in signs like words, icons, com-
mand buttons, interface layouts etc.) while they interact with the system [36].
This is why, designers should provide users with all possible means (e.g. hints,
explanations, etc.) to facilitate their sense-making of interface signs.
A comparable interactive meaning negotiation activity takes place in legal
hermeneutics. Meaning does not derive solely from the intentions of the legisla-
tor nor from the activity of the legal interpreter: meaning is constantly produced
within an interpretative dialogue among author, text and interpreter in specific
contextual conditions. Thus, interpretation cannot be unique, it can only be
coherent and correct according to shared criteria [38]. The theory of legal inter-
pretation provides not only an explanation for one of the possible meanings, but
also a methodology based on arguments, reasoning and values, that can justify
one possible interpretation in the light of a certain social context and of a cer-
tain case [1]. On the other hand, common sense or preunderstanding can support
legal experts to carry out their analysis to obtain a correct interpretation [9].
3
A comprehensive approach to communication theories goes well beyond the scope of
this paper.
78 A. Rossi and M. Palmirani

Similarly, the visualization of legal content should be sufficiently evocative to

produce in its addressees a visual perception comparable to the preunderstanding
activity, in order to ensure a correct decoding of the visual legal message. In fact,
the graphical representation has the goal of making legal terms more accessible
and understandable. However, if wrongly interpreted, it would create obscurity
in lieu of transparency. It is one thing to rely on a shared visual vocabulary
(e.g. the Highway Code) and to depict concrete objects (e.g. patents or technical
regulations). It is another thing to represent abstract or complex concepts with
arbitrary symbols. For instance, user studies reveal that misinterpretation of pri-
vacy icons can occur [19,30] (the latter on the icons in Fig. 3). In these cases, the
designer embedded a certain meaning in a symbol, which however did not meet
other individuals’ expectations and understanding, for instance due to different
cultural backgrounds. Moreover, the visualization should also provide the nec-
essary information to account for the methodology used to obtain a graphical
representation of a certain legal concept.
In this scenario, two different issues must be addressed: firstly, define the
criteria for the choice of a certain visual representation for a legal concept and,
secondly, define a methodology for the induction of the correct legal concept in
the decoder of a legal visualization.

4.2 Machine-Readable Representation and Interpretation

Even the analysis and consequent XML mark-up of legal documents cannot
abstain from interpretation. Although the expertise of legal knowledge engineers
who mark up the text guarantees the reliability of the annotation [25], in some
cases multiple semantic annotations which represent different legal interpreta-
tions are unavoidable [3]. This is why the metadata section of the XML docu-
ments provides relevant information on the interpretation (for instance about the
authoritativeness of the annotator) [24]. Furthermore, multiple metadata layers
about different interpretations of the same document are possible [4]. Despite
the possibility of multiple annotations, legal mark-up disambiguates uncertain
concepts and clarifies meanings [25], thus it can represent a solid basis to build
visualizations. Finally, the concepts of a certain domain, that are captured by
the document mark-up, can be formalized and organized in an ontology, which
lowers the personal bias in the selection, interpretation, and representation of
legal knowledge. This shared formalization can also contain the graphical rep-
resentation of its concepts, that will thus have a precise, stable, and machine-
interpretable meaning [23].

4.3 Iconography, Legal Design and Interpretation

Nevertheless, although it is a common understanding that images are a universal
language [7,12], visualized legal content allows greater freedom of interpretation
than written text: indeed, “[i]mages are potentially more anarchic than words”
[6, p. 89]. Because of the lack of a comprehensive theoretical framework for
visual legal communication, methodological questions on the creation, analysis
 From Words to Images Through Legal Visualization 79

and evaluation of legal visualizations have been raised [8,11] and some (yet
incomplete) answers have been suggested.
Iconographical and iconological methods developed by image disciplines have
been compared to legal hermeneutics, on the grounds that these disciplines aim
to uncover different layers of meaning and discover the deepest one [6]. Image
disciplines and hermeneutics have developed similar interpretative approaches:
the pre-iconographic description of the image elements resembles the preliminary
analysis of individual words and sentences of the legal text; the iconographical
analysis recalls the interpretation of the historical development, systematic anal-
ysis and context of the norm; finally, the iconological interpretation looks for the
deeper meaning and purpose of the picture, similarly to the teleological inter-
pretation of the law. Brunschwig’s seminal work [7] originates from the same
premises, but she also proposes a sound methodology for the creation of legal
visualizations. The author applies methods of “visual rhetoric” derived from
classical rhetoric (in particular from the elocutio process), to the Swiss Civil
Code and transforms norms into drawings, especially through the application of
“visual figures of speech” (i.e. visual association, visual synecdoche, visual sym-
bolization, etc.). The transformation process of text into pictures is inherently
arbitrary, but the correctness and understandability of the images depend on the
following principles: 1. application of graphical elements drawn from traditional
legal iconography; 2. their appropriateness to the time and place; 3. their appro-
priateness to the target audience (e.g. age, background, etc.); 4. compliance
with Gestalt psychology principles (e.g. simplicity, clarity, organization, etc.);
5. aesthetics.
The same principles are usually respected by legal designers. They draw best
practices from human-centered design and they usually cooperate with legal
experts and other individuals with diverse backgrounds [5,10] to graphically
elaborate concepts with the end-user of the legal document in mind. Although
legal design does not explicitly tackle legal interpretation, this approach guar-
antees, on the one hand, the correctness of the visual representation of legal
concepts thanks to the knowledge of legal experts [7] and, on the other hand, it
considers the characteristics of the user that could influence the interpretation
(i.e. age, education, culture) [33]. Indeed, the design process starts with empirical
studies (e.g. surveys and interviews) and observations that reveal users’ needs
and characteristics, so that designers do not project on them their own beliefs
and assumptions [5]. The process ends with user-testing, which is an empirical
evaluation of the legal visualized document, e.g. in terms of comprehension of
the legal meaning embedded in visualizations [26].

5 Suggestions for a Visual Legal Interpretation

Framework
As illustrated above, there still lacks a comprehensive framework for legal text
interpretation with the explicit aim of generating visualizations and for the inter-
pretation of visualized legal content. However, the approaches illustrated above
80 A. Rossi and M. Palmirani

(cf. Subsect. 4.3) represent an essential point of departure to propose empirical

solutions to the interpretation of legal visualizations, which is essential to our
current research. Brunschwig [7] proposes clear, practical guidelines for the legal
design process, but some aspects must be re-elaborated and integrated for our
particular research topic. Firstly, this approach was proposed for paper docu-
ments, whereas privacy policies live in an online and interactive environment.
Secondly, despite her sound methodology, the author does not offer a solution to
the possibility of multiple interpretation of pictorial norms. Thirdly, the degree
of comprehensibility and engagement of the legal drawings was not studied. By
contrast, the effectiveness of legal design principles and practices has been proven
[26–28], but a couple of aspects must be stressed to guarantee an appropriate
representation and interpretation of data protection notions. In the first place,
participatory design [34] must be the preferred framework so that ordinary users
are consulted not only before and after the design process, but also during it.
Should this not be the case, if a legal visual message is clear and unambiguous
for its creators (e.g. legal and design experts), all the other users (the mes-
sage decoders) will not necessarily assign the same meaning to it. In the second
place, the graphical creation cannot only represent the view of one single group.
Although less personal than the work of one person, it would not be representa-
tive of all the user-groups (e.g. designers, lawyers, computer scientists, laypeople,
etc.) and, thus, its meaning could be ambiguous.

5.1 Moving Forward: Participation, Representativeness,

Customization
We now propose an empirical methodology that is inspired by the aforemen-
tioned approaches and by current design practices to visualize the information
expressed in privacy policies. The methodology is based on three pillars, that we
believe should be integrated to propose an answer to the questions raised in the
introduction: participation, representativeness and customization.

Participation. The knowledge of legal experts ensures the correct interpre-

tation of the privacy concepts that must be rendered graphically, whereas the
expertise of graphic designers together with existing good practices guarantee
the quality of visualization. Since privacy and data protection notions are legal
but also technical in nature, in this specific context even the knowledge of IT
professionals is valuable. However, even ordinary users need to be involved in the
creation stage to ensure that the graphical representation of the legal concepts
mirrors their mental model. For instance, a technical expert and a layperson
might represent the concept of “personal data” differently4 , because diversity
4
Indeed, this tension became visible during a participatory workshop that we orga-
nized for the generation of privacy icons [21]. “Data” was initially represented as a
cylinder, which is part of a shared visual vocabulary in computer science. It was,
however, not understandable by those without technical background, thus it was
transformed into the less specialist, but more widespread representation of a file
folder.
 From Words to Images Through Legal Visualization 81

in experiences and knowledge produces different mental images of it. “Partic-

ipation stands in contrast with the cult of the specialist” [34, xi]: during the
design process, users should cooperate with experts to reach a compromise that
considers as many different points of view as possible. In this way, ordinary users
become message encoders who will probably shape the visual message in a more
comprehensible way for ordinary message decoders like themselves.

Representativeness. It is questionable whether the legal graphical represen-

tations will be interpreted in the same way as intended by the group of creators,
even if this is multidisciplinary and takes into account expert as well as non-
expert views. To reach the greatest level of agreement among message encoders
and decoders, an open consultation or a crowdsourcing experiment about the
correspondence between a certain graphical element and a certain legal concept
could be launched. Then, an accuracy measurement of image-concept matching
and of interpersonal agreement must be applied, in order to determine which
visualizations can be considered the most representative of certain concepts for
a significant number of individuals. Of course, it is necessary to determine what
this means according to the specific context. In the case of the data protection
icons, they aim to become a EU standard: they should, therefore, be sufficiently
and consistently evocative for individuals with dozens of different nationalities,
different ages, different backgrounds, etc., which is no easy task.

Customization. Although the outlined suggestions aim to find the highest

degree of agreement to ensure correspondent interpretations, it is objectionable
whether one single representation is ideal for every typology of user. Unlike con-
tracts that are usually destined to a specific user-group (e.g. businesspeople,
engineers, etc.), privacy policies address any type of person: from well-educated
teenagers to technology-illiterate pensioners. This is why a unique interpretation
is difficult to achieve, while customization could provide a viable solution. Pri-
vacy policies are not fossilized on paper, but can be rather conceived as online
graphical, possibly dynamic, interfaces. Thus, a certain degree of customization
that responds to the different needs and characteristics of different users can
be envisaged. Culture, age, technical proficiency, linguistic proficiency and legal
knowledge are some of the features that must guide the creation of the visual-
ization and, consequently, its interpretation. At the same time, users themselves
must have the possibility to initiate the customization. For instance, different
users could choose between different degrees of visualization: from a maximum
of textuality (e.g. for lawyers) to some degree of pictoriality (e.g. for teenagers).
Moreover, as seen above, interaction is a fundamental part of the interpre-
tative process: the user assigns meanings to the elements of an interface while
using it (in our case, the privacy policy) and can even adjust the meanings as
the interaction proceeds. Users could be given the opportunity to opt for the
kind of visualization they will better understand and remember, as exemplified
by the online learning platform Memrise5 . For instance, an old user and a young
5
https://www.memrise.com.
82 A. Rossi and M. Palmirani

user might need two different representations of the “justice” concept in order
to interpret it correctly: one might prefer the long-established, traditional sign
of a scale, whilst for the other the icon of Batman might be more meaningful.
All the information gathered from the interaction of users could then be sent
back to the source, confirming or rejecting the visualization proposed by the
encoders. The results of this iterative process can be considered for subsequent
re-elaborations (e.g. all children interpret the image of “Batman” as “justice”).

5.2 A Methodology for Legal Visualizations

In the current research, different disciplines interlace to offer a methodological
framework (see also [23]) for the generation and interpretation of visual elements
that represent legal information of many kinds. The proposed framework is open
to argument and does not intend to provide any final solution to the debate.
However, in this specific context of research, it does provide an answer to some
of the questions raised throughout the article.
First of all, the appropriateness of the visualization of a certain concept
is supported, to a certain extent, by the legal XML mark-up of the privacy
policy, the metadata that provides information on the mark-up process (e.g.
annotations, exceptions, context, jurisdiction, etc.), and by the legal ontology
that organizes and formalizes the meanings of a certain legal domain, e.g. the
EU data protection laws. Semantic web technologies that tranform information
into a machine-readable format can, thus, provide a first layer of interpretation
of a certain legal domain knowledge in a specific context that can inform the
generation of visual cues. Even the specific legal framework can provide direction
for the visualization of legal notions: for instance, it is the GDPR to mandate
which information must be provided to the final user and how they must be
visualized (e.g. through icons, as opposed to other visual elements).
Moreover, interdisciplinary cooperation of (design, legal, technical) experts
and non-experts in the stage of transforming concepts into visualizations also
decreases the chances of personal bias [5]. The expertise of designers ensures
that the proposed visual representation is coherent with the lessons drawn by
previous experiments and good practices in legal design. Moreover, even the
sense-making process of end-users (typical message decoders) is considered: in
the first place, by involving them in the design phase; in the second place, by
carrying out user testing to establish whether a certain image is as evocative as
intended by those who created it; in the third place, by allowing customization
that depends on the users’ profile and on their interaction with the graphical
interface.
Nevertheless, there exist cases where these measures could be insufficient to
guarantee easy recognition. For instance, if the concepts are abstract, visualiza-
tions can be arbitrary and only their standardization and widespread adoption
can guide a correct interpretation. The same holds when the symbol is compre-
hensible, but it is the notion to which it refers to be unfamiliar for the addressee
(e.g. the notion of “pseudonymization”). In the best case, the image will be
sufficiently evocative to suggest its meaning and create a preunderstanding in
 From Words to Images Through Legal Visualization 83

the user’s mind. Should this not be the case, this can be solved only through
education to privacy and data protection.

6 Conclusions and Future Work

In this position paper, the topic of interpretation of legal visualizations has been
discussed. We have illustrated the shortcomings of traditional legal communi-
cation and, on the contrary, the many advantages that the visualization of the
law grants. However, the pictorial representation of concepts opens up multiple
interpretations and, in the legal domain, this ambiguity must be faced. This is
why we have suggested an approach based on Semantic Web technologies and
human-centered design that aims to guarantee representativeness of privacy pic-
torial representations to the highest possible degree, also through engagement
and participation of multiple stakeholders in the design phase. Equally, some
practical considerations on the role of interaction to realize customization and
tailored visualizations have been proposed to address individual differences. In
conclusion, the generation of one single “right” image is not desirable: there
might be more than one easily intepretable picture, but it depends on the con-
text, the type of representation, the background of the decoder, etc. The pro-
posed methodology is being currently implemented and integrated with best
practices already in use, in order to test whether it can represent a solution for
the generation and interpretation of legal visualizations.

References
1. Alexy, R.: Interpretazione giuridica. In: Enciclopedia delle scienze sociali. Treccani
(1996)
2. Article 29 Data Protection Working Party: Guidelines on transparency under reg-
ulation 2016/679, 17/EN WP260, December 2017
3. Athan, T., Governatori, G., Palmirani, M., Paschke, A., Wyner, A.: LegalRuleML:
design principles and foundations. In: Faber, W., Paschke, A. (eds.) Reasoning
Web 2015. LNCS, vol. 9203, pp. 151–188. Springer, Cham (2015). https://doi.org/
10.1007/978-3-319-21768-0 6
4. Barabucci, G., Cervone, L., Palmirani, M., Peroni, S., Vitali, F.: Multi-layer
markup and ontological structures in Akoma Ntoso. In: Casanovas, P., Pagallo,
U., Sartor, G., Ajani, G. (eds.) AICOL -2009. LNCS (LNAI), vol. 6237, pp. 133–
149. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-16524-5 9
5. Berger-Walliser, G., Barton, T.D., Haapio, H.: From visualization to legal design:
a collaborative and creative process. Am. Bus. Law J. 54(2), 347–392 (2017)
6. Boehme-Nessler, V.: Pictorial Law: Modern Law and the Power of Pictures.
Springer, Berlin (2010). https://doi.org/10.1007/978-3-642-11889-0
7. Brunschwig, C.: Visualisierung von Rechtsnormen: legal design. Ph.D. thesis, Uni-
versity of Zürich (2001)
8. Brunschwig, C.R.: On visual law: visual legal communication practices and their
scholarly exploration. In: Schweihofer, E., et al. (eds.) Zeichen und Zauber des
Rechts: Festschrift für Friedrich Lachmayer, pp. 899–933. Editions Weblaw, Bern
(2014)
84 A. Rossi and M. Palmirani

9. Canale, D.: La precomprensione dell’interprete è arbitraria? Etica Politica 1, 1–42

(2006)
10. Curtotti, M., Haapio, H., Passera, S.: Interdisciplinary cooperation in legal design
and communication. In: Schweighofer, E., et al. (eds.) Co-operation. Proceedings
of the 18th International Legal Informatics Symposium IRIS, pp. 455–462 (2015)
11. Curtotti, M., McCreath, E.: Enhancing the visualization of law. In: Law via the
Internet Twentieth Anniversary Conference, Cornell University (2012)
12. Esayas, S., Mahler, T., McGillivray, K.: Is a picture worth a thousand terms?
Visualising contract terms and data protection requirements for cloud computing
users. In: Casteleyn, S., Dolog, P., Pautasso, C. (eds.) ICWE 2016. LNCS, vol. 9881,
pp. 39–56. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-46963-8 4
13. European Parliament, Council of European Union: Regulation (EU) 2016/679 of
the European Parliament and of the Council of 27 April 2016 on the protection
of natural persons with regard to the processing of personal data and on the free
movement of such data, and repealing Directive 95/46/EC (General Data Protec-
tion Regulation). O.J. L 119, 4 May 2016, pp. 1–88 (2016)
14. European Parliament. Committee on Civil Liberties, Justice and Home Affairs:
Draft report on the proposal for a regulation of the european parliament and of the
council on the protection of individuals with regard to the processing of personal
data and on the free movement of such data (general data protection regulation)
[com(2012) 0011 - c7–0025/2012 - 2012/0011 (cod)], 21 November 2013
15. Geist, A., Brunschwig, C., Lachmayer, F., Schefbeck, G.: Multisensory law and
legal informatics: a comparison of how these legal disciplines relate to visual law. In:
Strukturierung der Juristischen Semantik - Structuring Legal Semantics. Editions
Weblaw (2011)
16. Haapio, H., Hagan, M., Palmirani, M., Rossi, A.: Legal design patterns for privacy.
In: Schweighofer, E. (ed.) Data Protection/LegalTech Proceedings of the 21st Inter-
national Legal Informatics Symposium IRIS 2018, pp. 445–450. Editions Weblaw,
Berlin (2018)
17. Hagan, M.: User-centered privacy communication design. In: Twelfth Symposium
on Usable Privacy and Security (SOUPS 2016). USENIX Association (2016)
18. Hagan, M.: Law by design (2017). http://www.lawbydesign.co
19. Holtz, L.-E., Nocun, K., Hansen, M.: Towards displaying privacy information with
icons. In: Fischer-Hübner, S., Duquenoy, P., Hansen, M., Leenes, R., Zhang, G.
(eds.) Privacy and Identity 2010. IAICT, vol. 352, pp. 338–348. Springer, Heidel-
berg (2011). https://doi.org/10.1007/978-3-642-20769-3 27
20. Jakobson, R.: Closing statement: linguistics and poetics. In: Style in Language, pp.
350–377 (1960)
21. Legal Design Lab: Design workshop for eu general data protection regulation,
July 2017. http://www.legaltechdesign.com/design-workshop-for-eu-general-data-
protection-regulation/
22. Palmirani, M., Cervone, L.: Measuring the complexity of the legal order over time.
In: Casanovas, P., Pagallo, U., Palmirani, M., Sartor, G. (eds.) AICOL -2013.
LNCS (LNAI), vol. 8929, pp. 82–99. Springer, Heidelberg (2014). https://doi.org/
10.1007/978-3-662-45960-7 7
23. Palmirani, M., Rossi, A., Martoni, M., Hagan, A.: A methodological framework to
design a machine-readable privacy icon set. In: Schweighofer, E. (ed.) Data Protec-
tion/LegalTech Proceedings of the 21st International Legal Informatics Symposium
IRIS 2018, pp. 451–454. Editions Weblaw, Wien (2018)
 From Words to Images Through Legal Visualization 85

24. Palmirani, M., Vitali, F.: Akoma-ntoso for legal documents. In: Sartor, G., Palmi-
rani, M., Francesconi, E., Biasiotti, M. (eds.) Legislative XML for the semantic
Web. LGTS, vol. 4, pp. 75–100. Springer, Dordrecht (2011)
25. Palmirani, M., Vitali, F.: Legislative XML: principles and technical tools. Inter-
American Development Bank (2012)
26. Passera, S.: Beyond the wall of text: how information design can make contracts
user-friendly. In: Marcus, A. (ed.) DUXU 2015. LNCS, vol. 9187, pp. 341–352.
Springer, Cham (2015). https://doi.org/10.1007/978-3-319-20898-5 33
27. Passera, S.: Flowcharts, swimlanes, and timelines. J. Bus. Tech. Commun. 32,
229–272 (2017)
28. Passera, S., Haapio, H.: Transforming contracts from legal rules to user-centered
communication tools: a human-information interaction challenge. Commun. Des.
Q. Rev. 1(3), 38–45 (2013)
29. Passera, S., Haapio, H., Curtotti, M.: Making the meaning of contracts visible-
automating contract visualization. In: Proceedings of the 17th International Legal
Informatics Symposium IRIS 2014 (2014)
30. Pettersson, J.S.: A brief evaluation of icons in the first reading of the european
parliament on COM (2012) 0011. In: Camenisch, J., Fischer-Hübner, S., Hansen,
M. (eds.) Privacy and Identity 2014. IAICT, vol. 457, pp. 125–135. Springer, Cham
(2015). https://doi.org/10.1007/978-3-319-18621-4 9
31. Rossi, A., Palmirani, M.: A visualization approach for adaptive consent in the
European data protection framework. In: Parycek, P., Edelmann, N. (eds.) CeDEM
2017: Proceedings of the 7th International Conference for E-Democracy and Open
Government, pp. 159–170. Edition Donau-Universität Krems, Krems (2017)
32. Sartor, G., Palmirani, M., Francesconi, E., Biasiotti, M.A.: Legislative XML for
the Semantic Web: Principles, Models, Standards for Document Management, vol.
4. Springer, Dordrecht (2011). https://doi.org/10.1007/978-94-007-1887-6
33. Schaub, F., Balebako, R., Durity, A.L., Cranor, L.F.: A design space for effective
privacy notices. In: Eleventh Symposium On Usable Privacy and Security (SOUPS
2015), pp. 1–17 (2015)
34. Schuler, D., Namioka, A.: Participatory Design: Principles and Practices. CRC
Press, Boca Raton (1993)
35. de Souza, C.S.: Semiotic engineering: bringing designers and users together at
interaction time. Interact. Comput. 17(3), 317–341 (2005)
36. de Souza, C.S.: Semiotics. In: Soegaard, M., Dam, R.F. (eds.) The Encyclopedia
of Human-Computer Interaction, 2nd edn. The Interaction Design Foundation,
Aarhus (2013)
37. TNS Opinion & Social: Special Eurobarometer 431 Data Protection. Technical
report, European Commission, Directorate-General for Justice and Consumers,
Directorate-General for Communication (2015)
38. Viola, F., Zaccaria, G.: Diritto e interpretazione: lineamenti di teoria ermeneutica
del diritto, Laterza (2013)
Rules and Norms Analysis and
Representation
 A Petri Net-Based Notation
for Normative Modeling: Evaluation
on Deontic Paradoxes

Giovanni Sileno1,2(B) , Alexander Boer1 , and Tom van Engers1

1
Leibniz Center for Law, University of Amsterdam, Amsterdam, The Netherlands
g.sileno@uva.nl
2
LTCI, Télécom ParisTech, Université Paris-Saclay, Paris, France

Abstract. Developing systems operating in alignment with norms is

not a straightforward endeavour. Part of the problems derive from the
suggestion that law concerns a system of norms, which, in abstract, in
a fixed point in time, could be approached and expressed atemporally,
but, when it is contextualized and applied, it deals with a continuous
flow of events modifying the normative directives as well. The paper
presents an alternative approach to some of these problems, exemplified
by well-known deontic puzzles, by extending the Petri net notation, most
common in process modeling, to Logic Programming Petri Nets. The
resulting visual formalism represents in a integrated, yet distinct fashion,
procedural and declarative aspects of the system under study, including
normative ones.

Keywords: Contrary-to-duty · Deontic puzzles · Logic Programming

Petri nets · Normative modeling

1 Introduction
Puzzles and paradoxes are tools for bringing conceptualizations to their bound-
ary conditions, and are therefore relevant for testing formal notations. The deon-
tic logic community has devoted special attention to paradoxes constructed with
contrary-to-duty (CTD) structures (see e.g. [3,4,9,11,17,22]). These are “para-
doxes” because, although the normative statements look plausible in natural
language, when each sentence is formalized in standard deontic logic (SDL)—
the paper tiger of normative modeling—either the set of formulas is inconsistent,
or one of the formulas is a logical consequence of another formula (e.g. [2]).

Definition 1 (Contrary to Duty). A contrary-to-duty (CTD) structure is a

situation in which a primary obligation exists, and with its violation, a secondary
obligation comes into existence.

The importance of CTDs lies in more than just their theoretical aspects:
CTDs are fundamental to normative modeling, because they are at the base
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 89–104, 2018.
https://doi.org/10.1007/978-3-030-00178-0_6
90 G. Sileno et al.

of compensatory norms, prototypical in e.g. contracts. The problem carries def-

inite applicative concerns. This intuitively simple structure produces complex
structures of obligations, prohibitions or permissions applying to sequences of
violations or satisfactions relative to the conduct of agents in a regulated social
system. Moreover, the secondary obligation may directly contradict the primary
obligation. In this case, the conflict has to be solved to decide a course of action
(i.e. “I am obliged, but I am forbidden, so what should I do?”). This consid-
eration highlights the problem of specifying and treating preferences between
idealities, or, in more agentive terms, priorities between commitments.
In this work, we focus on testing CTD structures on Logic Programming
Petri Nets (LPPN).1 This modeling notation has been introduced in [25] with
the purpose of integrating and distinguishing in the same visual representa-
tion declarative aspects (concerning terminology, ontological constraints, nor-
mative directives, etc.) and procedural aspects (mechanisms, processes, courses
of actions, etc.) of the reference system. The resulting common representational
ground is proposed as a basis to support a continuous re-alignment in admin-
istrative organizations of representations of law (norms), of implementations
of law (services as business processes), and of action (behavioural scripts, use
cases, possibly intentionally characterized). The notation aims therefore to cover
a wider class of models (business processes embedded with normative positions,
scenarios issued from narratives, agent scripts, etc.) than what usually studied
by deontic logic. Furthermore, beside its visual power (in principle increasing
its accessibility), it enjoys computational properties as distributed computation,
i.e. it does not require the reference to a global state.2
The connection of normative modeling with Petri nets is not new; see e.g.
[18,19], and more recently [23]; however, these works mainly focus on events
and factual conditions, overlooking normative characterizations. Other works,
such as [24], have proposed using Petri nets to formalize contracts; with respect
to these, the present proposal is more specific, as it focuses on minimal CTDs
(through the lens of deontic puzzles), but also more general, as it considers the
inclusion of declarative bindings in the model.
The paper is organized as follows. In Sect. 2, we will present Logic Program-
ming Petri Nets (LPPN), delineating the notation and introducing informally
a simplified version of its semantics. In Sect. 3, we will review a series of exam-
ples, all, save the first, copied from the literature. For each of them, we will
propose models of the corresponding scenarios, and attempt to clarify part of
the issues encountered in SDL. Discussion and further developments end the
paper. A formalization of the propositional version of LPPN can be found in the
appendix.

1
Prototypes of LPPN interpreters are available on http://github.com/s1l3n0/pypneu
and http://github.com/s1l3n0/lppneu.
2
Cf. the recent extension to standard deontic logic by Gabbay and Straßer [6] integrat-
ing reactive constructs, an approach in many aspects dual to the present proposal.
 A Petri Net-Based Notation for Normative Modeling 91

2 Logic Programming Petri Nets

Petri nets are a simple, yet effective computational modeling representation fea-
turing an intuitive visualization (see Fig. 1). They consist in directed, bipartite
graphs with two types of nodes: places (visually represented with circles) and
transitions (with boxes). A place can be connected only to transitions and vice-
versa. One or more tokens (dots) can reside in each place. The execution of Petri
nets is also named “token game”: transitions fire by consuming tokens from their
input places and producing tokens in their output places.3

t1 t1 t1

p1 p3 p1 p3 p1 p3

p2 p2 p2

(a) not enabled transition, (b) enabled transition and (c) the transition has fired
before firing firing

Fig. 1. Example of a Petri net and of its execution (but also of a LPPN procedural
component when labels are propositions).

Despite their widespread use in computer science, electronics, business pro-

cess modeling and biology, Petri nets are generally considered not to be enough
expressive for reasoning purposes; in effect, they do not refer explicitly to any
informational or representational concept. In their simplest form, tokens are
indistinct, and do not transport any data. Nevertheless, usually modelers intro-
duce labels to set up a correspondence between the modeling entities and the
modeled entities. This practice enables them to read the results of a model exe-
cution in reference to the modeled system, and therefore it becomes functional
to the use of the notation, although it is not a requirement for the execution in
itself. Further interaction is possible if these labels are processed according an
additional formalism, as for instance with the Coloured Petri Net (CPN) nota-
tion [13], which, for many aspects, is a descendant of Predicate/Transition Nets
[7]. If its expressiveness and wide application provide reasons for its adoption,
the CPN notation introduces many details which are unimportant in our setting
(e.g. expressions on arcs); more importantly, it still misses the requirement of
processing declarative bindings, necessary, for instance, to model terminological
relationships. We opted therefore for an alternative notation.
Whereas Petri nets specify procedural mechanisms, LPPNs extend those
(a) with Prolog-like literals as labels, attached on places and transitions; (b)
with nodes specifying (logic) declarative bindings on places and on transitions.
3
For an overview on the general properties of Petri nets see e.g. [21].
92 G. Sileno et al.

The notation builds upon the intuition that places and transitions mirror the
common-sense distinction between objects and events (e.g. [1]), roughly reflect-
ing the use of noun/verb categories in language [14]: the procedural components
can be used to model transient aspects of the system in focus; the declarative
components to model steady state aspects, i.e. those on which the transient is
irrelevant or does not make sense (e.g. terminology, ontological constraints, etc.).

p9

p7 t3 p10

AND IMPLIES t2 p11

p4 p6
p8 t4
p5

(a) (b)

Fig. 2. Examples of LPPN declarative components: (a) defined on places, correspond-

ing to the Prolog/ASP code: p6 :- p4, p5. p5. (b) defined on transitions, instanta-
neously propagating the firing where possible (the IMPLIES label on black circles is left
implicit).

In this paper, for simplicity, we will consider only propositional labeling; with
this assumption, the execution model of the LPPN procedural component is the
same of Condition/Event nets, i.e. Petri nets whose places are not allowed to
contain more than one token. For this reason, the Petri net in Fig. 1 can be inter-
preted as an example of LPPN specifying a procedural mechanism. However, the
LPPN notation introduces also logic operator nodes (or l-nodes), which apply
on places or on transitions. An example of a sub-net with l-nodes for places
(small black squares) is given in Fig. 2a. These are used to create logic compo-
sitions of places (via operators as NEG, AND, OR, etc.), or to specify logic inter-
dependencies (via the logic conditional IMPLIES). Similarly, transitions may be
connected declaratively via l-nodes for transitions (black circles) as in Fig. 2b.
These connections may be interpreted as channels enabling instantaneous prop-
agation of firing. In this case, it is not relevant to introduce operators as AND
because, for the interleaving semantics, only one source transition may fire per
step. To simplify the visual burden, we might leave the IMPLIES label implicit,
exploiting the sense of the arrow to specify the direction of the relation. Oper-
ationally, these declarative components are treated integrating the stable model
semantics used in answer set programming (ASP) [15]. This was a natural choice
because process execution exhibits a prototypical ‘forward’ nature, and ASP can
be interpreted as providing forward chaining. A formalization of propositional
LPPNs can be found in the Appendix.
 A Petri Net-Based Notation for Normative Modeling 93

suspension

recognize recognize recognize recognize

violation initiation violation initiation
prohibition
to cross
prohibition
to cross
recognize recognize
satisfaction satisfaction
obligation obligation
to cross to cross

start crossing end start crossing end

crossing crossing crossing crossing

(a) (b)

Fig. 3. A minimal, conflicting and event-based contrary-to-duty (CTD) structure: a

secondary obligation is created after the violation of a primary obligation, and in
conflict with it. The CTD may be interpreted: (a) as an exception; or (b) as overriding
the primary obligation. The second is the most accepted option.

3 Deontic Exercises
3.1 Crossing or Not Crossing?
Let us start from this minimal, conflicting CTD structure:

You are forbidden to cross the road.

If you are crossing the road, (you have to) cross the road!

This rule of conduct is perfectly plausible: most parents say something similar to
their children at some moment. However, its translation in basic deontic logic is
not direct. The text suggests, in effect, an underlying model in terms of action: a
state-based interpretation would miss the implicit initiation/termination events
that make the action-wise prescription sound, and namely:
You are forbidden to cross the road.
If you have started to cross the road, you are obliged to finish crossing.
This transitional aspect can be easily mapped on a LPPN, separating the expe-
riential world from the institutional world, with the second synchronized to the
first via constituting links determining what counts as violation or satisfaction.4
4
With respect to constitutive rules, the LPPN notation enables to easily distinguish
classificatory constitutive rules (e.g. “a bike counts as a vehicle”) from constitutive
event rules (e.g. “raising a hand counts as making a bid”), as they are modeled
respectively using black boxes or black circles. Most formalizations of constitutive
rules consider only on the first aspect (e.g. [10]), cf. the overview in [27].
94 G. Sileno et al.

In principle, two modeling options are available in regard to the secondary

obligation; it can interpreted:

– as an exception, thus temporarily retracting the primary obligation (Fig. 3a);

– as overriding the primary obligation, which persists concurrently (Fig. 3b).

The second option requires an additional treatment, because it brings two con-
trary/opposite positions to hold concurrently. Similarly to what suggested in the
literature, this can be solved introducing an explicit ordering between positions,
which depends on how close to ideal is the world/context they are referring to
(see e.g. [16,22]). In the proposed Petri net an aspect of this mechanism—the
fact that the secondary obligation is put in force in response to the violation of
the primary one—is already reified in the topology. To capture the remaining
part, i.e. that the second is contextually overriding the first, we need to order
them in the opposite sense: the last obligation created is the one with most
priority and should be the only active, suspending the previous ones. This can
be done introducing an inhibiting arc (visualized in Fig. 3b as an arrow with a
circle-shaped head).5 The resulting design can be seen as a model of salience.

prohibition
to kill suspension

initiation
violation
obligation
to kill
gently
satisfaction violation

kill
gently kill not
gently

start killing
killing

kill

Fig. 4. Gentle murderer case.

5
Inhibiting arcs goes from places to transitions. If the input place of an inhibiting arc
is occupied, its output transition is disabled.
 A Petri Net-Based Notation for Normative Modeling 95

3.2 Gentle Murderer

The previous CTD model gives us the basic instruments to proceed. Let us start
from the classic case of the “gentle murderer”, given by Forrester [5]:

It is forbidden to kill,
but if one kills, one ought to kill gently.

This example is very similar to the previous one, except that the target of the
secondary obligation is subsumed by the target of the first one. Because our nota-
tion explicitly accounts for a declarative dimension for events, we can directly
map this relation (Fig. 4).

suspension

permission prohibition
fence fence initiation

obligation
exception white fence
violation

cottage
by the sea satisfaction violation

white not white

default
cottage fence fence
not by
the sea
fence

Fig. 5. White fence case.

3.3 White Fence

Now, we consider a static and extended variation proposed by Prakken and

Sergot [22], the “white fence” case:
There must be no fence.
If there is a fence, it must be a white fence.
If the cottage is by the sea, there may be a fence.

This example shows the importance of distinguishing exceptions from overriding

effects due to CTDs (and therefore supports the second interpretation of CTD
given in Sect. 3.1). In principle, a rule specifies a CTD if its premise is the
negation of the target of the obligation in the consequent of another rule. A rule
specifies an exception if it has as consequent the negation of the consequent of
another rule, and it has a lower priority than the first one (exceptions are by
definition subordinate to some normal conditions). In effect, the two rules can
96 G. Sileno et al.

be read as referring to a priority-based representation [26]. Considering part of

the “white fence” case in propositional form, we have:

Forb(fence) sea → Perm(fence)

which can be translated to the corresponding constraint-based representation:

¬sea → Forb(fence) sea → Perm(fence)

This treatment gives a hint as to how to deal with exceptions—that is, it helps
make explicit an enchaining of negations of the premises following the inverse
ordering of salience. The fastest solution to avoiding conflicts in the case of belief
revision is to not reify directly the default position (in this case, prohibition
against having a fence), but to generate it through a default rule [26]:

not sea → Forb(fence)

The resulting model is illustrated in Fig. 5.

Perm(D) Forb(D)

suspension

default Forb(A) initiation

Perm(A)

Obl (B)
exception violation
satisfaction violation

C A
monitor B

B not B

Fig. 6. Privacy act case.

3.4 Privacy Act

Recently, Governatori [8] has proposed the case of a Privacy Act (fictional, but
based on actual Australian normative provisions):
i. The collection of personal information is forbidden, unless acting on a court
order authorising it.
 A Petri Net-Based Notation for Normative Modeling 97

ii. The destruction of illegally collected personal information before accessing it

is a defence against the illegal collection of the personal information.
iii. The collection of medical information is forbidden, unless the entity collecting
the medical information is permitted to collect personal information.
The following deontic interpretation is proposed:
i. Forbidden A. If C, then Permitted A.
ii. If Forbidden A and A, then Obligatory B.
iii. Forbidden D. If Permitted A, then Permitted D.
A, B, C, and D in this specific case are actions; (ii) specifies a CTD, (i) and
(iii) provide rules based on a priority-based representation. As before, we extract
explicitly the defaults. The negation of permission of A in (iii) can be interpreted
as the prohibition of A, thus converging to the default in (i). For completeness,
we have reported in Fig. 6 the monitoring place from which the event B or not
B is recognized.

3.5 Detachment Principles

In the deontic logic literature, two types of “detachment principles” are recog-
nized as relevant. The first is called factual detachment (FD):

p ∧ Obl (q|p) → Obl (q) (1)

The second is known as deontic detachment (DD):

Obl (p) ∧ Obl (q|p) → Obl (q) (2)

In our framework, a conditional directive or commitment is seen as a sus-

ceptibility to a condition that creates or implies the directive depending upon
whether the connective is a causal or logical dependence. The two principles can
be then translated using the LPPN notation. Focusing on the logical dependence
case, the result is seen in Fig. 7. The pictures show that the first principle is sat-
isfied by the notation semantics; on the contrary, the second principle, which is
based on an anticipation of the normal conditional, is not satisfied.

3.6 Derived Obligation

Consider these two sentences:
– Bob’s promise to meet you commits him to meeting you.
– It is obligatory that if Bob promises to meet you, he does so.
Although in natural language their difference is arguable, in the literature they
have been formalized using two distinct deontic formulations:
– p → Obl (m)
– Obl (p → m)
98 G. Sileno et al.

Obl (q|p) p Obl (q|p) p Obl (p)

violation
initiation initiation satisfaction

Obl (q) Obl (q) not p

(a) (b)

Fig. 7. Factual and deontic detachments.

What is the difference between the two formulations in our framework? The
second formula can be translated with the consideration that the obligation of
something consists of two recognition rules about satisfaction and violation, by
default anchored respectively to something and to ¬something. As an object,
the conditional within the obligation can be transformed using the material
implication.6 The result are reported in Fig. 8. As we see in the picture, both
models are violated in the same situation (p and ¬m); however, the second
includes the recognition of a satisfied situation not accounted for in the first
(¬p). In other words, the first derived obligation precisely discriminates the
elements producing the violation. The second takes an explicit position also on
the satisfying elements.

3.7 Chisholm’s Paradox

At this point, we can finally model the “paradox” proposed by Chisholm [4]:

It ought to be that Jones goes (to the assistance of his neighbors).

It ought to be that if Jones goes, then he tells them he is coming.
If Jones doesn’t go, then he ought not tell them he is coming.
Jones doesn’t go.

This was seen as a paradox, because if we model it as:

i. Obl (go)
ii. Obl (go → tell )
iii. ¬go → Forb(tell )
iv. ¬go

6
The specific example from which we started is not based on a logic conditional, but
on a causal connective, at least in the case of “if Bob promises to meet you, then
he does so”. In this case, the use of material implication is not a perfect fit, as the
temporal shift between the promise and the meeting falsifies the derived constraint,
at least on a transient basis. On a steady state analysis, however, this simplification
may be applied.
 A Petri Net-Based Notation for Normative Modeling 99

p Obl (p → m)

initiation satisfaction violation

Obl (m)
p→m p ∧ ¬m

satisfaction violation OR

m not m ¬p m

(a) (b)

Fig. 8. Derived obligation modeled as p → Obl (m) or as Obl (p → m).

and we apply both deontic and factual detachments, we find an inconsistency.

More precisely, from (i) and (ii), using of deontic detachment, we derive Obl (tell ),
while from (iii) and (iv), using factual detachment, we derive Obl (¬tell ). How-
ever, representing this model using our notation as in Fig. 9, we do not find any
specific issue. The fact ¬go satisfies Obl (go → tell ), but violates Obl (go), and
for this reason, Forb(tell ) is instantiated. Depending on whether tell becomes
true, this may or may not be violated.

Obl (go)

satisfaction violation Forb(tell)

satisfaction violation
go ¬go
Obl (go → tell)
¬tell tell
go ∧ ¬tell
satisfaction
violation AND

go → tell

OR

Fig. 9. Chisholm’s paradox case.

100 G. Sileno et al.

prohibition
initiation dog initiation

prohibition obligation
satisfaction violation
warning sign warning sign

no dog dog

Fig. 10. Residential neighbourhood case.

3.8 Residential Neighbourhood

As we observed above, the difference in natural language between the two types
of derived obligations is arguable. In order to show what would happen if we
interpret the two as conditional obligations, we consider the non-agent version
of the Chisholm’s paradox proposed by Prakken and Sergot [22]:
There must be no dog.
If there is no dog, there must be no warning sign.
If there is a dog, there must be a warning sign.
There is a dog.
Written in this way, the modeling is straightforward (Fig. 10).

4 Discussion and Further Developments

At superficial level, the paper presented examples of application of a notation
introduced for wider modeling purposes [25]. This served as an important exer-
cise to (partially) evaluate its practical functionality, as one may capture the
subtle problems related to modeling in a certain domain (in this case normative
modeling) only by approaching the specific issues that have been raised by prac-
titioners and scholars in that domain. The introduction of defaults, exceptions,
and suspensions in the notation presented here is a preliminary proposal. Some
of these aspects have been treated more in more detail elsewhere (e.g. defaults in
priority-based rule-bases [26], proceeding along [12]), while others require further
investigation (e.g. suspension, cf. [20]).
Nevertheless, the exercise yields concrete practical and theoretical results.
From a practical point of view, it makes a case supporting normative modeling
with notations similar to those used in business process modeling, thus poten-
tially facilitating cross-fertilization between theoretical to operational settings.
From a theoretical point of view, we observed for instance that our conceptual
framework does not entail the deontic detachment principle, hinting at a more
general minimal commitment taken by the notation. In other words, i.e. the nota-
tion does not provide any rule a priori to conclude whether Obl (A) ∧ Obl (B) is
the same as Obl (A ∧ B). This neutral starting point can be used to evaluate the
 A Petri Net-Based Notation for Normative Modeling 101

alternative impact of different axioms proposed in the literature, in affinity with

approaches like input/output logic [17].
More importantly, while working on these exercises, we appreciated the cru-
cial interplay between static and dynamic aspects (one of the issues underlying
many deontic puzzles). The LPPN notation, requiring the explicitation of proce-
dural and declarative aspects, highly facilitated this task, but our exercise sug-
gests further research on the modeling methodology. For instance, in Chisholm’s
paradox, we considered go and tell as labels of places, but strictly speaking,
they should be attached to transitions (as in the privacy act case). With this
choice, we would require making explicit the occurrence of the events as places,
in order to evaluate the material implication. Does this simplification hint at a
more general pattern? We modeled the nodes concerning violation and satisfac-
tion as l-nodes only when their transformational nature was certain. In general,
however, they may be transformational (when they simply identify whether a
violation or satisfaction holds at the moment) or reactive (when they reify the
fact that a violation occurred in that moment, or when they cause any change
on the inputs, e.g. removing the obligation)—note again the interplay between
static and dynamic aspects. In the future, an investigation of boundary cases
may help in formulating a more general theory on how to decide upon the level
of abstraction, and about whether it is possible to identify or elaborate general
patterns, that, depending on circumstances, are read in one form or in the other.

A Formalization

Here we present a simplified version of the LPPN notation considering only

a propositional labeling. We start from the definition of propositional literals
derived from ASP [15], accounting for strong and default negation.

Definition 2 (Literal and Extended literals). Given a set of propositional

atoms A, the set of literals L = L+ ∪ L− consists of positive literals (atoms)
L+ = A, negative literals (negated atoms) L− = {−a | a ∈ A}, where ‘−’ stands
for strong negation.7 The set of extended literals L∗ = L ∪ Lnot consists of
literals and default negation literals Lnot = {notl| l ∈ L}, where ‘not’ stands for
default negation.8

We denote the basic topology of a Petri net as a procedural net.

Definition 3 (Procedural net). A procedural net is a bipartite directed

graph connecting two finite sets of nodes, called places and transitions. It can
be written as N = P, T, E, where P = {p1 , . . . , pn } is the set of place nodes;
T = {t1 , . . . , tm } is the set of transition nodes; E = E + ∪ E − is the set of arcs
connecting them: E + from transitions to places, E − from places to transitions.

7
Strong negation is used to reify an explicitly false situation (e.g. “It does not rain”).
8
Default negation is used to reify a situation in which something cannot be
retrieved/inferred (e.g. ‘It is unknown whether it rains or not’).
102 G. Sileno et al.

LPPNs consists of three components: a procedural net specifying causal or

temporal relationships, and two declarative nets specifying respectively logical
dependencies at the level of objects or ongoing events (on places), and on impulse
events (on transitions). Furthermore, propositional LPPNs build upon a boolean
marking on places (like condition/event nets).

Definition 4 (Propositional Logic Programming Petri Net). A proposi-

tional Logic Programming Petri Net LPPN prop is a Petri Net whose places and
transitions are labeled with literals, enriched with declarative nets of places and
of transitions. It is defined by the following components:

– P, T, PE  is a procedural net; PE stands for procedural edges;

– CP : P → L∗ and CT : T → L are labeling functions, associating literals
respectively to places and to transitions;
– OP = {¬, −, ∧, ∨, →, ↔, . . .} is a set of logic operators.
– LP and LT are sets of logic operator nodes (in the following called l-nodes)
respectively for places and for transitions.
– CLP : LP → OP maps each l-node for places to a logic operator; similarly,
CLT : LT → OP does the same for l-nodes for transitions.
−
– DE LP = DE + LP ∪ DE LP is the set of arcs connecting l-nodes for places to
−
places; similarly, DE LT = DE + LT ∪ DE LT for l-nodes for transitions and
transitions.9
– M : P → {0, 1} returns the marking of a place, i.e. whether the place contains
(1) or does not contain (0) a token.

Note that if LP ∪LT = ∅, we have a strictly procedural LPPN prop , i.e. a standard
binary Petri net. If T = ∅, we have a strictly declarative LPPN prop , that can be
directly mapped to an ASP program.
With respect to the operational semantics, the execution cycle of a LPPN
consists of four steps: (1) given a “source” marking M , the bindings of the declar-
ative net of places entail a “ground” marking M ∗ ; (2) an enabled transition is
selected to pre-fire, so determining a “source” transition-event e; (3) the bindings
of the declarative net of transitions entail all propagations of this event, obtain-
ing a set of transition-events, also denoted as the “ground” event-marking E ∗ ;
(4) all transition-events are fired, producing and consuming the relative tokens.
The steps (1) and (3) are processed by an ASP solver: the declarative net of
places (respectively transitions) is translated as rules, tokens (transition-events)
are reified as facts; the ASP solver takes as input the resulting program and, if
satisfiable, it provides as output one or more ground marking (one or more sets
transition-events to be fired). For the steps (2) and (4), the operational seman-
tics distinguishes the external firings (started by the execution) from the internal
firing, immediately propagated (triggered by the declarative net of transitions).

9
Note that DE −
LT ⊆ (T ∪ P ) × LT , i.e. these edges go from transitions and places
(modeling contextual conditions) to l-nodes for transitions.
 A Petri Net-Based Notation for Normative Modeling 103

Definition 5 (Enabled transition). A transition t is enabled in a ground

marking M ∗ if a token is available for each input places:

Enabled (t) ≡ ∀pi ∈ •t, M ∗ (p) = 1

Similarly to what marking is for places, we consider an event-marking for tran-

sitions E : T → {0, 1}. E(t) = 1 if the transition t produces a transition-event
e. Each step s has a “source” event-marking E.
Definition 6 (Pre-firing). An enabled transition t pre-fires at a step s if it
selected to produce a transition-event:

∀t ∈ Enabled (T ) : t pre-fires ≡ E(t) = 1

As we apply an interleaving semantics for the pre-firing, the interpreter selects

only one transition to pre-fire per step; for any other t , E(t ) = 0.

Definition 7 (Firing). An enabled transition t fires by propagation, consuming

a token from each input place, and forging a token in each output place:

∀t ∈ Enabled (T ) : t fires ≡
E ∗ (t) = 1 ↔ ∀pi ∈ •t : M  (pi ) = 0 ∧ ∀po ∈ t• : M  (po ) = 1

References
1. Breuker, J., Hoekstra, R.: Core concepts of law: taking common-sense seriously.
In: Proceedings of Formal Ontologies in Information (2004)
2. Broersen, J., van der Torre, L.: Ten problems of deontic logic and normative rea-
soning in computer science. In: Bezhanishvili, N., Goranko, V. (eds.) ESSLLI 2010-
2011. LNCS, vol. 7388, pp. 55–88. Springer, Heidelberg (2012). https://doi.org/10.
1007/978-3-642-31485-8 2
3. Carmo, J., Jones, A.: Deontic logic and contrary-to-duties. In: Gabbay, D.M.,
Guenthner, F. (eds.) Handbook of Philosophical Logic, vol. 8, pp. 265–343.
Springer, Dordrecht (2002). https://doi.org/10.1007/978-94-010-0387-2 4
4. Chisholm, R.M.: Contrary-to-duty imperatives and deontic logic. Analysis 24(2),
33–36 (1963)
5. Forrester, J.W.: Gentle murder, or the adverbial Samaritan. J. Philos. 81(4), 193–
197 (1984)
6. Gabbay, D.M., Straßer, C.: Reactive standard deontic logic. J. Log. Comput. 25(1),
117–157 (2015)
7. Genrich, H.J.: Predicate/transition nets. In: Brauer, W., Reisig, W., Rozenberg,
G. (eds.) ACPN 1986. LNCS, vol. 254, pp. 207–247. Springer, Heidelberg (1987).
https://doi.org/10.1007/978-3-540-47919-2 9
8. Governatori, G.: Thou shalt is not you will. Technical report, NICTA (2015)
9. Governatori, G., Rotolo, A.: Logic of violations: a gentzen system for reasoning
with contrary-to-duty obligations. Australas. J. Log. 4, 193–215 (2006)
10. Grossi, D., Meyer, J.J.C., Dignum, F.: Classificatory aspects of counts-as: an anal-
ysis in modal logic. J. Log. Comput. 16(5), 613–643 (2006)
104 G. Sileno et al.

11. Hansen, J., Pigozzi, G., Van Der Torre, L.: Ten philosophical problems in deontic
logic. Normative Multi-agent, pp. 1–26 (2007)
12. Horty, J.F.: Rules and reasons in the theory of precedent. Legal Theory 17(1),
1–33 (2011)
13. Jensen, K.: Coloured Petri Nets: Basic Concepts, Analysis Methods and Practical
Use. Springer, Heidelberg (1996). https://doi.org/10.1007/978-3-662-03241-1
14. Kemmerer, D., Eggleston, A.: Nouns and verbs in the brain: implications of lin-
guistic typology for cognitive neuroscience. Lingua 120(12), 2686–2690 (2010)
15. Lifschitz, V.: What is answer set programming? In: Proceedings of the 22th AAAI
Conference on Artificial Intelligence (2008)
16. Makinson, D.: Five faces of minimality. Stud. Log. 52, 339–379 (1993)
17. Makinson, D., Van Der Torre, L.: Input/output logics. J. Philos. Log. 29, 383–408
(2000)
18. Meldman, J., Fox, S.: Concise petri nets and their use in modeling the social work
(Scotland) Act 1968. Emory Law J. 30, 583–630 (1981)
19. Meldman, J., Holt, A.: Petri nets and legal systems. Jurimetr. J. 12(2), 65–75
(1971)
20. Meneguzzi, F., Telang, P., Singh, M.: A first-order formalization of commitments
and goals for planning. In: Proceedings of the 27th AAAI Conference on Artificial
Intelligence, pp. 697–703 (2013)
21. Murata, T.: Petri nets: properties, analysis and applications. Proc. IEEE 77(4),
541–580 (1989)
22. Prakken, H., Sergot, M.: Contrary-to-duty obligations. Stud. Log. 57(1), 91–115
(1996)
23. Purvis, M.A.: Dynamic modelling of legal processes with petri nets. Ph.D. thesis,
University of Otago (1998)
24. Raskin, J.F., Tan, Y.H., van der Torre, L.: How to model normative behavior in
Petri nets. In: Proceedings of the 2nd ModelAge: Workshop on Formal Models of
Agents, pp. 223–241 (1996)
25. Sileno, G.: Aligning law and action. Ph.D. thesis, University of Amsterdam (2016)
26. Sileno, G., Boer, A., van Engers, T.: A constructivist approach to rule bases. In:
Proceeding of the 7th International Conference on Agents and Artificial Intelligence
(ICAART 2015) (2015)
27. Sileno, G., Boer, A., van Engers, T.: Revisiting constitutive rules. In: Proceedings
of the 6th Workshop on Artificial Intelligence and the Complexity of Legal Systems
(AICOL 2015) (2015)
 Legal Patterns for Different Constitutive Rules

Marcello Ceci(&), Tom Butler, Leona O’Brien, and Firas Al Khalil

GRCTC, University College Cork, Cork, Ireland

{marcello.ceci,tbutler,leona.obrien,
firas.alkhalil}@ucc.ie

Abstract. The research for solutions for compliance is mainly focused on the
representation of regulative rules, i.e. the imperatives that the industry is asked
to comply to. Yet, a relevant part of the legal knowledge contained in regulation
cannot be expressed in terms of deontic statements, and is instead represented as
constitutive rules. This concept was first introduced by philosophers of language
such as J.L. Austin and J.R. Searle and further developed in legal philosophy,
where constitutive statements are classified in categories according to their legal
effects. The present paper presents a heuristic approach for the representation of
alethic statements as part of a methodology aimed at ensuring effective trans-
lation of the regulatory text into a machine-readable language. The approach is
based on a classification of constitutive statements contained in the work of legal
philosophers A.G. Conte and G. Carcaterra. The methodology includes an
intermediate language, accompanied by an XML persistence model, and intro-
duces a set of “legal concept patterns” to specifically represent the different
constitutive statements. The paper identifies five patterns for the corresponding
constitutive statements found in financial regulations: legal definitions, com-
mencement rules, amendments, relative necessities, and party to the law
statements.

Keywords: Constitutive rules
Legal patterns
Legal definitions

Meta-rules
SBVR

“Comprendere un diritto significa sapere che cosa esso è, sapere

che cosa è significa possederne la definizione.”
Understanding a right implies knowing what it is, and knowing what
it is implies possessing its definition.
Carcaterra [9, p. 25].

1 Introduction

Assessing compliance means checking the correspondence of an activity with a set of

norms. Here, norm is intended as the interpretation of a rule contained in a regulation.
To assess compliance, we need two pieces of information: one about the activity, and
the other about the interpretation of the regulation. In order to speed up the process of
compliance assessment, we thus need a machine-readable interpretation of the law.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 105–123, 2018.
https://doi.org/10.1007/978-3-030-00178-0_7
106 M. Ceci et al.

Great are the potentials of semantic web technologies to express semantics of legal
texts, especially in terms of legal references (e.g. LegalDocML [24]) and legal scope
(e.g. LKIF [8]). Several rule languages exist that manage legal rules, and rest on solid
logical foundations (e.g. LegalRuleML, see survey [13]). Unfortunately, those layers of
technology are difficult for lawyers to grasp, and the related solutions are still out of
their reach.
The goal of the research presented in this paper is to represent regulations for GRC
tasks in financial industry. To achieve this, we developed a Regulatory Interpretation
Methodology (RIM) to guide a Subject Matter Expert (SME, e.g. the legal expert) and a
Semantic Technology Engineer (STE) in a collaborative process of transformation of
the regulatory text into machine-readable information [1].
To represent the semantics of regulatory requirements in a machine-readable format
with a SME-friendly process we built an intermediate language based on SBVR
(Semantics of Business Vocabulary and Business Rules [25]). SBVR is a powerful
instrument for building a vocabulary representing business activities [34, p. 14], but
unfortunately it isn’t suitable – as is – for the representation of legal rules in a machine-
readable format: besides SBVR being designed for human-to-human communication
across a business and not for automatic reasoning, some of its components are falling
short in capturing legal concepts, such as constitutive rules.
Philosophy of language [3, 32] identified two types of rules: regulative rules and
constitutive rules. Previous work describes our approach to regulative rules [12]. The
present paper discusses the representation of constitutive rules, which, despite not
being requirements themselves, still play more than a marginal role in compliance
assessment.

1.1 Scope
The present paper focuses on the issue of representing alethic statements (including
legal definitions, meta-rules, statements of facts) in a machine-readable way. The
proposed solution employs an intermediate language based on SBVR and follows the
classification drawn in legal philosophy for constitutive statements. Because the
research is focused on compliance, constitutive statements are seen as complementary,
and represented only to the extent necessary to define and specify the effects of reg-
ulative statements. Legal philosophy identifies different categories of constitutive rules,
and this paper follows these classifications in order to represent their semantics. The
paper introduces the concept of legal concept patterns, that work as templates to
represent constitutive norms with fixed effects, and presents five legal concept patterns
for capturing five types of constitutive rules.
This paper is structured as follows: Sect. 2 introduces the concept of constitutive
rules and the relevant doctrine on them. Section 3 introduces legal concept patterns,
used to represent constitutive norms as explained in Sect. 4.
 Legal Patterns for Different Constitutive Rules 107

2 Constitutive Rules

In legal theory, constitutive norms1 are the result of declarative acts [6]. These norms
introduce new abstract classifications of existing facts and entities. Those classifications
are called institutional facts (e.g. marriage, money, private property) and they emerge
from an independent ontology of “brute” physical facts. Differently from regulative
rules, constitutive norms have no deontic content: they do not introduce obligations,
prohibitions or permissions. Instead, they typically take the following form:

ðaÞ counts as ðbÞ in context ðcÞ

In order to capture these rules, it is thus necessary to identify three elements:

– a material (or previously identified) phenomenon (token);
– an abstract concept that is created by the constitutive rule itself (type);
– a limited area of application (context).
Two additional elements, the (alethic) modality and the legal source, are in com-
mon with regulative rules. Constitutive norms are also called “determinative rules”
[18]. In LegalRuleML, that class is represented as ConstitutiveStatement node element.
In Normative Multi-Agent systems, they are formalized as belief rule of normative
agents: from a knowledge representation point of view, they behave as data abstraction
in programming languages [7]. An investigation of the logic underlying “count-as”
statements is performed by Grossi [20], including a wide survey of the existing
attempts to provide a formalization for constitutive norms and count-as conditionals
[21]. In it, the diversity of formal approaches is evident. Seven strands of research are
identified [21, p. 429]: contextual aspects of counts-as; classificatory aspects of counts-
as; counts-as and actions; counts-as and conventions; counts-as as grounded on dedi-
cated agents; counts-as as related to regulative norms; counts-as as related to the
definition of legal terms.
This multiplicity of approaches suggests that a unique formalization is not capable
of representing all the possible aspects of constitutive norms. Furthermore, the dis-
tinction [19] of counts-as statements in classificatory counts-as, proper classificatory

1
The concept of constitutivity, as distinguished from the regulative effects of norms, was first
introduced by John Rawls [27], with the following distinction: “justifying a practice and justifying a
particular action falling under it… [by meaning for] practice any form of activity specified by a
system of rules which defines offices, roles, moves, penalties, defenses, and so on, and which gives
the activity its structure”. Austin [3] investigated the phenomenon of the performative utterances,
defining them as: “Utterances […] that […] do not ‘describe’ or ‘report’ or constate anything at all,
are not ‘true or false,’ and the uttering of [which] is, or is a part of, the doing of an action, which
again would not normally be described as, or as ‘just,’ saying something” (pp. 5–6). The concept of
performative utterances was later refined by Searle [32] into that of speech acts and constitutive rules,
defined as follows: “[R]egulative rules regulate antecedently or independently existing forms of
behaviour […]. But constitutive norms do not merely regulate, they create or define new forms of
behaviour. The rules of football or chess, for example […] create the very possibility of playing such
games” (p. 33).
108 M. Ceci et al.

counts-as, ascriptive counts-as and constitutive counts-as, is a proof of the hetero-

geneity of such statements, not only in their form but also in their effects.
SBVR’s restricted language does not explicitly include institutional facts or con-
stitutive norms, although it has a rule category called structural (or definitional) rules
that can be used for the purpose. These rules are represented in the language through
alethic statements instead of deontic statements.
The rest of this section will present part of the research conducted in philosophy of
law on the subject of constitutive rules with the aim of identifying a taxonomy of
constitutive rules.

2.1 On the Constitutivity of Rules

Over the last decades the Italian school of legal philosophy introduced distinctions
within constitutive statements [9, 14, 16, 22]. According to Carcaterra [9, p. 61], the
constitutive statement “produces, at the very moment when it enters into force, the effect
that is its scope and content”. Carcaterra distinguishes two meanings of constitutivity in
legal norms: according to the first meaning, constitutivity is a process downright creator
of legal effects (or states of affairs). Rules carrying such constitutivity correspond to
Searle’s declarative speech acts [33, pp. 16–20]. In the second meaning, constitutivity is
the process of creation of acts and facts with a specific legal meaning. Such meaning of
constitutivity differs from that of declarative speech acts, rather resembling the creative
attitude of Searle’s “rules that are constitutive of speech acts”.
According to this distinction, we can distinguish two types of constitutivity: thetic-
constitutivity (from hέri1, “affirmation”) directly creates its object (a legal effect or
state of affairs), while eidetic-constitutivity (from eἶdo1, shape) rather creates the
abstract concept of its object, thus making such legal effect or state of affairs possible in
the legal system. An example of thetic-constitutive rule (a declarative speech act)
would be a marriage celebrated by a public officer (as it creates the legal bond between
two individuals) or the divorce sentence of a judge (as it removes it), while examples of
an eidetic-constitutive rule (a statement constitutive of speech acts) would be the laws
that establish such procedures and effects (marriage and divorce).
Conte [14, pp. 82–83] classifies those rules as conditions for their regulated entity:
then, an eidetic-constitutive rule is a necessary condition and a thetic-constitutive rule
is a sufficient condition. These two types of constitutive rules are different, and this
duality cannot be reduced or simplified, as argued by Roversi [29].

2.2 Relative Necessities as Constitutive Statements

According to Sartor [31], “in many cases, when a legal text uses the words must, ought,
may, or can, it does not express obligations or permissions in the sense discussed
above, but it conveys a completely different notion, which is parasitical on the idea of a
normative conditional.”
Consider those cases in which, for example, the law says that a petition or contract
must or must not be done in a certain way, or that it can or cannot contain certain terms.
In these cases, the law establishes what we may call a relative necessity: it establishes
that certain requirements have (or don’t have) to be satisfied for a certain legal result to
 Legal Patterns for Different Constitutive Rules 109

be obtained in a certain way. Often, the specification of this result is left to further
normative propositions. For instance, suppose that in a legal text, after stating that
“whoever appropriates the property of others is going to be punished as a thief”, it is
stated that “the appropriator must have the intention of getting permanent possession of
the stolen object”. Clearly, there is no legal obligation to have such intention. The
“must” signals a necessity, relative to the normative antecedent which determines
subjection to punishment for theft. It indicates that the elements explicitly contained in
the antecedent of the rule on theft are not really sufficient to produce the effect indicated
in that rule: a further element, namely, the intention to appropriate, is also required to
instantiate the precondition of the rule.
We may use the term anankastic – from the Greek word Ἀmάcjη, necessity – to
characterise the (anankastic) propositions expressing this kind of necessity. As we may
have normative propositions expressing anankastic connections, we may also have
propositions denying (excluding) such connections. However, the basic and constant
meaning of the anankastic must consist in what Sartor calls relative necessity, that
corresponds to the combination of the following propositions (1) and (2):

ð1Þ if A then B

ð2Þ C must be realized for B to be determined according to ð1Þ

Being considered equivalent to the following proposition (3):

ð3Þ if A and C then B:

According to Conte [15, p. 362], anankastic-constitutive rules create a (necessary)

condition for their regulated entity, rather than being themselves a condition.

2.3 A Taxonomy of Constitutive Statements

The considerations of the previous paragraphs allow us to identify three types of
constitutive rules, namely:
– rules that directly constitute new entities and are sufficient conditions for the new
entity to exist (thetic-constitutive rules);
– rules that merely create the possibility of new entities and are necessary conditions
for the new entity to exist (eidetic-constitutive rules);
– rules that, without constituting new entities, introduce necessary conditions for them
to exist (anankastic-constitutive rules).
Azzoni, scholar of Conte, completed this taxonomy adding noetic-constitutive rules
(directly constituting new entities with necessary and sufficient conditions, e.g. the
Grundnorm of a legal system), metathetic-constitutive rules (introducing a sufficient
condition, e.g. the rule of the House of Lords Act 1999 that grants the right to
membership to 90 hereditary peers) and nomic-constitutive rules (introducing a nec-
essary and sufficient condition, e.g. the rule in Article 12 of the Constitution of Ireland
saying that “The President is elected by direct vote of the people”) [4, p. 161]. Azzoni’s
110 M. Ceci et al.

taxonomy is represented in Table 1. Introducing noetic and nomic constitutive rules,

however, exalts some critical points in the taxonomy [30, pp. 1289 ff.], which is outside
of the scope of the present paper.

Table 1. The six types of constitutive statements as identified by Azzoni [4].

Declarative speech act Constitutive of speech act
Necessary Eidetic Anankastic
Sufficient Thetic Metathetic
Nec + Suf Noetic Nomic

This taxonomy suggests that, when dealing with constitutive rules, we need to ask
two questions:
– Is the posed condition necessary, sufficient, or both?
– Does it create its effect directly or indirectly?
Answering these questions helps in defining the effects of constitutive rules2. Specif-
ically, answering the first question helps to determine its logical formulation. In our
research, the second question helps in distinguishing rules that affect the legal source
from rules that have only affect the single interpreted rules or entities.
In a more recent work [6], conceived for rationalization of legislative drafting for
automation purposes, what we define so far as “constitutive statements” are classified
alternatively as constitutive rules or metarules (see Table 2).

Table 2. Constitutive rules and metarules found in legal texts as identified by Biagioli [5].
Classes Rules Arguments
Constitutive rules
Definition Term definiendum, definiens
Procedure addressee, counterpart, action, object
Creation Institution addressee
Organization addressee
Attribution Power addressee, counterpart, activity, object
Liability addressee, counterpart, activity, object
Status addressee, object
Metarules
Application Inclusion partition
Exclusion partition
Modification Repeal partition, position, out, in
Insertion partition, position, out, in
Substitution partition, position, out, in

2
It is however necessary to be careful in the classification of constitutive rules because it can change
depending on the perspective taken [29]: there are views where all rules are constitutive, or none of
them are.
 Legal Patterns for Different Constitutive Rules 111

2.4 Constitutive Rules for Compliance

In financial regulations, our research has so far identified five types of constitutive
statements. We will now present these, describing how each of them specifies the
generic form “(a) counts as (b) in context (c)” introduced at the beginning of this
section.
– Legal definitions: these rules, often contained in the first article of regulations,
specify the meaning (a) of specific terms (b) that are found throughout the regu-
lative text or in a subpart of it (c). When terms specifically appear in legal defini-
tions, the interpretation of their meaning cannot be arbitrary: every time they occur
in the text, they must be understood as meaning the exact combination of words (or
sentences) that appear in the definition (a). In this sense, legal definitions are
creators of intermediate legal concepts, one of the ways to represent constitutivity
[28]. Legal definitions are thus eidetic constitutive rules in the distinction made by
Carcaterra [10]: they are in fact called “stipulated definition” by Guastini [22]. They
are thus constitutive of speech act, not speech act themselves: to have the latter we
need also the rules that tell us the legal valence of that concept [30, p. 1278]. When
exhaustive (intensional definitions), legal definitions are actually nomic constitutive
rules, as they do not only define what something is (the definiens), but also what
something isn’t (the negation of the definiens). This is true only for intensional
definitions, while extensional definitions are, strictly speaking, not eidetic rules
[30, p. 1278]. For our representation we treat all legal definitions as declarative
speech acts, and therefore intensional definitions are noetic constitutive rules while
extensional definitions are thetic constitutive rules.
– Commencement rules: these rules indicate a (directly or indirectly identified) time
parameter (a) as the starting point for the validity (b) of the regulation (or part of it)
(c). Commencement rules are eidetic constitutive rules in the distinction made by
Carcaterra [10], and they are really descriptive, as they do not constitute alone the
legal effect: to constitute it, in fact, we also need the rules that describe the legal
valence of the concept of validity. For reasons of simplicity in our representation,
however, in our work we treat commencement rules as thetic constitutive rules.
– “Party-to-the-law” statements: these rules extend the subjective or objective
dimension (b) of the norm (c), by identifying new addressees (a) to it. Imposing a
sufficient condition, and its effects being mediated (by the other conditions of the
target rule) makes it a metathetic constitutive rule.
– Relative necessities: Rules restricting the subjective or objective dimension (b) of
the norm (c), by identifying a limitation (a) to it. These rules, as explained in
Sect. 2.2, pose new limit on the subjective or objective aspect of another rule. They
are similar to legal definitions except they do not create a new intermediate concept.
Because they impose a necessary condition and their effect is mediated (by the other
conditions in the target rule) they are anankastic constitutive rules.
– Amendments: these rules modify the legal source (c), either by adding, removing, or
modifying (a) the textual content (b). Some theorists don’t classify these as con-
stitutive rules but rather as meta-rules [5]. According to Carcaterra [9], repeal laws
are definitely thetic, because their effect (the disappearance of the norm from the
legal system) is immediate.
112 M. Ceci et al.

With help from the classification of Carcaterra et al. explained in Sect. 2.3, we can
group together these constitutive statements depending on their category, and model
their content and effects consequently. We see two groups here: one is composed by
commencement and amendment, who are speech acts, immediately laying their effects
either on a legal source or on single rule statements, and the second is composed by
legal definition, relative necessity and party to the law, which are rules constitutive of
speech acts (and in fact they lay their effects mainly in the vocabulary section and in
single rule statements). While legal definitions pose necessary and sufficient conditions,
and explicitly introduce an intermediate legal concept (which in turn translates to an
autonomous vocabulary entry in SBVR), relative necessities and party-to-the-law
statements introduce necessary and sufficient conditions respectively, and do not
explicitly create intermediate legal concepts. In SBVR these two statements can be
represented in the rulebook, by limiting or extending one or more factors of one or
more conditions, or in the vocabulary, creating an intermediate concept or “conve-
nience form” (see Sect. 4.2) that applies within the context.
In order to support the lawyer’s work on regulations, it is necessary to capture – and
represent in the formal model – the semantics of those five types of constitutive rules.
This activity complements the work on regulative rules in producing a complete rep-
resentation of the semantics of a regulation.

3 Legal Concept Patterns

The specific needs of the research presented in this paper suggested a rather heuristic
approach in representing constitutive rules: the focus on regulatory compliance, and
thus on regulative norms, means that the constitutive norms are ancillary norms, used to
specify and extend the semantics of the requirements. For the same reason, the abstract
model does not involve the representation of Hohfeldian powers and thus it is not
possible to represent rules attributing powers.
The use of SBVR as a basis for the intermediate language, and the creation of the
Regulatory Interpretation Methodology, creates the possibility of an ad-hoc solution
where the templates for constitutive rules are specified in the methodology docu-
mentation (the “Protocol”) in the style of a user manual, and their semantics can be
specified within the vocabulary part, through general entries that constitute a template
(i.e. thing1 counts as thing2 in context). The literature supports this approach for rep-
resenting constitutive rules since, as explained in the previous section, it highlights the
risks and limitations of a omni-comprehensive, generic approach to their
representation.
The transformation of regulatory language into SBVR is aimed at providing the
knowledge engineers with an unambiguous, understandable text while at the same time
maintaining the implicit legal knowledge that is expressed by the original legal frag-
ment. This, however, is not always possible: some legal concepts exist, that can be
expressed only by a specific combination of words. Also, sometimes a certain com-
bination of words has a specific legal meaning, corresponding to a precise legal figure.
In these cases, the risk exists that aspects of the legal figure are lost in the passage from
the legal text to the machine-readable information. In order to store the semantics of
 Legal Patterns for Different Constitutive Rules 113

these legal figures in the knowledge base we need specific patterns which, in turn, take
into account the limitations coming from the targeted formal language backing the
knowledge base.
For example, saying that “law x applies to entity/activity y” doesn’t necessarily
mean that the law performs some particular action: if the statement following the “party
to the law” pattern, then it means that for the entity or activity y new obligations apply,
i.e., the entity or activity y must comply to the obligations of the law x. In the com-
putable model, the rule should therefore be represented as “it is necessary that
entity/activity y counts as addressee in law x”, but can a computer scientist in charge
with the formal model (an STE) independently do this when he reads the original “law
x applies to entity/activity y”?
A second example is the sentence “in the present law, a handshake has the value of
an agreement” which corresponds to the constitutive statement “handshake counts as
agreement in Law X”. Being an extensional legal definition, that statement introduces a
sufficient condition (“it is necessary that handshake counts as agreement in ”)
but not also a necessary condition (“it is impossible that [something that is] not [a]
handshake counts as agreement in ”), which would be the case in presence of an
intensional definition such as “A meeting of minds with the understanding and
acceptance of reciprocal legal rights and duties as to particular actions or obligations
counts as agreement in Italian Law”. Should the computer scientist (STE) know about
legal definitions, their status of thetic-constitutive rules, and the distinction between
extensional and intensional definitions, in order to correctly translate the structured
English into the formal model? Or should instead be the legal expert (SME) the one in
charge of this identification, and deliver to the STE not only the sentence in structured
English, but also an indication in the Terminological Dictionary that the sentence
follows a specific template with specific semantics?
To represent sentences and forms with specific legal meaning, we thus introduce
legal concept patterns. Legal concept patterns are related to similar figures, known in
the literature e.g. as “technical relations” [5, 17] or “logical relations” (e.g. Hohfeldian
relations [23]).
Legal concept patterns are created in the form of a verb concept with generic verb
concept roles (e.g. the Legal Definition pattern “definiens counts as definiendum in
context”). When the SME meets a rule that follows one such pattern (e.g. “handshake
counts as agreement in ”), a verb concept entry is created and the
applicable pattern is indicated as a specific attribute. The roles played by the verb
concept roles in the pattern definition are important, as they determine the classification
of the instances found in the single rules: e.g. the Legal Definition pattern “definiens
counts as definiendum in context” tells the STE where to locate the information related
to “definiens”, “definiendum” and “context” in the rule “handshake counts as agree-
ment in ” (i.e. handshake is the definiens, agreement is the
definiendum, and is the context)3.

3
It is also possible to extend the basic patterns into more complex forms by further specifying its verb
concept roles, even introducing verb concepts as roles (e.g. adding the vocabulary entry “person1
shakes hands with person2” with the attribute “general concept: handshake” results in the more
complex pattern “person1 shakes hands with person2 counts as agreement in ” – see Fig. 1).
114 M. Ceci et al.

Generally, legal concept patterns enhance the interaction between SME and STE
during the iterative process of translation by allowing both users to refer to concepts
that are defined both in their legal valence (for the SME) and in their formal model (for
the STE). When a STE finds “handshake counts as agreement” and doesn’t know how
to model it in the machine-understandable knowledge base, the STE can ask the SME
to point at a legal concept pattern to specify the intended legal meaning. The SME
would refer e.g. to the legal definition pattern that indicates “definiens counts as
definiendum in context”, and this would at the same time specify the role of the terms
handshake and agreement in the formal model, for the STE, and remind the SME that
e.g. according to the Protocol (RIM – Regulatory Interpretation Methodology) the
context of legal definitions must be explicitly stated.
Legal concept patterns thus assist SMEs in conveying legal content in a way that is
understandable by the STE, without abandoning the legal language constructs that
express a specific meaning4.

It is obligatory that each agreement for a Real Estate Contract is recorded.

Source: Italian Civil Law

It is necessary that handshake counts as agreement in Italian Civil Law

Rulebook
Vocabulary
Person shakes hands with Person
Synonymous Form: Handshake

Agreement for Contract Real Estate Contract

Equivalent to: Contract has Agreement General Concept: Contract

Fig. 1. Example of an implicit ontology built using SBVR. In the example, any occurrence of
the verb concept “person shakes hands with person” that is referred to a Real Estate Contract is
implicitly inferred as being subject to duty of recording. Additional vocabulary entries such as
the noun concept “Real Estate Contract” and the verb concept “Agreement for Contract” enrich
the implicit ontology further.

4
By combining the SBVR attributes “general concept” and “synonymous (form)”, the Legal Concept
Patterns, and verb concept roles, the SMEs effectively build taxonomies – and even simple
ontologies – covering portions of the knowledge base (see Fig. 1). Those ontologies are built
independently, but can be linked together (e.g. through a common term or pattern). In this way, the
burden of enriching the ontology is shared between the SME and the STE, with the first building
modules of a legal ontology to express legal concepts, leaving to the latter only the task of merging
and consistency checking.
 Legal Patterns for Different Constitutive Rules 115

In regulatory interpretation, legal concept patterns can be used for three purposes:
– When defined in the protocol, to help representing the most important legal fig-
ures. A number of legal concept patterns are introduced in the RIM documentation,
that can be used to express common legal concepts. Their use however is not
compulsory: the SME can decide to ignore them for their first iterations, relying on
them only to disambiguate resulting vocabulary and rules when required by the STE
(see the last point).
– When defined by SMEs within their SBVR transformation, to represent recur-
ring (also non-legal) patterns easily (see business definitions and convenience forms
in Sect. 4.2).
– When used within SME-STE iterations, to disambiguate concepts and keep track
of the incremental process. Because legal concept patterns are documented both in
their legal model (for the SME to understand) and in their machine-understandable
formal model (for the STE to understand), they are the common ground that allows
the feedback between the two and the progressive refinement of the knowledge
base.

4 Legal Concept Patterns for Constitutive Rules

We model constitutive rules according to the doctrine explained in Sect. 2, not using
formal logics as suggested in the state-of-the art [19, 20] – as this would be too abstract
for an SME to use – but rather modelling single types of constitutive rules as pre-
defined legal concept patterns. Table 3 shows the list of legal concept patterns currently
available for constitutive rules:

4.1 General Properties of a Constitutive Rule

In this section we are going to present the common attributes of constitutive rules, and
then introduce the five legal concept patterns that specify those attributes for the five
types of constitutive rules identified in the table above, with an indication of the
meaning of the variables and on the effect that these constitutive rules have on the
regulative rules contained within the context of application of these constitutive rules.
As illustrated in Fig. 2, all constitutive rules share five main attributes:
– Modality
– Source
– Token
– Type
– Context
Every constitutive statement has exactly one constitutive modality out of three possible
values (possibility, impossibility, necessity). This attribute indicates the modality in
which the original rule entry in the rulebook has been modelled by the SME: a
necessity statement can in fact be transformed into an equivalent impossibility state-
ment (and vice versa), but it still makes sense, for authoring purposes, to specify which
116 M. Ceci et al.

Table 3. The list of legal concept patterns currently available for constitutive rules. For each of
them, we show the syntax that these patterns normally use in legal language, followed by the
syntax used in our research to capture those statement in a uniform way.

Alethic Modality
Necessity / Impossibility / Possibility

A (Token)
Constitutive Rule
B (Type)
Disjoint With

C (Context)
Regulatory Rule

Source
Deontic Modality
Obligation / Prohibition / Permission

Fig. 2. Ontology of a constitutive rule.

 Legal Patterns for Different Constitutive Rules 117

modality was originally employed. The modality also identifies necessary and/or suf-
ficient conditions: sufficient conditions are in fact represented as necessities, while
necessary conditions are represented as impossibilities of the opposite. For example,
the sufficient condition of a legal definition is “it is necessary that definiens counts as
definiendum in context”, while its necessary condition is “it is impossible that not
definiendum counts as definiens in context”.
Source is a general element in SBVR, and our application uses LegalDocML [24]
for its representation in a machine-understandable format.
Token and type change depending on the legal concept pattern being used, as the
present work does not define their semantics (and the semantics of “counts as”) at a
generic level [20]. This is when the classification proposed by the legal theory and
presented in Sect. 2 turns out useful, as it can be used to guide the design choices.
While the dichotomy thetic/ipothetic has not been given too much importance (the
distinction itself being criticized also in the theory [30, p. 1291]), the distinction of the
type of condition being posed (necessary or sufficient) played a major role in distin-
guishing and modelling those rules.
Finally, every constitutive statement has a context. In legal theory, the context of a
constitutive rule is used to identify the limits within which the constitutive effects of the
rule take place. In our approach, the concept of context is used in a slightly different
way: it represents the domains where the rule is relevant. This difference becomes
evident when dealing with commencement rules (see below): while in the legal theory
the context of a commencement rule is the entire legal system (jurisdiction), in our
approach it is used to indicate which legal fragments have their coming into force date
affected by the commencement rule. Context can be specified in terms of themes,
activities, rulebooks, or sources. For legal rules, the context must include a legal
source. The context determines which regulative rules are affected by the constitutive
rule.

4.2 The Legal Concept Pattern for Legal Definitions

Legal definitions are eidetic-constitutive rules because they create the possibility of a
speech act. In legal theory they are called constitutive definitions and reduced to a
thetic-constitutive rule and an agreed definition. Semantic Web frameworks treat them
as technical relations [17]. In our model we treat extensional definitions as sufficient
conditions and intensional definitions as necessary and sufficient conditions. In this
latter case we need two rules to represent this condition: the first stating the necessity of
it and the second stating the impossibility of the opposite (see Sect. 4.1).
Our approach captures legal definitions by using the rule that is attached to a
vocabulary entry (the definiendum). Token and type are definiens and definiendum
respectively, and context is the context of the definition, i.e. the act(s) where it applies.
For regulative rules whose source is within such context, the definiens and definiendum
are thus equivalent: the definiens can be replaced with the definiendum, and vice versa.
Figure 3 shows the replacement of the word “relevant bank”. Other examples:
• (defining a noun concept) It is necessary that trade has value more than
counts as Relevant Trade in .
118 M. Ceci et al.

It is obligatory that each relevant bank makes public a price for each share issued by that bank.
Source: MiFIR

It is necessary that bank that operates in more than 1 country counts as relevant bank in MiFIR

Rulebook
Vocabulary
BankOfIreland
General Concept: Bank BankOfIreland operates in 27 countries

Fig. 3. Example of the relationship between an operative rule and the legal definition of a noun
concept. Please note that the two vocabulary entries do not result from an interpretation of the
regulatory text but rather from internal company data.

• (defining a verb concept) It is necessary that person helps person that commit crime
counts as person participates in crime in .

Legal Definitions vs. Business Definitions and Convenience Forms

When modelling the representation of concept definitions, it is important to distinguish
legal definitions (explicitly introduced by legal texts) from business definitions (created
within the company or industry, or by the SMEs themselves).
In our approach, business definitions also cover what in SBVR is defined as
“convenience form”: in regulatory texts, it is common that parts of text are repeated
several times. For example, locutions such as “Market operators and investment firms
operating a trading venue” and “bonds, structured finance products, emission allow-
ances and derivatives traded on a trading venue” are found several times in the MiFIR5
regulatory text. To make the work less repetitive, and to increase human readability of
the interpreted rules, the SME can use convenience forms for these locutions (e.g.
“trading operator” for the first, and “traded instrument” for the second) in the rulebook,
and then define them in the terminological dictionary. As a result, during the process of
interpretation the SME divides the regulatory statement into one requirement (regu-
latory rule, a deontic statement in the rulebook) and one or more definitions (consti-
tutive rule, a structural statement in the terminological dictionary). This process is also
very valuable towards automatizing some phases of the translation: once defined, those
forms can in fact be automatically detected by NLP tools, thus easing the process of
writing SBVR rulebooks. Please note that, from a legal-philosophical point of view,
legal definitions are themselves convenience forms [28, 29, p. 112], only with an
authoritative value derived from the legal system they are part of.

5
Regulation (EU) No. 600/2014 of the European Parliament and of the Council of 15 May 2014 on
markets in financial instruments and amending Regulation (EU) No. 648/2012.
 Legal Patterns for Different Constitutive Rules 119

In our approach the distinction between legal definitions on one side, and business
definitions and convenience forms on the other, is explicitly stated in the rulebook, as
every rule is either a legal rule (legal definition) or a business rule (a business definition
or a convenience form). Further distinction between the latter two is to be found in the
context: while business rules are valid within a certain class of actions (e.g. pertaining
to a specific business activity or industry, or to a specific company), convenience forms
are valid within a certain rulebook (because they are subjective, ad-hoc solutions for the
simplification of the interpretation job at hand).

4.3 The Legal Concept Pattern for Commencement Rules

Commencement rules are thetic-constitutive rules, as they directly modify the validity
of the norms contained in the legal source indicated as context [26]. Token is a date,
while type is either a start or an end date of validity. Context is the legal source whose
efficacy times are affected. Regulative rules within the context have their applicability
restricted by the start and/or end date, as shown in Fig. 4. Examples:
– (start date) MIFID II Comes into force on January 10th, 2014.
It is necessary that January 10th, 2014 counts as efficacy start in MIFID II.
– (end date) MIFID I Stops being into force on January 10th 2014.
It is necessary that January 10th, 2014 counts as efficacy end in MIFID I.
– (start and end date) Article 13 is in force from January 1st, 2014 to December 31st,
2014.
It is necessary that January 1st, 2014 counts as efficacy start in Article 13.
It is necessary that December 31st, 2014 counts as efficacy end in Article 13.
– (end and start date) Article 13 is suspended from June 1st 2014 to June 15th 2014.
It is necessary that June 1st, 2014 counts as efficacy end in Article 13.
It is necessary that June 14th, 2014 counts as efficacy start in Article 13.

Regulatory Rule
source
Source validityStart validityEnd

o Reg. A June 1st, 2015 July 15th, 2015

context
o Reg. B efficacyStart efficacyEnd

source
Commencement Rule

Fig. 4. Ontological model of the relationship between a regulatory rule and a commencement
rule. The dashed properties of the regulatory rule are inferred.
120 M. Ceci et al.

4.4 The Legal Concept Pattern for Party-to-the-Law Statements

Party-to-the-law statements are metathetic-constitutive rules, as they create the possi-
bility of “new speech acts” by specifying a new category of addressees for an existing
speech act. Token is one (or more) person(s); type is “the addressee of a norm” or “the
co-responsible of the breach” or other similar legal liability figures; context is the norm
or legal text. Regulative rules within the context have now potentially new addressees.
Examples:
– (context as a legal source) It is necessary that Person that participates in crime
counts as subject to the law in Article 13
– (context as an interpreted norm) It is necessary that Person that participates in
crime counts as subject to the law in Obligation_L_143.

4.5 The Legal Concept Pattern for Relative Necessities

Relative necessities (see Sect. 2.2) are the converse of party-to-the-law statements, as
they identify necessary conditions for the regulative rule. On this aspect, they are
anankastic-constitutive rules. Token is the new requirement; type is the target condition
contained in the context; context is the target regulatory statement(s) or legal source.
Regulatory statements within the context have their condition extended with the new
requirements.
Because they are necessary conditions, they are better expressed as impossibility
statements, stating the impossibility of negating the condition. Example:
– It is prohibited that appropriator steals object
– It is impossible that person that not has intention to keep object counts as appro-
priator in Rule1.

4.6 The Legal Concept Pattern for Amendments

Amendments are thetic-constitutive rules, as they immediately create new states of
affairs (insert/remove/modify regulation from legal system). Token is the new text, type
is the old text, and context is the target location as a structural element (e.g. article,
clause). Regulative rules within the context are replaced/repealed. Examples:
– (substitution) It is necessary that counts as in
.
– (new text) It is necessary that, counts as new text in
.
– (repeal) It is necessary that repeal counts as
in .

4.7 Possibility Statements

SBVR possibility statements can be used for representing exceptions [13]. They can
also be used for representing eidetic (or, in some cases, metathetic) constitutive rules or
rules that attribute Hohfeldian powers (“ESMA will publish technical standards on
 Legal Patterns for Different Constitutive Rules 121

what constitutes a prevalent market condition”): because these rules introduce events
that may or may not happen, the purpose is not to trigger automatic conclusions out of
those statements, but only to record the eventuality of them to happen and the legal
relevance attributed by the law to such administrative acts. Because the scope of our
research is regulatory compliance, this information is only marginally relevant and thus
needs to be recorded but not semantically enriched. From a theoretical point of view,
such statements are constitutive to the extent to which they attribute a new power (as
noted previously, Hohfeldian powers are not represented in our approach). In all other
cases, e.g. when they foresee the publishing of some documentation, these statements
have no constitutive power as they are statements de jure condendo and not de jure
condens [11, p. 19, 30, p. 1273].

5 Conclusions

The paper presented results from applied research on compliance regarding the rep-
resentation of alethic statements in an intermediate language that is human-readable
and that can be mapped to a machine-understandable language. The solution applies
notions from philosophy of language and philosophy of law to AI & Law, identifying
different types of constitutive statements.
The paper claims that, in order to capture the different legal effects of these
statements, we need to represent them through distinct models, with different semantics
but a similar syntax, emanation of a general “constitutive rule” pattern.
Legal concept patterns are thus conceived to fill the gap between the SME and STE
in the process of translating the regulatory text into machine-readable information, a
process that is collaborative and iterative. This process, and this solution, are part of the
Regulatory Interpretation Methodology (RIM), that governs the translation process.
Applications of legal concept patterns and the RIM include: building a knowledge
base and exploring it; modelling the effects of alethic statements in the document
metadata (for metarules) or in the rulebook/vocabulary (for constitutive rules); mapping
the knowledge to external ontologies such as FIBO or FIRO (Financial Industry
Regulatory Ontology [2]) for reasoning and queries; mapping to a (defeasible) rule
language.
In FIRO, reasoning capabilities rely on axiomatization of rules through conditions
and factors. The model for regulative rules is explained in previous work [2], while the
model for constitutive rules is currently under construction.
The next step for the research is to find a logical formulation for the types of
constitutive statement, similarly to what has been done [13] for regulative rules. This
will allow the definition of the logical expressivity necessary to represent them in a rule
language. Outside of constitutive rules, but towards the same goal of logical formal-
ization, the attention will focus on the formalization of keywords (especially logical
operators and quantifiers). The research is also investigating the application of NLP
techniques to speed up the translation process, especially in the most repetitive tasks.

Acknowledgments. This work is mainly supported by Enterprise Ireland (EI) and the Irish
Development Authority (IDA) under the Government of Ireland Technology Centre Programme.
122 M. Ceci et al.

References
1. Abi-Lahoud, E., O’Brien, L., Butler, T.: On the road to regulatory ontologies. In: Casanovas,
P., Pagallo, U., Palmirani, M., Sartor, G. (eds.) AICOL -2013. LNCS (LNAI), vol. 8929,
pp. 188–201. Springer, Heidelberg (2014). https://doi.org/10.1007/978-3-662-45960-7_14
2. Al Khalil, F., Ceci, M., Yapa, K., O’Brien, L.: SBVR to OWL 2 mapping in the domain of
legal rules. In: Alferes, J.J.J., Bertossi, L., Governatori, G., Fodor, P., Roman, D. (eds.)
RuleML 2016. LNCS, vol. 9718, pp. 258–266. Springer, Cham (2016). https://doi.org/10.
1007/978-3-319-42019-6_17
3. Austin, J.L.: How to Do Things with Words, 2nd edn. Oxford University Press, Oxford
(1962)
4. Azzoni, G.: Condizioni costitutive. Rivista Internazionale di Filosofia del Diritto 63, 160–
191 (1986)
5. Biagioli, C.: Modelli Funzionali delle Leggi: Verso Testi Legislativi Autoesplicativi.
European Press Academic Publishing, Firenze (2009)
6. Biagioli, C., Sartor, G.: Regole e atti linguistici nel discorso normativo: studi per un modello
informatico-giuridico. Nuovi Modelli Formali del Diritto. Il Ragionamento Giuridico
nell’Informatica e nell’Intelligenza Artificiale, CLUESP (1993)
7. Boella, G., van der Torre, L.: Regulative and constitutive norms in normative multiagent
systems. In: KR 4, pp. 255–265 (2004)
8. Boer, Alexander, Winkels, Radboud, Vitali, Fabio: MetaLex XML and the legal knowledge
interchange format. In: Casanovas, Pompeu, Sartor, Giovanni, Casellas, Núria, Rubino,
Rossella (eds.) Computable Models of the Law. LNCS (LNAI), vol. 4884, pp. 21–41.
Springer, Heidelberg (2008). https://doi.org/10.1007/978-3-540-85569-9_2
9. Carcaterra, G.: Le Norme Costitutive. Giuffrè, Milan (1974)
10. Carcaterra, G.: La Forza Costitutiva delle Norme. Bulzoni, Rome (1979)
11. Carcaterra, G.: Le regole del Circolo Pickwick. Nuova Civiltà delle Macchine 3, 16–23
(1985)
12. Ceci, M., Al Khalil, F., O’Brien, L.: Making sense of regulations with SBVR. In: RuleML
2016 Challenge, Doctoral Consortium and Industry Track hosted by the 10th International
Web Rule Symposium (2016)
13. Ceci, M., Al Khalil, F., O’Brien, L., Butler, T.: Requirements for an intermediate language
bridging legal text and rules. In: MIREL 2016 Workshop, held within JURIX 2016, Nice,
France, 13 December 2016 (2016)
14. Conte, A.G.: Konstitutive regeln und deontik. In: Morscher, E.S. (ed.) Ethik. Akten des
Fünften Internationalen Wittgenstein-Symposiums (Kirchberg am Wechsel, 1980), pp. 82–
86. Hölder-Pichler-Tempsky, Wien (1981)
15. Conte, A.G.: Materiali per una tipologia delle regole. Materiali per una Storia della Cultura
Giuridica 15, 345–368 (1985)
16. Conte, A.G.: Regola costitutiva in Wittgenstein. In: Conte, A.G. (ed) Filosofia del
Linguaggio Normativo. I, pp. 237–54. Torino (1° ed. 1981) (1995)
17. Francesconi, E.: A description logics framework for advanced accessing and reasoning over
normative provisions. Artif. Intell. Law 22, 291–311 (2014)
18. Gordon, T.F., Governatori, G., Rotolo, A.: Rules and norms: requirements for rule
interchange languages in the legal domain. In: Paschke, A., Governatori, G., Hall, J. (eds.)
Rule Interchange and Applications, pp. 282–296. Springer, Berlin (2009)
19. Grossi, D., Meyer, J.J.C., Dignum, F: Modal logic investigations in the semantics of counts-
as. In: Proceedings of the Tenth International Conference on Artificial Intelligence and Law
(ICAIL 2005), pp. 1–9. ACM, June (2005)
 Legal Patterns for Different Constitutive Rules 123

20. Grossi, D., Meyer, J.J.C., Dignum, F.: Classificatory aspects of counts-as: an analysis in
modal logic. J. Logic Comput. 16(5), 613–643 (2006)
21. Grossi, D., Jones, A.J.I.: Constitutive norms and counts-as conditionals. In: Gabbay, D.,
Horty, J., Parent, X., van der Meyden, R., van der Torre, L. (eds.) Handbook of Deontic
Logic and Normative Systems. College Publications, Milton Keynes (2013)
22. Guastini, R.: Six concepts of “constitutive rule”. In: Eckhoff, T., Friedman, L.M., Uusitalo,
J. (eds.), Vernunft und Erfahrung im Rechtsdenken der Gegenwart (Reason and Experience
in Contemporary Legal Thought: Proceedings of the 11th World Congress of IVR, Helsinki
1983), pp. 261–269. Rechtstheorie Beiheft 10 (1986)
23. Hohfeld, W.N.: Some fundamental legal conceptions as applied in judicial reasoning. Yale
Law J. 23, 16–59 (1913)
24. OASIS: Akoma Ntoso Version 1.0 Part 1: XML Vocabulary. Committee Specification Draft
01/Public Review Draft 01, Standards Track Work Product, 14 January 2015
25. Object Management Group: Semantics of Business Vocabulary and Rules. May 2015. http://
www.omb.org/spec/SBVR/1.3/
26. Palmirani, M., Ceci, M., Radicioni, D., Mazzei, A.: FrameNet model of the suspension of
norms. In: Proceedings of the 13th International Conference on Artificial Intelligence and
Law, pp. 189–193. ACM (2011)
27. Rawls, J.: Two concepts of rules. Philos. Rev. 64, 3–32 (1955). https://doi.org/10.2307/
2182230
28. Ross, A.: Tû-tû. Harv. Law Rev. 70(5), 812–825 (1957). https://doi.org/10.2307/1337744
29. Roversi, C.: Costituire: Uno Studio di Ontologia Giuridica. Giappichelli, Turin (2012)
30. Roversi, C.: Sulla duplicità del costitutivo. RIFD, Quaderni della Rivista Internazionale di
Filosofia del Diritto 8, 1251–1295 (2012)
31. Sartor, G.: Fundamental legal concepts: a formal and teleological characterisation. Artif.
Intell. Law 21, 101–142 (2006)
32. Searle, J.R.: Speech Acts: An Essay in the Philosophy of Language. Cambridge University
Press, Cambridge (1969)
33. Searle, J.R.: A taxonomy of illocutionary acts. In: Gunderson, K. (ed.) Language, Mind and
Knowledge, pp. 344–369. University of Minnesota, Minneapolis (1975)
34. Van Haarst, R.: SBVR Made Easy. Conceptual Heaven, Amsterdam (2013)
 An Architecture for Establishing Legal
Semantic Workflows in the Context
of Integrated Law Enforcement

Markus Stumptner1(&) , Wolfgang Mayer1 , Georg Grossmann1,

Jixue Liu1, Wenhao Li1, Pompeu Casanovas2,3, Louis De Koker2 ,
Danuta Mendelson4 , David Watts2, and Bridget Bainbridge2
1
University of South Australia, Adelaide, Australia
Markus.Stumptner@unisa.edu.au
2
La Trobe Law School, La Trobe University, Melbourne, Australia
3
Autonomous University of Barcelona, Barcelona, Spain
4
Deakin Law School, Deakin University, Melbourne, Australia

Abstract. Traditionally the integration of data from multiple sources is done on

an ad-hoc basis for each analysis scenario and application. This is a solution that
is inflexible, incurs high costs, leads to “silos” that prevent sharing data across
different agencies or tasks, and is unable to cope with the modern environment,
where workflows, tasks, and priorities frequently change. Operating within the
Data to Decision Cooperative Research Centre (D2D CRC), the authors are
currently involved in the Integrated Law Enforcement Project, which has the
goal of developing a federated data platform that will enable the execution of
integrated analytics on data accessed from different external and internal sour-
ces, thereby providing effective support to an investigator or analyst working to
evaluate evidence and manage lines of inquiries in the investigation. Technical
solutions should also operate ethically, in compliance with the law and subject
to good governance principles.

Keywords: Natural language processing of legal texts

Law enforcement investigation management

1 Introduction

This paper presents ongoing research of the Australian government-funded Data to

Decisions Cooperative Research Centre (D2D CRC).1 Australia’s national security and
law enforcement agencies are faced with a deluge of intelligence and other data. Data
from sources as varied as financial transactions, immigration movements, vehicle

A previous version of this paper was presented at the Third Workshop on Legal Knowledge and
the Semantic Web (LK&SW-2016), EKAW-2016, November 19th, Bologna, Italy.

1
http://www.d2dcrc.com.au/.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 124–139, 2018.
https://doi.org/10.1007/978-3-030-00178-0_8
 An Architecture for Establishing Legal Semantic Workflows 125

registrations, call charge records, criminal histories, airline data, social media services,
etc. results in a flood of information in disparate formats and with widely varying
content. In Australia, such data is often held by individual federal, state or territory
agencies and inter-agency access to and sharing of data is generally subject to multiple
laws and complicated rules and agreements [20, 21]. Accessing relevant data as well as
linking and integrating them in a correct and consistent way remains a pressing
challenge, particularly when underlying data structures and access methods change
over time. In addition to this challenge, a large volume of data needs to be handled.
Usually only a fraction of current volumes can be analyzed. The Big Data challenge is
to extract maximum value from this flood of data through the use of smart analytics and
machine enablement.
Traditionally the integration of data from multiple sources is done on an ad-hoc
basis for each analytical scenario and application. This is a solution that is inflexible,
costly, entrenches “silos” that prevent sharing of results across different agencies or
tasks, and is unable to cope with the modern environment, where workflows, tasks, and
priorities frequently change. Working within the D2D CRC, one group of authors of
this article are currently involved in the Integrated Law Enforcement Project, which has
the goal of developing a federated data platform to enable the execution of integrated
analytics on data accessed from different external and internal sources, in order to
provide effective support to an investigator or analyst working to evaluate evidence and
manage lines of inquiry in the investigation. This will be achieved by applying foun-
dational semantic technologies based on the meta-modelling of data models and
software systems that permit alignment and translation by use of model-driven trans-
formations between the different APIs, services, process models and meta-data repre-
sentation schemes that are relied upon by the various stakeholders. It will also provide
easily adapted interfaces to third party data sources currently outside of the stake-
holders’ reach, such as financial transactions. The other group of authors are involved
in the D2D CRC’s Law and Policy Program, which aims to identify and respond to the
legal and policy issues that arise in relation to the use of Big Data solutions by
Australian law enforcement and national security agencies.
A 2015 systematic ACM review and mapping [1] of papers on online data mining
technology intended for use by law enforcement agencies identified eight main prob-
lems being addressed in the literature: (i) financial crime, (ii) cybercrime, (iii) criminal
threats or harassment, (iv) police intelligence, (v) crimes against children, (vi) criminal
or otherwise links to extremism and terrorism, (vii) identification of online individuals
in criminal contexts, and (viii) identification of individuals. The survey also included an
array of technologies capable of application to Open Source Intelligence (OSINT), i.e.
data collected from publicly available sources in the fight against organized crime and
terrorism: Artificial Intelligence, Data Fusion, Data Mining, Information Fusion, Nat-
ural Language Processing, Machine Learning, Social Network Analysis, and Text
Mining.
Data integration in this context raises serious legal compliance and governance
challenges. While the Onlife Manifesto considers the use of self-enforcing technologies
as the exception, or a last resort option, for coping with the impact of the information
revolution [2], nothing prevents the regulation of OSINT in accordance with existing
legislation and case law, international customary law, policies, technical protocols, and
126 M. Stumptner et al.

best practices [3]. Indeed, compliance with existing laws and principles is a pre-
condition for the whole process of integration, as information acquisition, sharing and
analysis must occur within the framework of the rule of law.
We have taken this complex set of issues into account in our paper on architecture
and information workflows. In order to foster trust between citizens and national
security and law enforcement agencies, a commitment to transparency and respect for
privacy must be preserved. However, addressing these issues in practice is difficult; in
order to achieve a good outcome a more nuanced approach may be required. For
example, an insistence upon ‘full transparency’ may not be desirable for citizens and
law enforcement agencies alike if it undermines operational secrecy. Rather, the goal is
to identify an outcome that maintains public accountability, understanding that to do so
requires effort. The identification of relevant legal, regulatory and policy requirements
is the starting point of this process.

2 Architectural Challenges
2.1 Data Integration and Matching
Many prototypes for data matching exist [4]. Matching systems rely either on hand-
crafted rules or use simple lexical similarity and concept tree based similarity measures.
Complex data structures and entire Service API interface specifications are not covered.
Besides extensions for complex structures, simplification of human input and incre-
mental match maintenance are open issues for further research.
Mapping of relational data sources to semantic models is still a predominantly
manual activity. Standards, such as XML-DB and RDB2RDF can represent only
syntactic mappings. Academic tools (e.g., Karma) allow mapping of relational sources
to rich semantic models based on past mappings. Incremental match maintenance (if
the model on either side evolves) and support for query APIs and meta-data attributes
are not supported in the current tools.
Linked data uses standards such as RDF and OWL for linking knowledge sources
in the Web. Although links can be established manually or with the help of various
application- and source-specific algorithms, dealing with the semantic interpretation of
links spanning multiple sources, the integration of data models, meta-data, and the
possible unintended consequences of linking entities is often left to application
programmers.
NIEM2 has emerged as a standard for information exchange between government
agencies in the U.S.A. The standard specifies data models for specific message
exchanges (in XML format) between two endpoints, and covers core data elements that
are commonly understood and defined across domains (e.g. ‘person’, ‘location’) as well
as community specific elements that align with individual domains (e.g. immigration,
emergency management, screening). However, the standard is weak in relation to meta-
data and provenance information, and security considerations are orthogonal. More-
over, the architecture is designed for enterprise application integration, not Big Data

2
National Information Standard Model, https://www.niem.gov/.
 An Architecture for Establishing Legal Semantic Workflows 127

analytic interfaces. There is no equivalent standard in Europe yet although the EU

General Data Protection Regulation (GDPR) (2016/679), which comes into force in
May 2018, is motivating its development [5].
Current research focuses on schema mapping where a relationship between data
specifications is established [4, 6] in a semi-automated way and requires a domain
ontology. Current challenges include dealing with the evolution of schema- and data
structure, for example, how to re-establish links across data sources that have been
changed, integrating data on the record level – not only on the schema level, and
integration of data streams which is particularly important for integration of events.

2.2 Meta-Data
Meta-data management is addressed in various proprietary ways in most commercial
databases, intelligence tools, and Big Data platforms. A federated meta-data mecha-
nism is required that spans multiple vendor tools and can capture and manage meta-
data such as provenance, data quality, and linkage information at the right granularity
(attribute/fact level) for a policing and intelligence context.
Linking data and data access processes to related legal policies and workflows is
required but often not provided explicitly. Although there are a growing number of
databases that use licenses (CC), most of them do not contain any reference to licenses
[7]. Yet, there are some research attempts to compose them [8] or to facilitate their use
within a copyright term bank [10], or through a general framework [9].
Meta-data approaches for linked data platforms, such as Resource Description
Framework (RDF) annotations, are not standardized and possess no widely agreed-
upon semantics. The W3C is currently working on security standards for linked data.3
However, this standard will be generic and may not meet the specific needs of intel-
ligence and policing applications (e.g. it will lack the capacity to establish and preserve
the provenance of the meta-data, which is critical when dealing with data that is
sourced across governmental or organisational boundaries). Temporal aspects and the
degree of confidence in meta-data are also not considered.
Information governance is as important in Big Data initiatives as in traditional
information management projects. Gartner has identified information governance as
one of the top three challenges of Big Data analytics initiatives.4 SAS is also singling
out information governance and data quality as major challenges to the success of
analytics projects.5
In traditional information management initiatives, the focus has been on absolute
control of the data attributes such as accuracy, consistency, completeness and other
data quality dimensions. Initiatives such as meta-data management and master data
management have assisted in creating ‘single versions of the truth’ for sharing infor-
mation assets [33].

3
https://www.w3.org/Metadata/.
4
Gartner Data & Analytics Summit 2017. 20–21 February/Hilton Sydney.
5
SAS Institute (Suite of Analytics Software), esp. Best Practices in Enterprise Data Governance,
https://www.sas.com.
128 M. Stumptner et al.

Big Data initiatives involving the placement of disparate data sets into ‘data lakes’
for analysis have significant information governance issues as they have the potential to
force knowledge of contextual awareness and semantic understanding. Once analytical
models have been created their operationalization will be restricted without data
curation and lineage metadata.

2.3 Workflow Orchestration

Commercial tools predominantly rely on proprietary implementations of analytic tool
chains and workflows. Although there is some support for exchange of analytical
processes (e.g., through an UIMA specification) in commercial tools (Leidos TeraText)
tools are confined to a federation comprising a single vendor’s analytic tool chain in
many cases. Some platforms, e.g. SAS IM, can access Hadoop file systems and gen-
erate analytic scripts for such Map-Reduce architectures. Big Data frameworks
(Hadoop, YARN, RabbitMQ) support efficient and flexible data pipelines but require
configuration and custom code for integration with vendor tools for each individual
application.
Scientific workflow tools (Kepler, Taverna) provide process design and limited
execution monitoring and (coarse-grained) provenance mechanisms. However, exten-
sions and novel meta-data models are required to suit the specific tools for
policing/analytics and fine-grained provenance.
If more than a few selected data sources are accessed in a data pipeline, data access
itself can become an issue since end-user analysts cannot be expected to know all of the
underlying technical data models and languages used by the individual systems (e.g.,
(Geo)SPARQL, (Geo)JSON, XML, etc.). Effective end-user analytic processes there-
fore must abstract from source-specific query languages at the user level. There is
ample academic work on query translation, but mostly for relational databases [11].
Existing analytic tools/products assume that all data is available locally, and require
(at least partial) ETL/ingest and indexing of data to facilitate linking, matching, search,
deduplication, and quality checking. If federated cross-platform analytic processes are
to be supported, systematic mechanisms are required for managing cached and repli-
cated data across vendor tool chains, and local updates made in an intelligence tool
should be fed back into the federation.
Multiple workflows and policies may apply to analysis tasks (in particular to
identity resolution), individual cases and agencies, and workflows may differ based on
availability of data, timeline, and security credentials. As such, in addition to executing
a given analytic process it is also desirable to suggest and configure suitable analytic
processes specific to the analyst’s situation.

3 Architectural Overview

The overall architecture for an Integrated Law Enforcement System is described in

Fig. 1. The architecture comprises 8 broad categories of services depicted as blue
rectangles.
 An Architecture for Establishing Legal Semantic Workflows 129

Fig. 1. Architecture overview

The chosen approach is consistent with the TOGAF Architecture Development

Method [28], in that requirements management within the project is ongoing and
changing stakeholder requirements, changes in partnerships, and technological inno-
vation result in updates to the specifics. With regard to managing strategies according
to the approaches listed in [27], the setting of the project can be categorised as a bottom
up approach (due to the independence of different agencies), open specifications, and a
focus on technical integration and a reference architecture approach. Centralisation
efforts that may lead to a more top-down approach are underway but at this point not
clear in terms of their impact.
The project has defined an open architecture for data/meta-data management and
analytic processes. As such it will translate the best practices from Enterprise Appli-
cation Integration to the “Big Data” analytic pipelines. The work addresses aspects
related to data and metadata modelling and storage, modelling and execution of ana-
lytic processes across multiple analytic tools and data sources. Central to this archi-
tecture is a framework for effective semi-interactive entity linking and querying of
linked data. The project has been working on a comprehensive data management
framework that relies on a well-defined shared data and meta-data model supported by
vendor-agnostic interfaces for data access and execution of processes comprising
analytic services offered by different tools. In the following sections we discuss the
main elements of the architecture.
130 M. Stumptner et al.

3.1 Knowledge Hub Services

The central Knowledge Hub Services area comprises the data stores and related data
management services. It governs the repository of data held in the node of the federated
architecture and exposes data and the schema via query services to front-end interfaces.

3.1.1 Data Stores

The Knowledge Hub’s data stores implement a polyglot architecture comprising of
multiple technologies tailored to different categories of data. It is partitioned into the
Curated Data area which includes databases that hold information that has been con-
firmed by a user, and the Generated Data area which hold data that have been derived
by automatic enrichment processes but not yet confirmed.
The Knowledge Graph store holds the collection of linked entities. It holds facts
and meta-data about entities and their links whose veracity has been confirmed. Data in
this store is predominantly structured, linked, and associated with meta-data, providing
a semantic entry point for operating on the base data. The information is represented in
terms of an Ontology describing the main entity types, relation types, their attributes,
and meta-data attributes. The ontology itself is also represented explicitly and can be
queried. Due to differing access rules and access patterns for data/meta-data and
ontology, entities/links/meta-data are stored separately from the Ontology. The linked
data store implements a directory of entities and links enriched with appropriate meta-
data and source information such that detailed information can be obtained from
authoritative sources that may be external to the system. This approach is needed as
data in the law enforcement domain is dispersed among a number of systems owned
and operated by different agencies. As such no centrally controlled database can fea-
sibly be put in place in the foreseeable future.
The Bulk Data store holds documents and binary objects (e.g., videos), and the
enrichments store holds derived information that may facilitate analysis and can be
promoted into the curated data area. As curated and generated data exhibit different
lifecycles, volume, and access patterns, these are held in separate data stores.

3.1.2 Ontology and Meta-Data

The contents of the data stores are governed by an Ontology that describes the domain-
specific taxonomy of entity types, their properties, and relationships. Akin to a schema
definition in a relational database, the ontology acts as a reference for knowledge
organization and aids in the integration of information stemming from external sources,
where it acts as a reference for linking and translating information into a form suitable
for the knowledge hub. The adoption of a linked data approach supported by an
ontology provides the flexibility needed to evolve data and schema over time and
supports semantic interpretation of the linked elements.
The ontology has been designed specifically for the law enforcement domain and
includes a broad spectrum of data types, including structured entities such as persons,
organizations, vehicles, communication events, and their relationships; and unstruc-
tured data including text documents, video and audio recordings. The ILE ontology is
too large to reproduce it in full in this paper; it comprises 19 high-level domain
concepts which are further refined into a total of *140 concepts and a taxonomy
 An Architecture for Establishing Legal Semantic Workflows 131

of *400 specialized relationship types. The full ontology has been documented in
[23]. These domain concepts are closely aligned with the draft National Police Infor-
mation Model (NPIM), complemented with relevant aspects drawn from the NIEM
standard6 and concepts related to case management. The provenance model is an
extension of PROV-O [24].
The ontology is complemented by a meta-data ontology that defines the meta-data
attributes that are associated with each entry in the knowledge hub. Meta-data infor-
mation includes information about provenance, links to entries in external systems and
document store, temporal qualification, security and access control descriptors, and
information about acquisition process and modification events. Meta-data is one of the
cornerstones of information management in the knowledge hub, as the process and
timing of information acquisition must be documented meticulously in order to satisfy
the legal requirements related to evidence collection. Meta-data elements can be
attached to each element (entity, property, link). This relatively fine-grained approach
has been selected to be able to support entity linking and merging of information
stemming from multiple external sources. The resulting data and meta-data model serve
as the foundation for information use, governance, data quality protocols, analytic
pipelines, exploration and justification of the results.

3.1.3 Query Services

The services exposed by the Knowledge Hub component relate to querying, keyword
search, and meta-data access. At the time of writing, the implementation of the
architecture supports structured and keyword search queries about related entities.
Graph matching and graph analytics are not currently implemented directly in the hub
and must be performed in external tools.
The architecture embraces location- and representation transparency principles that
enable front-end applications to access data irrespective of where the data is stored and
rely on a uniform representation defined by an ontology irrespective of how the data is
stored in a source system. If a knowledge hub node cannot satisfy a query based on the
information it holds, queries to external system can be spawned to acquire the
requested information. For this purpose, each knowledge hub maintains a registry of
sources that can be queried for information. Adapters for each source type abstract from
the specific technologies and message formats used to access the sources. The results
obtained from other systems are then expressed in the common representation and
merged into a consolidated list of results using entity linking and ranking algorithms.

3.2 Ingestion and Integration Services

Ingestion services provide functions for importing (bulk) data from external systems,
such as legacy case management systems and document repositories. The ingestion
process follows a pipeline architectural style where data is processed and enriched in
stages: data is acquired from the source system, the content is indexed, entities are
extracted, annotated with meta-data and provenance information, and linked to relevant

6
https://www.niem.gov/.
132 M. Stumptner et al.

existing information in the knowledge hub. The results are represented in terms of the
ontology and added to the knowledge hub.
The heterogeneity in representation and data format among different sources pre-
sents challenges related to information interpretation and transformation into the
ontology used within the data platform. Declarative data transformation methods are
employed to convert the different external representations into a common data model
and link the resulting structure to the ontology governing the knowledge hub. Ontology
matching techniques [11] are used in a semi-interactive process to match and convert
user supplied data, and graph mining and matching techniques have been developed to
improve the mapping of implicit relationships discovered between extracted entities.
The mapping from proprietary data representations to the representation used in the
knowledge hub is performed via declarative transformation specified in the Epsilon
Transformation Language framework [25]. This approach enables flexible configura-
tion of the transformation rules as sources are added. Moreover, the explicit repre-
sentation of the transformation facilitates analysis of impact of changes as the internal
ontology evolves.
Interoperability with existing data sources and systems can be achieved by con-
structing executable mappings from the (meta-)data model to the individual system’s
data models and APIs. Our work goes beyond existing Extract-Transform-Load
(ETL) and data access approaches in that the mapping will facilitate bi-directional
communication to allow for propagating updates to/from the federated knowledge hub,
and model-driven mapping technologies will facilitate maintenance of schema trans-
formations in case source systems undergo extensions or data format changes. We
intend to rely on proven meta-modelling techniques for early detection and semi-
automated resolution of mapping problems at the interfaces to legacy systems. This
approach will help avoid problems related to failing ETL processes and subtle issues
arising from changing data sources. Currently, these issues are predominantly left for
manual resolution by software engineers.
The pursued approach is well-established in Enterprise Application integration but
has only recently been considered in the form of “Big Service” integrated pipelines
[12], where a shared architecture comprising of an integrated shared data model and
implementation-agnostic service APIs is described. This project aims to translate this
idea to the policing and intelligence domain where Big Data requirements are preva-
lent. Given that the number of sources relevant to policing and intelligence has been
increasing, unless systematic data management and access mechanisms are imple-
mented, data quality, provenance and maintenance issues are likely to worsen if the
current siloed approach is continued.

3.3 Analytic Services

Analytic services related to the data platform include entity extraction from unstruc-
tured text [26], entity linking, similarity calculation and ranking. Entity linking and
ranking methods that are effective for sparse data are being developed. Moreover, the
inclusion of meta-data attributes in the resolution and ranking calculations is a dis-
tinguishing feature of this work. Services provided by commercial tools, such as
 An Architecture for Establishing Legal Semantic Workflows 133

network analysis and entity linking/resolution solutions, can be integrated in the

modular architecture.
The project will provide an efficient, open orchestration platform for Big Data
applications that can incorporate capabilities from multiple tools made available
through services. Integration of multiple COTS tools will furthermore enable analysts
to use familiar languages and tools for exploring information and semi-interactive
linking/resolution while relying on capabilities beyond their single tool to perform
search, linking, analysis tasks. For static information models, specific-purpose tool
chains could be coded by software engineers; however, for end-user specified analytic
pipelines and evolving linked data models (e.g. case-specific taxonomies) a compre-
hensive data and processing architecture is needed.
For efficiency of certain analytic tasks, data may need to be replicated and indexed
in multiple locations, subject to security and data access considerations. Keeping these
cached copies consistent is a non-trivial problem that we will address through lever-
aging incremental data propagation and eventual consistency mechanisms. This
capability will benefit end users who can rely on automation of complex data man-
agement processes, avoid batch ETL jobs and error-prone manual coding of data/meta-
data transformation pipelines, and improve data consistency and meta-data capture.

3.4 Automation Services

Automation services provide workflow and process orchestration functions. Workflow
services will facilitate the enactment of work processes such as acquiring authorization
and warrants. Supported by a set of user-facing “widgets” and a library of tasks and
processes, common standard processes can be planned and automatically triggered and
executed. Process standardization for common tasks may improve efficiency and
compliance with policies and legal requirements. In the domain of law enforcement,
some processes vary depending on the context of the investigation. A task ontology for
investigation planning could capture the semantics of key process steps as well as
machine-interpretable descriptions of the roles of various actors and data objects, e.g.
evidence, in the investigation process. Automatic configuration of workflows and
tracking of their execution may reduce the potential for errors. Moreover, appropriate
provenance and chain of evidence can be established by linking the information
obtained in the course of an investigation with the process activity and timeline that led
to its acquisition. At the time of writing, the automation services component is not yet
implemented in software.

3.5 Security
Information security is one of the main concerns in a system designed for law
enforcement. Trust in the sharing platform is paramount as an absence of trust and
security protocols will prevent sharing of most data. Any data sharing platform in this
domain needs to be capable of operating in a multi-agency environment where each
agency may have its own security and information sharing policies and protocols. It is
difficult to envision a single system and access control policy that would simultane-
ously satisfy all stakeholders’ requirements.
134 M. Stumptner et al.

The approach taken in the architecture presented here rests on two principles: (i) a
fine-grained security model where access privileges can be associated with each
individual fact that in the knowledge hub (akin to the Accumulo database management
systems), and (ii) a federated network of linked knowledge hubs that collaborate to
provide access to information. The granular access privilege model enables precise
control of what information can be disclosed (e.g. some attributes and relationships
associated with selected entities may be classified or restricted whereas others may be
accessible to all authorized users). The federated architecture aims to build trust in the
sharing platform by maintaining control of data access within each individual source
organization. Queries are dispatched to multiple nodes and executed under the local
node’s access policy. The results are then transmitted and collated at the originating
node where a query was posed. At the time of writing, the precise access control model
and full implementation of the access federation remain the subject of future work.

3.6 Legal Workflow Processing

Another key concern is the incorporation of constraints into the workflow execution to
ensure compliance with laws, policies and procedures, including agency legislation and
applicable privacy requirements. Natural Language parsing can be used to elicit event
specifications that could then be translated into business rules in an executable formal
language and issued to an event processor in the knowledge hub [19]. These rules
would be used to check and guarantee conformance of analytic processes/workflows
and data usage.
At present, there exists a substantive body of works on law and semantic languages
—LegalXML, LegalRuleML, RDF and legal ontologies in OWL modelling interoper-
ability and reasoning [13]. However, in the field of policing, law enforcement and
security, research and experience have shown that this relationship is by no means
simple. Law and policies are subject to contextual and dynamic interpretation across
different jurisdictions, legal systems and policy environments [22]. Legal requirements
cannot be comprehensively itemized/programmed, nor can the rule of law, privacy and
data protection be hard-coded, in particular, where judgement or an exercise of dis-
cretion is required [14, 15]. Thus, indirect strategies [16] and design tactics [17],
including privacy engineering and risk management, will need to be employed to ensure
both compliance with the Australian legal system, and beyond that, the proactive
adoption of privacy and data protection safeguards and ethical principles. [18].

3.7 Example
The system has not been tested with end-users users, but the following provides an
example illustrating its application and the legal challenges faced by the designers. In
the context of police investigations, work processes can be supported by partially
automated workflows that help investigators carry out activities efficiently while, at the
same time, ensuring that each activity is linked to appropriate supporting information.
For example, the planning underpinning an application for a search warrant of premises
by an officer of the Australian Federal Police in terms of the Crimes Act 1914
(Cth) could be partly automated.
 An Architecture for Establishing Legal Semantic Workflows 135

The common law imposed significant restrictions on the use of search and seizure
powers by government officials and constables based on the inviolability of property
interests:
Against that background, the enactment of conditions, which must be fulfilled before a search
warrant can be lawfully issued and executed, is to be seen as a reflection of the legislature’s
concern to give a measure of protection to those interests. To insist on strict compliance with the
statutory conditions governing the issue of a search warrant is simply to give effect to the
purpose of the legislation.7

The law therefore provides a range of control measures to protect the rights of
individuals affected or potentially affected by a search warrant. Apart from the pro-
cedures prescribed by the Crimes Act 1914, outlined in this example, other Acts may
also be relevant, for example the Australian Federal Police Act 1979 (Cth) and the
Privacy Act 1988 (Cth). There would also be internal agency procedures to be fol-
lowed, including the Commonwealth Director of Public Prosecution’s Search Warrant
Manual for obtaining and executing warrants under Commonwealth law. A range of
legal questions may therefore arise and these may differ from case to case.
Section 3E of the Crimes Act 2014 sets out a number of requirements that must be
met before a valid search warrant can be issued. In broad terms, a successful search
warrant application involves two steps. The first is to assemble the necessary material
necessary to enable the applicant to present sufficiently persuasive material, on oath or
affirmation, to enable an ‘issuing officer’ to be satisfied ‘that there are reasonable
grounds for suspecting that there is, or there will be within the next 72 h, any evidential
material at the premises.’8 The second is for the issuing officer, once so satisfied, to
address the requirements set out in Section 3E(5) in the warrant itself. These include a
description of the offence to which the warrant relates,9 a description of the premises to
which the warrant relates10 and the kinds of evidential material to be searched for.11
The background processes that are needed to support compliance with requirements
such as these could be automated using a federated data architecture. Investigation
planning could be supported by an ontology describing goals and activities that may be
conducted in the course of an investigation. Each goal would be associated with
supporting sub-goals and activities as well as information requirements. If an element is
added to the investigation plan, subordinate elements could be automatically added
and, where possible, executed automatically. This would require a careful preparatory
work, as each line of investigation may rise separated legal issues.
For example, if an investigator adds a line of inquiry (e.g. representing the criminal
history, if any, of the owner of the premises) to the investigation plan in the case
management system, the integrated information architecture as described in this paper
would automatically enable a search of its knowledge hub for relevant information,
including the criminal history of the subject, whether the person owns a registered

7
George v Rockett (1990) 170 CLR 104 at pp. 110–111.
8
S 3E(1) of the Crimes Act 1914.
9
S 3E(5)(a) of the Crimes Act 1914.
10
S 3E(5)(b) of the Crimes Act 1914.
11
S 3E(5)(c) of the Crimes Act 1914.
136 M. Stumptner et al.

firearm and, where relevant, whether previous applications were made for warrants
relating to the same person or premises, and their outcomes. The system would then
populate an application for an arrest warrant, which, after being sworn or affirmed by
the applicant, could be presented, potentially automatically, to the issuing officer and
the applicant would be informed of the outcome. If the outcome is positive, execution
of the warrant must be planned. To facilitate risk mitigation, the system could further
determine whether there is information indicating that other persons reside at the same
address, whether they pose a potential threat, or whether there are any potential threats
linked to surrounding premises and their occupants that should be considered. Such
information may be obtained from data sources within law enforcement, such as a case
management system containing structured person records, documents and notes; and
from sources external to the agency, such as council rate bills, electoral roll informa-
tion, and information extracted from social media posts by the subject. Linking this
information to the corresponding elements in the investigation plan facilitates a more
comprehensive and efficient consideration of available information and automation
facilitates appropriate execution of mandated investigation practices.
Many legal questions arise in relation to the design of an effective automated
system, for example: What are the different types of information that an investigator
would wish to access? Where relevant information is highly sensitive to another
investigation conducted by a different team, is the investigator entitled to access that
information? Once collected for purposes of the warrant, can personal information of
the occupants and residents in the area be used and stored for future investigations?
Where answers to these legal questions may be clear, they may differ across Australian
states and territories [29, 30], and the architecture should be flexible enough to
accommodate all legal requirements. In this project we are therefore addressing two
separate but linked sets of problems: (i) the coexistence of both artificial and human
decision-making and information processes; and (ii) the modelling of specific legal
requirements arising from different legal and government sources [31, 32]. We are
considering a blended “RegTech” perspective12 to be applied to law enforcement and
security with the due legal protections in place.

4 Intention and Future Work

This paper outlined the data management architecture for supporting law enforcement
agencies under development at the D2D CRC. Although some of it is confidential at a
granular level, the architectural overview, meta-data driven integration, and legal
workflow processing can be disclosed for academic and scientific discussion. We have
shown that extensible domain ontologies and semantic meta-data are an essential pillar
of long-term data management in a domain where the variety of data and complex
analytical processes are dominant. In the law enforcement domain, work processes,
mandated procedures-, approval- and data acquisition processes are just as important as
the collected information. Moreover, in the age of advanced data analytics, discoveries

12
http://www.investopedia.com/terms/r/regtech.asp.
 An Architecture for Establishing Legal Semantic Workflows 137

are increasingly based on automated collection and analysis of data. As a result,

questions related to the way in which these processes are conducted and under what
circumstances their results may be used are increasingly important and must be con-
sidered simultaneously with the design of the business processes and supporting
information systems. We contend that policy models for law enforcement and national
security purposes should and can be based on an appropriate understanding and
implementation of all relevant legal requirements. We intend to further explore the
specific requirements pertaining to policing and the wider law enforcement context in
Australia and devise appropriate models and implementations that can support and
govern the information sharing activities conducted within and across organizational
boundaries.

References
1. Edwards, M., Rashid, A., Rayson, P.: A systematic survey of online data mining technology
intended for law enforcement. ACM Comput. Surv. (CSUR) 48(1), 15 (2015)
2. Pagallo, U.: Good onlife governance: on law, spontaneous orders, and design. In: Floridi, L.
(ed.) The Onlife Manifesto. Being Human in a Hyperconnected Era, pp. 161–177. Springer,
Cham (2015). https://doi.org/10.1007/978-3-319-04093-6_18
3. Casanovas, P.: Cyber warfare and organised crime. A regulatory model and meta-model for
open source intelligence (OSINT). In: Taddeo, M., Glorioso, L. (eds.) Ethics and Policies for
Cyber Operations. PSS, vol. 124, pp. 139–167. Springer, Cham (2017). https://doi.org/10.
1007/978-3-319-45300-2_9
4. Alexe, B., ten Cate, B., Kolaitis, P.G., Tan, W.C.: Designing and refining schema mappings
via data examples. In: Proceedings ACM SIGMOD International Conference on Manage-
ment of Data, pp. 133–144 (2011)
5. Rodríguez-Doncel, V., Santos, C., Casanovas, P., et al.: Legal aspects of linked data – The
European framework. Comput. Law Secur. Rev. (2016). https://doi.org/10.1016/j.clsr.2016.
07.005
6. Bellahsene, Z., Bonifati, A., Rahm, E.: Schema Matching and Mapping. Data-Centric
Systems and Applications. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-
16518-4. ISBN 978-3-642-16517-7
7. Rodriguez-Doncel, V., Gómez-Pérez, A., Mihindukulasooriya, N.: Rights declaration in
linked data. In: Hartig, O., et al. (eds.) COLD. CEUR, vol. 1034 (2013). http://ceur-ws.org/
Vol-1034/RodriguezDoncelEtAl_COLD2013.pdf
8. Governatori, G., Rotolo, A., Villata, S., Gandon, F.: One license to compose them all. In:
Alani, H., et al. (eds.) ISWC 2013. LNCS, vol. 8218, pp. 151–166. Springer, Heidelberg
(2013). https://doi.org/10.1007/978-3-642-41335-3_10
9. Cardellino, C., et al.: Licentia: a tool for supporting users in data licensing on the web of
data. In: Proceedings of the 2014 International Conference on Posters & Demonstrations
Track, vol. 1272. CEUR-WS.org (2014)
10. Rodríguez-Doncel, V., Santos, C., Casanovas, P., et al.: A linked term bank of copyright-
related terms. In: Rotolo, A. (ed.) Legal Knowledge and Information Systems, pp. 91–99.
IOS Press, Amsterdam (2015)
11. Knoblock, C.A., Szekely, P.A.: Exploiting semantics for big data integration. AI Mag. 36(1),
25–38 (2015)
138 M. Stumptner et al.

12. Xiaofei, X., Sheng, Q.Z., Zhang, L.-J., Fan, Y., Dustdar, S.: From big data to big service.
IEEE Comput. 48(7), 80–83 (2015)
13. Casanovas, P., Palmirani, M., Peroni, S., van Engers, T., Vitali, F.: Special issue on the
semantic web for the legal domain, guest editors editorial: the next step. Semant. Web J. 7
(2), 1–13 (2016)
14. Boella, G., Humphreys, L., Muthuri, R., van der Torre, L., Rossi, P.: A critical analysis of
legal requirements engineering from the perspective of legal practice. In: Seventh IEEE
Workshop on Requirements Engineering and Law, pp. 14–21. IEEE RELAW (2014)
15. Koops, B.J., Leenes, R.: Privacy regulation cannot be hardcoded. A critical comment on the
‘privacy by design’ provision in data-protection law. Int. Rev. Law Comput. Technol. 28(2),
159–171 (2014)
16. Casanovas, P., Arraiza, J., Melero, F., González-Conejero, J., Molcho, G., Cuadros, M.:
Fighting organized crime through open source intelligence: regulatory strategies of the
CAPER Project. In: Proceedings of the 27th Annual Conference on Legal Knowledge and
Information Systems, JURIX-2014, pp. 189–199. IOS Press, Amsterdam (2014)
17. Colesky, M., Hoepman, J.H., Hillen, C.: A critical analysis of privacy design strategies. In:
IEEE Symposium on Security and Privacy Workshops, pp. 33–40 (2016). https://doi.org/10.
1109/spw.2016.23
18. Maurushat, A., Bennet-Moses, L., Vaile, D.: Using ‘big’ metadata for criminal intelligence:
understanding limitations and appropriate safeguards. In: Proceedings of the 15th
International Conference on Artificial Intelligence and Law, pp. 196–200. ACM (2015)
19. Selway, M., Grossmann, G., Mayer, W., Stumptner, M.: Formalising natural language
specifications using a cognitive linguistic/configuration based approach. Inf. Syst. 54, 191–
208 (2015)
20. Bennet Moses, L., Chan, J., De Koker, L., et al.: Big Data Technology and National Security
- Comparative International Perspectives on Strategy, Policy and Law Australia. Data to
Decisions CRC (2016)
21. Parliamentary Joint Committee on Law Enforcement. In: Inquiry into the Gathering and Use
of Criminal Intelligence (2013). http://www.aph.gov.au/*/media/wopapub/senate/
committee/le_ctte/completed_inquiries/2010-13/criminal_intelligence/report/report.ashx
22. Pagallo, U.: Online security and the protection of civil rights: a legal overview. Philos.
Technol. 26, 381–395 (2013)
23. Grossmann, G., et al.: Integrated Law Enforcement Platform Federated Data Model.
Technical report, Data to Decisions CRC (2017)
24. Lebo, T., et al.: Prov-o: The PROV Ontology. W3C Recommendation (2013)
25. Kolovos, D.S., Paige, R.F., Polack, F.A.C.: The epsilon transformation language. In:
Vallecillo, A., Gray, J., Pierantonio, A. (eds.) ICMT 2008. LNCS, vol. 5063, pp. 46–60.
Springer, Heidelberg (2008). https://doi.org/10.1007/978-3-540-69927-9_4
26. Del Corro, L., Gemulla, R.: Clausie: clause-based open information extraction. In:
Proceedings of the 22nd International Conference on World Wide Web. ACM (2013)
27. Mondorf, A., Wimmer, M.A.: Requirements for an architecture framework for pan-european
e-government services. In: Scholl, H.J., et al. (eds.) EGOVIS 2016. LNCS, vol. 9820,
pp. 135–150. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-44421-5_11
28. Open Group Standard TOGAF Version 9.1 Document Number: G116. ISBN
9789087536794
29. Watts, D., Bainbridge, B., de Koker, L., Casanovas, P., Smythe, S.: Project B.3. In: A
Governance Framework for the National Criminal Intelligence System (NCIS), Data to
Decisions Cooperative Research Centre, La Trobe University, 30 June 2017
 An Architecture for Establishing Legal Semantic Workflows 139

30. Bennet-Moses, L., de Koker, L.: Open secrets: balancing operational secrecy and
transparency in the collection and use of data for national security and law enforcement
agencies. Melb. Univ. Law Rev. 41(2) (2017)
31. Bainbridge, B., de Koker, L., Watts, D., Mendelson, D., Casanovas, P.: Identity Assurance,
‘Pattern of Life’ and Big Data Analytics Report. Project B.1: Identity Assurance, Law and
Policy Program. Data to Decisions Cooperative Research Centre, La Trobe University, May
2017
32. Mayer, W., Stumpfner, M., Casanovas, P., de Koker, L.: Towards a linked information
architecture for integrated law enforcement. In: Poblet, M., Plaza, E., Casanovas, P. (eds.),
Linked Democracy: Artificial Intelligence for Democratic Innovation, IJCAI-2017 Work-
shop, August, Melbourne, CEUR, pp. 15–37 (2017). http://ceur-ws.org/Vol-1897/
33. Berson, A., Dubov, L.: Master Data Management and Data Governance, 2nd edn. McGraw-
Hill Education, New York (2010)
 Contributions to Modeling Patent Claims
When Representing Patent Knowledge

Simone R. N. Reis1(&) , Andre Reis2 , Jordi Carrabina3 ,

and Pompeu Casanovas1,4
1
Institute of Law and Technology, Faculty of Law,
Universitat Autònoma de Barcelona (UAB), Bellaterra, Spain
simonerosa.nunes@e-campus.uab.cat,
pompeu.casanovas@uab.cat
2
Institute of Informatics, UFRGS, Porto Alegre, Brazil
andre.reis@inf.ufrgs.br
3
Cephis, Universitat Autònoma de Barcelona, Bellaterra, Spain
jordi.carrabina@uab.cat
4
CRC D2D, La Trobe Law School, Bundoora, Melbourne, Australia
p.casanovasromeu@latrobe.edu.au

Abstract. This paper discusses the modeling of patent claims in ontology

based representation of patent information. Our contributions relate to the
internal structure of the claims and the use of the all-element rule for patent
coverage. Starting from the general template for the structure of the claim, we
present contributions to (1) visualization of claims, (2) storing claim information
in a web semantics framework, and (3) evaluating claim coverage using
Description Logic.

Keywords: Patent
Ontology
Claim structure
All-element rule

Web semantics
Description logics

1 Introduction

Patents are a form of intellectual property aimed to protect human inventions. Basi-
cally, the inventor discloses her invention through a patent document and in exchange
he or she receives temporary exclusivity rights to practice (manufacture, sell, etc.) the
invention. The patent document has several sections, namely: abstract, description,
claims and drawings. The section that legally determines what is protected by a patent
is the claim section [1, 2]. In this sense, the exact wording of the claims is very
important and should receive a closer attention [3, 4]. The wording of the claims is
firstly written by the inventor (or his/her patent attorney or agent), but this is not
necessarily the final writing. During the application process, claim wording will
probably suffer modifications in the final text, as requested by an examiner associated
with the government office that issues the patent rights if they are granted. The mod-
ifications of the text are often necessary so that the claims represent an original
invention for which the inventor receives exclusivity rights.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 140–156, 2018.
https://doi.org/10.1007/978-3-030-00178-0_9
 Contributions to Modeling Patent Claims 141

Several works have been proposed in claim interpretation, i.e. legally determine the
invention that is protected by the claims of a patent. The PhD thesis of Fabris [5]
includes an analysis of claim interpretation in Europe and Brazil. The book edited by
Manzo [6] covers claim interpretation in several countries, containing chapters written
by local specialists. However, neither Fabris [5] nor Manzo [6] cite the all-element rule
widely used in patent interpretation. The so-called all-element rule states that a product
is covered by a patent when the product contains all the elements recited in at least one
of the claims of the patent.
Notice that modeling patent knowledge is different from modeling patent law
knowledge. Ramakrishna and Paschke [14] use first a semi-formal representation
format for representing legal (procedural) norms, enriched by an annotated semantic
representation later on. However, they do not address the modeling of patent claims,
they only address the representation of legal texts. They propose the use of the so-
called “Elementary Pragmatics” (EP) approach to modelling patent laws and associated
pragmatics, i.e. jurisprudence texts complementing the text of the law.
Wang [7] describes how to legally avoid patent infringement, based on claim
interpretation and explicitly describes the all-element rule. The all-element rule is also
explicitly described in the work of Schechter and Thomas [8]. Considering the all-
element rule, we have identified that there is a lack of works that make an effort to
(1) visually analyze claim interpretation, (2) store claim information in a web semantics
framework, and (3) evaluating claim coverage using different types of logic such as
Description Logic and Deontic Logic. Our work intends to present contributions to
these three aspects, as described in the next paragraphs.
It is worth to notice that visualization and visual analytics constitute a recent trend in
legal analysis [16]. “Visual law” trends have been already introduced to represent
content in legal semiotics [32], human rights law [33], contracts [34, 35], user-centered
privacy design [36], risks and security [37], crisis mapping [15], and co-regulatory
mobile applications [38]. A better “visualized” interface has been proposed for e-
government executable modelling [39], patent statistics representation for quality
management [40], and adaptive consent in the European data protection framework [43].
We will concentrate on the logical representation concerning patent coverage
visualization content. We choose intuitive tools to face it for practical reasons, as we
had in mind usable and shareable interfaces with end-users. Visual analysis of claims
using Venn diagrams [9] is advanced by Brainard [10], but only the relationship among
claims is discussed, pointing out that dependent or derived claims are more restrictive.
Brainard´s work lacks a more precise discussion of infringement by comparing the
elements of an object to the elements of a claim. In this work, we propose the repre-
sentation of elements of a claim as sets visually represented as Venn diagrams. This is
different of the representation proposed by Brainard, where only complete claims were
represented as sets. We argue that by representing individual elements of a claim as
sets, the all-element rule can be better visualized, and the inventors can better under-
stand what is protected and what is not protected by a specific claim.
Considering web semantics frameworks for patent information modeling, many
approaches have been proposed. The work of Ramakrishna and Paschke [14] is more
devoted to the representation of knowledge of patent law than to represent the
knowledge of patent documents. Other works [11, 17–31] that are devoted to model
142 S. R. N. Reis et al.

patent documents knowledge do not address the internal structure of the claims. Among
these works, a comprehensive approach to model patent document knowledge pre-
sented by Giereth et al. [11] stands out. The second contribution of this paper is to
extend the ontology proposed by Giereth et al. to include information about the internal
structure of the claims, which is absent from the work of Giereth et al.
Concerning the processing of patent concepts an algorithm to compute concept
difference was proposed by Karam and Paschke [12]. The third contribution of this
paper is to describe how the all-element rule can be considered when processing patent
document information represented as Description Logic. We also discuss Deontic
Logic in this context.
This paper is organized as follows. Section 2 describes the structure of a patent,
including the internal structure of claims. Claim structure is discussed in Sect. 3.
A similar presentation for the relationship among claims is made in Sect. 4. Our
contributions to represent ontological information of claims are described in Sect. 5.
The second contribution of the paper, representing patent claim information with web
semantics, is discussed in Sect. 5.1. Processing claim information using Description
Logic, is described in Sect. 5.2. Conclusions are presented in Sect. 6.

2 The Structure of a Patent

In this section we discuss the structure of a patent. In order to illustrate this discussion,
we refer to Fig. 1, which presents the major patent upper level concepts and relations
for patent structure as proposed by Giereth et al. [11]. Notice that the structure of a
patent, according to Giereth et al. is composed of some non-textual content (i.e. figures)
and some textual content consisting of title, abstract, description and claim (sic).
The modelling proposed by Giereth et al. lacks a more detailed description of patent
claims. Both from an internal structure of a claim point of view, as well as from a
standpoint from the relationship among the different claims in the patent. In fact a
patent may have several claims, which can make references to previous claims, forming
a complex layered claiming structure, as it will be discussed later. It is very rare for a
patent to have just a single claim, a typical patent has more than one claim. When a
patent has a single claim (instead of having multiple claims), this is a sign that the
patent has been poorly written.

2.1 The Structure of a Patent Claim

The internal phrasal structure of a patent claim must follow a well-defined pattern,
according to the World Intellectual Property Organization (WIPO). This structure is
discussed in the WIPO Patent Drafting Manual [13]. The general structure is shown in
Fig. 2a, while an example is shown in Fig. 2b. The general structure of a patent claim
consists of (i) a preamble, (ii) a transitional phrase (or transition) (iii) and the body of
the claim, which recites one or more elements.
The preamble defines the nature of the invention being claimed. For instance, it
might be common to have almost identical claims in a single patent where the only
distinction is the preamble. It would be possible for a patent to claim very similar
 Contributions to Modeling Patent Claims 143

Fig. 1. Major patent upper level concepts and relations for patent structure as proposed by
Giereth et al. [11] (the figure presented here is a verbatim partial reproduction of Fig. 2 from
[11])

Preamble, transition: An apparatus, comprising:

Element (#e1); a plurality of printed pages;
Element (#e2); and a binding configured to hold the printed pages together; and
Element (#e3). a cover attached to the binding.
(a) general structure (b) one example

Fig. 2. The structure of a patent claim, according to WIPO [13].

inventions in the form of a method, an algorithm or a computer program, just by having

claims with different preambles. From a patent drafting point of view, most patent
offices worldwide require the preamble to be separated from the transition by a comma,
as shown in Fig. 2.
The transitional phrase determines the relationship among the nature (defined by
the preamble) and the elements of the body. In the example of Fig. 2(b), the transition
is the verb comprising. Notice that from a punctuation point of view, the transition is
separated from the preamble by a comma and separated from the body by a colon.
The body of the claim consists of one of more elements and the relationship among
them. In the example of Fig. 2, the body of the claim consists of three distinct elements:
e1 (a plurality of printed pages); e2 (a binding configured to hold the printed pages
together); and e3 (a cover attached to the binding). The elements may be separated by a
144 S. R. N. Reis et al.

semicolon or by a line break, but there is no obligation to indicate explicitly the

different elements of a claim.

3 Claim Interpretation with the All-Element Rule

Claim interpretation is done with the so-called all-element rule. The all-element rule
states that an object is covered by a patent when the object physically presents all the
elements verbally recited in a single claim of the patent.
A visual interpretation of the all-element rule using a Venn Diagram is presented in
Fig. 3. The preamble defines the nature of what is being claimed. For instance, the
nature of the claim in Fig. 2(b) is an apparatus. Notice that the word apparatus is
chosen because it has a broad general sense. What is claimed is defined by the nature

Nature
e2
e1
claimed

e3

Fig. 3. Visual Interpretation of the all-element rule.

(preamble) of the claim, restricted by the elements recited in the claim. In the case of
the claim in Fig. 2(b), it has three elements (e1, e2 and e3). These three elements
restrict the nature of the claim, and the claimed invention is given by the intersection
among the three elements. This is illustrated in Fig. 3.

1. Is the object an apparatus?

2. Does the object comprise a plurality of printed pages?
3. Does the object comprise a binding configured to hold the printed pages together?
4. Does the object comprise and a cover attached to the binding?

Fig. 4. Questions to detect coverage of claim in Fig. 2b using the all-element rule.

The interpretation of a patent with the all-element rule can be done through a set of
questions about the nature and the presence of the elements in the object. In order to
decide if an object is covered by the claim in Fig. 2(b), the questions in Fig. 4 should
be asked.
A given object is covered by the claim in Fig. 2(b) when the answer for each of the
questions in Fig. 4 is yes. When the answer for any of these questions is no, the object
is not covered by the claim in Fig. 2(b).
 Contributions to Modeling Patent Claims 145

4 Relationship Among the Claims of a Patent

A patent typically contains more than one claim. The reasoning for this, from a patent
strategy is that patent claims can possibly be invalidated during patent litigation. This
way, having more than one claim makes the patent more robust during litigation.
Typically, a well written patent has more than one claim.
The claims of a patent can be of two different types: independent and dependent
claims. An independent claim does not make reference to any other claim in the patent.
A dependent claim makes reference to at least one of the previous claims of the patent.
For example, we illustrate an independent claim and a derived claim in Fig. 5. The
claim presented in Fig. 5(a) is an independent claim comprising two elements. The
derived (or dependent) claim presented recited in Fig. 5(b) adds a third element to the
two elements inherited through the reference to the first claim. In this sense, the claim
in Fig. 5(b) has three elements: one recited in the claim and two inherited by reference.
Notice that the claim in Fig. 5(b) could be rewritten in independent format by textually
reciting all the three elements in the body of the claim. The claim in Fig. 2(b) could be
a possible rewriting of the claim in Fig. 5(b) in independent form. In this sense, the
claims from Figs. 2(b) and 5(b) have the same legal coverage, as they have the same
nature and the same elements.

1) An apparatus, comprising: 2) An apparatus, according to

a plurality of printed pages; and claim 1, further comprising:
a binding configured to hold the printed pages a cover attached to the
together. binding.
(a) an independent claim (b) a derived claim

Fig. 5. An independent claim and a derived claim.

The coverage of the claims in Fig. 5 can be discussed by using the all-element rule.
This coverage is illustrated in Fig. 6(a) for the claim in Fig. 5(a). As shown in the
figure, the claim in Fig. 5(b) adds a third element to the two already existing in Fig. 5
(a). The added element acts as a restriction or limitation, further limiting the coverage
of the claim from Fig. 5(b), with respect to the claim from Fig. 5(a).
As a consequence of the all-element rule, as illustrated in Fig. 6, a derived claim
tends to be narrower than the claim it refers to. This is always true when the reference is
made to add new elements to an existing claim.
The interpretation of a patent with the all-element rule can be done through a set of
questions about the nature and the presence of the elements in the object. In order to
decide if an object is covered by the claim in Fig. 5(a), the questions in Fig. 7 should
be asked.
A given object is covered by the claim in Fig. 5(a) when the answer for all the three
questions in Fig. 7 is yes. When the answer for any of these questions is no, the object
is not covered by the claim in Fig. 5(a). Concerning the claim in Fig. 5(b), it is
equivalent to the claim in Fig. 2(b). This way, the set of questions to detect coverage of
the claim in Fig. 5(b) is the same set of questions as shown in Fig. 4 for the equivalent
claim in Fig. 2(b).
146 S. R. N. Reis et al.

Apparatus Apparatus
e2 e2
e1 e1
claim of
claim of
Fig. 5(a)
Fig. 5(b)

e3
(a) Coverage of the claim in Fig. 5(a) (b) Coverage of the claim in Fig. 5(b)

Fig. 6. Coverage of the claims from Fig. 5. Element e1 is “a plurality of printed pages”;
element e2 is “a binding configured to hold the printed pages together”; and element e3 is “a
cover attached to the binding”.

1. Is the object an apparatus?

2. Does the object comprise a plurality of printed pages?
3. Does the object comprise a binding configured to hold the printed pages together?

Fig. 7. Questions to detect coverage of claim in Fig. 5a using the all-element rule.

5 Patent Ontologies and Claims

Several approaches have been proposed for ontological modeling in the patent domain
[11, 12, 17–31]. However, none of these approaches discusses the structure of the
claims in detail. Additionally, the relationship among a set of claims is not discussed in
these previous publications either.
In the following we discuss some extensions of our view to the two most relevant
approaches we found for ontological patent information representation. Giereth et al.
[11] carried out a broad ontological approach for patent knowledge based on web
semantics (OWL). Karam and Paschke [12] set a description logic approach to model
patents claims and performing differences. Both can benefit from our analysis of patent
claim structure and the all-element rule.

5.1 Representing Patent Claim Information with WEB Semantics

Giereth et al. [11] is based on web semantics (OWL) including (i) patent metadata,
(ii) patent classification (according classifications proposed by patent offices),
(iii) patent internal structure, and (iv) patent drawings.
As previously discussed, Fig. 1 reproduces the model for patent internal structure
proposed by Giereth et al. [11]. Notice that their proposed model for internal structure
only makes reference to a single Claim as part of its textual content (TextualIPContent,
in Fig. 1). We propose to have a more detailed model that includes a list of independent
claims. Each dependent claim would be linked to the claim it refers to (i.e., the parent
claim), possibly through a dependence tree. For each claim, the structure composed of
preamble (nature), transitional phrase (or transition), and the body of the claim would
 Contributions to Modeling Patent Claims 147

Table 1. Proposed xml markers for patent claim content storage.

xml marker Role in claim description
<claims> </claims> Claim section
<claim> </claim> Individual claim
<parent> </parent> Parent claim
<preamble> </preamble> Preamble
<transition> </transition> Transition
<elements> </elements> Element section
<element> </element> Individual element

be stored. The claim would contain a list of all of its elements. The suggested xml
markers are shown in Table 1.

5.2 Processing Claim Information with Description Logic

Karam and Paschke [12] proposed a description logic approach to model patents claims
and performing differences. According the authors, “their work was motivated by an
application in the context of patent applications valuation. In this context one needs to
compare the claims of the patent with previous patents solving a similar problem”.
This affirmation is true, but it can be more precisely described when the all-element rule
is taken into account. The result section of Karam and Paschke [12] is composed of two
examples, which we will discuss in the following sub-sections considering patent claim
structure and the all-element rule.

5.2.1 Discussion About Example 1 from Karam and Paschke [12]

The example 1 from Karam and Paschke [12] discusses a comparison of a patent
application against a previous patent. The patent application “is for a chair with only
one leg having a seat made only of a light material”, which indicates that this
description is being claimed, i.e. it can be found in a claim of the application, even if
the provided text in the example is not in the form of a claim. This claim is represented
in Description Logic by the formula in Eq. (1). The previous patent “describes a chair
with three legs and having one seat made of light wood”, which is represented in
Description Logic by the formula in Eq. (2).

¼ 1hasLeg u 9hasSeat:ð8hasMaterial:LightÞ ð1Þ

¼ 3hasLeg u 9hasSeat:ð8hasMaterial:ðWood u LightÞÞ ð2Þ

Karam and Paschke [12] proposed an algorithm to compute the difference between
concepts expressed in Description Logic. Their algorithm returns as answer the formula
in Eq. (3), which reads “at most one leg” for the difference between Eqs. (1) and (2).
We agree that the differences between the two concepts reside mainly in the number of
legs presented by the chairs in the application and in the previous patent. However, we
do not agree that the difference can be expressed as “at most one leg”, in a way that is
meaningful for patent interpretation. In fact, the expected difference would be two legs,
148 S. R. N. Reis et al.

which does not correspond to Eq. (3). Unfortunately, Karam and Paschke [12] do not
verbalize the meaning of Eq. (3), which makes difficult to check for possible typos in
the equation.

 1hasLeg ð3Þ

The comparison between the individual elements of the patent application and the
previous patent in example 1 are shown in Table 2. From an all-element rule stand-
point, the question is whether the claimed elements in the application were already
present in the previous patent. From Table 2, it is clear that element e2 is present in
both the application and the previous patent. It is also easy to see that the previous
patent has the element e3 of the application, even if it uses wood as the light material.
The difference resides in element e1, concerning the number of legs, but the difference
is not “at most one leg”. From a patent law point of view, the following concerns can
be raised. First, the examiner would probably not allow the word only in a claim as only
one leg means not having more than one leg, and the form not having is not a valid
form of claiming. The patent agent should use the wording exactly one leg, which is
admissible. Second, any patent agent would try to use the broader formulation at least
one leg while claiming, but then the previous patent would be valid prior art, as a chair
with three legs has at least one leg.

Table 2. The comparison of the elements in the patent application and in the previous patent in
example 1 from Karam and Paschke [12].
Element Element in the application Element in the previous patent
e1 Only one leg Three legs
e2 One seat One seat
e3 Seat of light material Seat of light wood

5.2.2 Discussion About Example2 from Karam and Paschke [12]

The example 2 from Karam and Paschke [12] discusses a comparison between a patent
application for a watch and an existing physical watch. The application is for “a watch
with at least two displays that are bright”, which is expressed in Description logic by
the formula in Eq. (4). The existing physical watch has one display that is analogical
and has one display that is not analogical, which is written in Description Logic
through the formula in Eq. (5). According to Karam and Paschke [12], their algorithm
to compute differences between formulas returns the formula in Eq. (6), as answer for
the difference between Eqs. (4) and (5). The formula in Eq. (6) reads “all displays are
bright”. We agree with the difference provided as answer in this example by Karam and
Paschke [12].

[ ¼ 2hasDisplay u 8hasDisplay:Bright ð4Þ

9hasDisplay:Analogical u 9hasDisplay::Analogical ð5Þ

 Contributions to Modeling Patent Claims 149

8hasDisplay:Bright ð6Þ

The comparison between the individual elements of the patent application and the
existing physical watch in example 2 are shown in Table 3. From an all-element rule
standpoint, the question is once again if the claimed elements in the application were
already present in the existing watch. From Table 3, it is clear that element e1 is present
in the existing watch. However, element e2 is not present in the existing watch, so it is
not a valid prior art for the patent, meaning that the patent could potentially be granted
due to the novelty provided by element e2.

Table 3. The comparison of the elements in the patent application and in the existing watch in
example 2 from Karam and Paschke [12].
Element Element in the application Element in the existing watch
e1 Has two or more displays Has two displays
e2 All displays are bright Not present

Notice that Karam and Paschke [12] introduce an algorithm to compute the dif-
ference between two concepts expressed through Description Logic to be used in patent
valuation. They also propose to use the difference algorithm to compute the difference
between two concepts. However, they do not discuss whether the difference algorithm
can be applied altogether with the all-element rule. In fact, Karam and Paschke [12]
conclude their work stating that “a direction for future work would be to investigate the
decision making process based on the results returned by the difference and empirical
rules derived from experts decisions”. We believe that the discussion provided herein
helps to work into this direction by clarifying the role of the all-element rule in experts’
decision-making.

6 Representing the All-Element Rule with Descriptive Logic

and Deontic Logic

In this section, we express the all-element rule in terms of Descriptive Logic [12] and in
terms of Deontic Logic [41]. This is done in the next two subsections. The importance
of representing claims in Descriptive Logic and in Deontic logic is justified by the
existence of frameworks for automatically computing logic difference between con-
cepts, such as the one presented in [12].

6.1 All-Element Rule in Description Logic

The all-element rule takes the general form of Eq. (7) when written in Description
Logic. There will be one of these formulas for each claim in the patent, taking into
account all the elements in the claim. Equation (7) is read in such a way that an object
covered by the claim has to have all elements recited in the claim; meaning that an
150 S. R. N. Reis et al.

object that “has element e1 and has element e2 and has element en” is covered by the
claim. In order for a physical object not to be covered by the claim represented by
Eq. (7) it must follow the conditions given by Eq. (8). Equation (8) is read in such a
way that an object not covered by the claim should not have all of the elements recited
in the claim; meaning that an object that “either has not element e1 or has not element
e2 or has not element en” is not covered by the claim. The same applies for a previous
patent not describing prior art.

¼ hasElement:e1 u hasElement:e2 u . . . u hasElement:en ð7Þ

¼ :hasElement:e1 t :hasElement:e2 t . . . t :hasElement:en ð8Þ

6.2 All-Element Rule in Standard Deontic Logic

Considering Deontic Logic [41], we have a similar modeling as well. There are early
developments by Nitta [44] and Roberts (as quoted in [45]). From our perspective, the
all-element rule takes the general form of Eq. (9) when written in Standard Deontic
Logic. There will be one of these formulas for each claim in the patent, taking into
account all the elements in the claim. Equation (9) is read in such a way that an object
covered by the claim has an obligation O to have all elements recited in the claim;
meaning that an object that “element e1 is obligatory and element e2 is obligatory and
element en is obligatory” for an object covered by the claim. In order for a physical
object to be permitted, i.e. not covered, by the claim represented by Eq. (9) it must
follow the conditions given by Eq. (10). Equation (10) is read in such a way that an
object is permitted P (i.e., the object is not covered by the claim) when the object does
not have all the elements recited in the claim; meaning that the object is permitted if
“either it has not element e1 or it has not element e2 or it has not element en”. The
same applies for a previous patent not describing prior art.

Oe1 ^ Oe2 ^ . . . ^ Oen ð9Þ

Pð:e1 _ :e2 _ . . . _ :enÞ ð10Þ

7 A Practical Example

As a practical example we examine the claims in the United States Patent 5960411,
also known as the one-click patent by Amazon [42]. The first five claims of the patent
are reproduced in Table 4. The nature of all claims is a method. The first claim is an
independent claim; the four other claims are derived directly from claim 1 by adding an
extra element. The way the claims are derived results in the Venn diagram presented in
Fig. 8.
Notice that the first claim can be considered as composed of eight distinct elements.
Figure 9 highlights these eight elements by using the proposed xml markers listed in
Table 1. Figure 9 also provides the markup for all the tags presented in Table 1,
 Contributions to Modeling Patent Claims 151

Table 4. The first five claims in United States Patent 5960411 [42].
1. A method of placing an order for an item comprising:
under control of a client system,
displaying information identifying the item; and
in response to only a single action being performed, sending a request to order the item
along with an identifier of a purchaser of the item to a server system;
under control of a single-action ordering component of the server system,
receiving the request;
retrieving additional information previously stored for the purchaser identified by the
identifier in the received request; and
generating an order to purchase the requested item for the purchaser identified by the
identifier in the received request using the retrieved additional information; and
fulfilling the generated order to complete purchase of the item
whereby the item is ordered without using a shopping cart ordering model
2. The method of claim 1 wherein the displaying of information includes displaying
information indicating the single action
3. The method of claim 1 wherein the single action is clicking a button
4. The method of claim 1 wherein the single action is speaking of a sound
5. The method of claim 1 wherein a user of the client system does not need to explicitly
identify themselves when placing an order

Method
c1
c3
c2 c5
c4

Fig. 8. Relationship among the first five claims in United States Patent 5960411 [42]
represented as a Venn diagram.

highlighting the hierarchy of the information. The fact that all the elements are
explicitly indicated allows for a more easy application of the all-element rule, including
visualization of claim hierarchy, as presented in Fig. 8.

8 General Discussion

This paper presented several different aspects of modeling knowledge about patent
documents. The discussion was focused on modeling patent claims, especially in the
light of the all-element rule. In this regard, we have demonstrated that organizing the
claims by having a clear view of the elements, parent claims and the process of
152 S. R. N. Reis et al.

<claims>
<claim #1>
<preamble>A method of placing an order for an item </preamble>
<transition> comprising: </transition>
<elements>
<element>under control of a client system,</element>
<element>displaying information identifying the item; and </element>
<element>in response to only a single action being performed, sending a request to
order the item along with an identifier of a purchaser of the item to a
server system; </element>
<element>under control of a single-action ordering component
of the server system, receiving the request; </element>
<element>retrieving additional information previously stored for the purchaser
identified by the identifier in the received request; and</element>
<element>generating an order to purchase the requested item for the purchaser identified
by the identifier in the received request using the retrieved additional
information; and</element>
<element>fulfilling the generated order to complete purchase of the item</element>
<element>whereby the item is ordered without using a shopping cart ordering
model. </element>
</elements>
</claim>
<claim #2> <parent> #1 </parent>
<preamble>The method of claim 1 </preamble>
<transition>wherein </transition>
<elements>
<element> the displaying of information includes displaying information indicating
the single action. </element>
</elements>
</claim>
<claim #3> <parent> #1 </parent>
<preamble>The method of claim 1</preamble>
<transition>wherein </transition>
<elements> <element>the single action is clicking a button. </element></elements>
</claim>
<claim #4> <parent> #1 </parent>
<preamble>The method of claim 1</preamble>
<transition>wherein </transition>
<elements> <element> the single action is speaking of a sound. </element> </elements>
</claim>
<claim #5> <parent> #1 </parent>
<preamble>The method of claim 1 </preamble>
<transition>wherein</transition>
<elements>
<element>a user of the client system does not need to explicitly identify themselves
when placing an order. </element> </elements>
</claim>
</claims>

Fig. 9. Proposed xml markers for patent claim content storage.

derivation from parent claims allow for a better visualization of claim hierarchy
through Venn diagrams. The clear view of the elements also allows creating claim
descriptions for frameworks that process claims expressed as Description Logic or
Deontic Logic. An example of such frameworks is the one proposed in [12]. We
acknowledge that our observations rely on the validity of the all-element rule. In the
next section we discuss the prevalence of the all-element rule for several countries.
 Contributions to Modeling Patent Claims 153

8.1 Prevalence of the All-Element Rule

One could argue that the observations made herein rely on the all-element rule, and
therefore are only valid in countries where such rule applies. We agree, and still our
work is useful in several countries where it applies, notably the United States of
America. We also highlight that WIPO (World Intellectual Property Organization)
makes extensive use of this rule, including in Wipo Patent Drafting Manual [13]. We
also know by the experience of some co-authors of this paper, that WIPO uses ques-
tions based on the all-element to test student knowledge on patent infringement and
prior art. Also, the Spanish patent law was recently modified and the new law includes
a mention to claim elements that can be understood as an adoption of the all-element
rule. For all these reasons, we believe that the all-element rule will be increasingly
followed by European countries in the future. This way, we believe the all-element rule
will be used in a larger number of national patent offices in the future, making the work
presented here more relevant.

9 Conclusion

This paper aims at extending ontologies and ontological frameworks to represent patent
documentation. We highlight the need to store and process information about patent
claims. Claims are the section of a patent with the function of defining what is covered
by the patent. Our first contribution shows how its representation can benefit from
using the all-element rule, visualizing them through Venn diagrams. The second
contribution extends the ontology proposed by Giereth et al. to include information
about the internal structure of the claims. Our third contribution describes how the all-
element rule can be considered when processing patent document information repre-
sented as Description Logic and Standard Deontic Logic. Considering that the claim
section represents the core of patent document knowledge, we believe these three
contributions will be essential for an effective knowledge modeling of patent docu-
ments. In the next future, we plan to extend these arguments further, and explore other
possibilities for modeling patents (e.g. in non-Standard Logic). Another open area for
future work is the possibility of using natural language processing methods for the
automatic identification and extraction of the proposed claim models from patents
described in natural language.

Acknowledgements. This research was partially funded by Meta-Rule of Law (DER2016-

78108-P, Spain). André Reis was supported by Brazilian funding agencies CAPES (Grant BEX
0466/15-8) and CNPq (Grant 312086/2016-4).
154 S. R. N. Reis et al.

References
1. Corcoran, P.: It is all in the claims! [IP Corner]. IEEE Consum. Electron. Mag. 4(3), 83–89
(2015)
2. Rackman, M.I.: Inventors: protect thyself: careful attention to the claims section will go far
toward establishing patent validity and extending the scope of protection. IEEE Spect. 15(2),
54–60 (1978)
3. Emma, P.: Writing the claims for a patent. IEEE Micro 25(6), 79–81 (2005)
4. Osenga, K.: Linguistics and patent claim construction. Rutgers Law J. 38, 61–108 (2006)
5. Guerra Fabris, R.: La determination de l´objet du Brevet en Droit Bresilien et Europeen.
Université de Strassbourg. Ph.D. thesis, 22 June 2012
6. Manzo, E.: Patent Claim Interpretation - Global Edition, 2014–2015 edn. LegalWorks, 24
October 2014
7. Wang, S.-J.: Designing around patents: a guideline. Nat. Biotechnol. 26(5), 519–522 (2008)
8. Schechter, R., Thomas, J.: Principles of Patent Law (Concise Hornbook Series). West
Academic Publishing, St. Paul (2007)
9. Parks, H., Musser, G., Burton, R., Siebler, W.: Mathematics in Life, Society, and the World.
Prentice Hall, Upper Saddle River (2000)
10. Brainard, T.D.: Patent claim construction: a graphic look. J. Pat. Trademark Off. Soc. 82,
670 (2000)
11. Giereth, M., et al.: A modular framework for ontology-based representation of patent
information. In: Proceedings of the 2007 Conference on Legal Knowledge and Information
Systems: JURIX 2007, The Twentieth Annual Conference, pp. 49–58. IOS Press (2007)
12. Karam, N., Paschke, A.: Patent valuation using difference in ALEN. In: 25th International
Workshop on Description Logics, p. 454 (2012)
13. Wipo Patent Drafting Manual. Available from WIPO at the following page address. http://
www.wipo.int/edocs/pubdocs/en/patents/867/wipo_pub_867.pdf
14. Ramakrishna, S., Paschke, A.: A process for knowledge transformation and knowledge
representation of patent law. In: Bikakis, A., Fodor, P., Roman, D. (eds.) RuleML 2014.
LNCS, vol. 8620, pp. 311–328. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-
09870-8_23
15. Poblet, M.: Visualizing the law: crisis mapping as an open tool for legal practice. J. Open
Access Law 1, 1 (2013)
16. Casanovas, P., Palmirani, M., Peroni, S., van Engers, T., Vitali, F.: Special issue on the
semantic web for the legal domain guest editors’ editorial: the next step. Semant. Web J.
7(2), 1–13 (2016)
17. Soo, V.-W., Lin, S.-Y., Yang, S.-Y., Lin, S.-N., Cheng, S.-L.: A cooperative multi-agent
platform for invention based on patent document analysis and ontology. Expert Syst. Appl.
31(4), 766–775 (2006)
18. Bermudez-Edo, M., Noguera, M., Hurtado-Torres, N., Hurtado, M.V., Garrido, J.L.:
Analyzing a firm’s international portfolio of technological knowledge: a declarative
ontology-based OWL approach for patent documents. Adv. Eng. Inform. 27(3), 358–365
(2013)
19. Trappey, C.V., Trappey, A.J., Peng, H.Y., Lin, L.D., Wang, T.M.: A knowledge centric
methodology for dental implant technology assessment using ontology based patent analysis
and clinical meta-analysis. Adv. Eng. Inform. 8(2), 153–165 (2014)
20. Lim, S.-S., Jung, S.-W., Kwon, H.-C.: Improving patent retrieval system using ontology. In:
30th Annual Conference of IEEE, 2004, IECON 2004, vol. 3, pp. 2646–2649. Industrial
Electronics Society (2004)
 Contributions to Modeling Patent Claims 155

21. Ghoula, N., Khelif, K., Dieng-Kuntz, R.: Supporting patent mining by using ontology-based
semantic annotations. In: Proceedings of the IEEE/WIC/ACM International Conference on
Web Intelligence (WI 2007). IEEE Computer Society, Washington, DC, USA, pp. 435–438
(2007)
22. Zhi, L., Wang, H.: A construction method of ontology in patent domain based on UML and
OWL. In: 2009 International Conference on Information Management, Innovation
Management and Industrial Engineering, Xi’an, pp. 224–227 (2009)
23. Law, K.H., Taduri, S., Law, G.T., Kesan, J.P.: An ontology-based approach for retrieving
information from disparate sectors in government: the patent system as an exemplar. In:
2015 48th Hawaii International Conference on, System Sciences (HICSS), Kauai, HI,
pp. 2096–2105 (2015)
24. Trappey, Amy J.C., Trappey, C.V., Chiang, T.-A., Huang, Y.-H.: Ontology-based neural
network for patent knowledge management in design collaboration. Int. J. Prod. Res. 51(7),
1992–2005 (2013)
25. Hsu, A.P.T., Trappey, C.V., Trappey, A.J.C.: Using ontology-based patent informatics to
describe the intellectual property portfolio of an e-commerce order fulfillment process. In:
ISPE CE 2015, pp. 62–70 (2015)
26. Nédellec, C., Golik, W., Aubin, S., Bossy, R.: Building large lexicalized ontologies from
text: a use case in automatic indexing of biotechnology patents. In: Cimiano, P., Pinto, H.S.
(eds.) EKAW 2010. LNCS (LNAI), vol. 6317. Springer, Heidelberg (2010). https://doi.org/
10.1007/978-3-642-16438-5
27. Li, M., Zheng, H.T., Jiang, Y., Xia, S.T.: PatentRank: an ontology-based approach to patent
search. In: Lu, B.-L., Zhang, L., Kwok, J. (eds.) ICONIP 2011. LNCS, vol. 7062. Springer,
Heidelberg (2011). https://doi.org/10.1007/978-3-642-24955-6
28. Wang, F., Lin, L.F., Yang, Z.: An ontology-based automatic semantic annotation approach
for patent document retrieval in product innovation design. In: Rui, H. (ed.) Applied
Mechanics and Materials, vol. 446, pp. 1581–1590. Trans Tech Publications, Zurich (2014)
29. Calvert, J., Joly, P.B.: How did the gene become a chemical compound? The ontology of the
gene and the patenting of DNA. Soc. Sci. Inf. 50(2), 157–177 (2011)
30. Li, Z., Tate, D.: Interpreting design structure in patents using an ontology library. In: ASME
2013 International Design Engineering Technical Conferences and Computers and
Information in Engineering Conference, IDETC/CIE (2013)
31. Zhai, D., Liu, C.: Research on patent warning index-system ontology modeling and its
application. In: Proceedings of the 4th International Conference on Innovation and
Management, vols. I and II, pp. 2051–2055 (2007)
32. Cyras, V., Lachmayer, F., Schweighofer, E.: Visualization as a tertium comparationis within
multilingual communities. Baltic J. Mod. Comput. 4(3), 524 (2016)
33. Hagan, M.: The human rights repertoire: its strategic logic, expectations and tactics. Int.
J. Hum. Rights 14(4), 559–583 (2010). https://doi.org/10.1080/13642980802704312
34. Haapio, H.: Contract clarity and usability through visualization. In: Marchese, F.T., Banissi,
E. (eds.) Knowledge Visualization Currents, pp. 63–84. Springer, London/Heidelberg
(2013). https://doi.org/10.1007/978-1-4471-4303-1_4
35. Haapio, H., Hagan, M.D.: Design patterns for contracts. In: Networks. Proceedings of the
19th International Legal Informatics Symposium IRIS, pp. 381–388 (2016)
36. Hagan, M.: User-centered privacy communication design. In: Twelfth Symposium on
Usable Privacy and Security (SOUPS 2016), Denver, Colorado, June 22th–24th (2016).
https://www.usenix.org/conference/soups2016/workshop-program/wfpn/presentation/hagan
37. Hall, P., Heath, C., Coles-kemp, L.: Critical visualization: a case for rethinking how we
visualize risk and security. J. Cybersecur. 1(1), 93–108 (2015)
156 S. R. N. Reis et al.

38. Poblet, M., Teodoro, E., Gonzalez-Conejero, J., Varela, R., Casanovas, P.: A co-regulatory
approach to stay safe online: reporting inappropriate content with the MediaKids mobile
app. J. Fam. Stud. (2016). https://doi.org/10.1080/13229400.2015.1106337
39. Olbrich, S., Simon, C.: Process modelling towards e-government – visualisation and
semantic modelling of legal regulations as executable process sets. Electron. J. E-gov. 6(1),
43–54 (2008). www.ejeg.com
40. Laub, C.: On legal validity—using the work of patent courts for quality management: the
statistical reutilization of patent court appeal decisions. J. World Intellect. Prop. 16.3–4
(2013), 168–188 (2013)
41. Hilpinen, R., McNamara, P.: Deontic logic: a historical survey and introduction. In: Gabbay,
D., Horty, J., Parent, X., van der Meyden, R., van der Torre, L. (eds.) Handbook of Deontic
Logic and Normative Systems, p. 80. College Publications, London (2013)
42. Hartman, P., Bezos, J.P., Kaphan, S., Spiegel, J.: Method and system for placing a purchase
order via a communications network. United States Patent 5960411
43. Rossi, A., Palmirani, M.: A visualization approach for adaptive consent in the European data
protection framework. In: Parycek, P., Edelmann, N. (eds.) Proceedings of the 7th
International Conference for E-Democracy and Open Government, CeDEM 2017, pp. 159–
170. Donau-UniversitŠt Krems, Krems (2017)
44. Nitta, K., Nagao, J., Mizutori, T.: A knowledge representation and inference system for
procedural law. New Gener. Comput. 5(4), 319–359 (1988)
45. Jones, A.J., Sergot, M.: Deontic logic in the representation of law: towards a methodology.
Artif. Intell. Law 1(1), 45–64 (1992)
 Modeling, Execution and Analysis
of Formalized Legal Norms in Model
Based Decision Structures

Bernhard Waltl(B) , Thomas Reschenhofer, and Florian Matthes

Software Engineering for Business Information Systems, Department of Informatics,

Technische Universität München, Boltzmannstr. 3,
85748 Garching bei München, Germany
b.waltl@tum.de, {reschenh,matthes}@in.tum.de

Abstract. This paper describes a decision support system to represent

the semantics of legal norms. The focus is on designing and software-
technical implementing of a comprehensive system to support model
based reasoning on legal norms and enabling end-users to create, main-
tain, and analyze semantic models, i.e. ontologies, representing structure
and semantics of norms.
A model based expression language (MxL) has been developed to
coherently support the formalization of logical and arithmetical opera-
tions. MxL is intended to define complex, nested, strongly-typed, and
functional operations. The paper summarizes research on the design and
implementation of a legal expert system built upon model based deci-
sion structures. Thereby, three different components, namely a model
store, a model execution component, and an interaction component have
been developed. The formalization, execution, and analysis is shown on
German child benefit regulations.

Keywords: Legal expert system · Deductive reasoning

Rule-based system · End-user centered · Domain specific language

1 Introduction

The formalization of normative texts, regarding to various aspects, e.g., propo-

sitional, deontological, temporal, defeasible, etc., is well studied [1–3] and led
to valuable principles allowing complex formal reasoning on given laws or con-
tracts. The execution of algorithmically processable formalizations requires the
translation of the textual representation, which is a complex and manual task [4].
Advances in natural language processing have at most led to tools or algorithms
supporting end-users during analysis, interpretation, and application of legal
rules [5]. The user perspective during the formalization process of legal rules
has earned rather less attention within the last decades [6]. Higher attention has
user-enabled formalization of arguments drawn [7].

c Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 157–171, 2018.
https://doi.org/10.1007/978-3-030-00178-0_10
158 B. Waltl et al.

The paper presents selected related approaches and ontologies for legal rea-
soning in Sect. 2. In Sect. 3 the model based formalization is introduced and
illustrated on a concrete example from the German tax law. The design and
implementation with a strong focus on end-user perspective of the resulting deci-
sion support system is summarized in Sect. 4. The automated analysis of decision
structures and their representation are presented in Sect. 5. Finally critical and
concluding remarks are discussed in Sect. 6.

2 Related Work
Within the last decades many attempts have been made to formalize legal sys-
tems, respectively legal rules. Thereby, the contributions made by the AI and
law community have significantly improved the understanding of the possibilities
and limitations of formalization [1].

2.1 A Quick Glance at Legal Expert Systems

Legal expert systems (LES) are well-studied throughout the domain of artifi-
cial intelligence and law [1]. Highly tailored to the legal domain the usage of
those systems was left up to experts with knowledge in both legal sciences and
computer science.
The reasoning within legal expert systems, can be differentiated into several
categories, such as rule-based systems (e.g., [7]), case-based reasoning (e.g., [8]),
neural nets (e.g., [9]), fuzzy logic (e.g., [10]), Bayesian networks (e.g., [11]), etc.
An additional step towards more comprehensive LES was reasoning on (legal)
ontologies [12–14]. Expressing logical constraints and relationships between those
knowledge objects can be done with description logics, e.g., web ontology lan-
guage (OWL). Thereby, OWL allows the definition of constraints and axioms in
ontologies. Many prior attempts used the W3C standard for modeling ontologies
based on RDF and OWL. OWL suffers of the possibility to formalize arithmeti-
cal or complex logic operations. Since OWL is an description logic, it was not
designed to be used for arithmetical expressions or as first-order (or higher order)
predicate language.
Latest approaches on the formalization of legal norms using LegalRuleML
have investigated structured processes to move from a natural language to a
controlled natural language focusing on the pragmatics of legal reasoning [15].
In the model proposed by Ramakrishna they need two roles, namely legal prac-
titioner and a knowledge engineer, to formalize the semantics in an independent
rule representation format, namely LegalRuleML. Another very interesting app-
roach was implemented by [16]. The authors focused on the implementation of
a rule management architecture integrating a defeasible inference engine.

2.2 End-User Oriented Decision and Reasoning Systems

Enabling end-users to understand, analyze, and model use cases, which are rel-
evant to them, e.g., user stories, has become an important paradigm in the
 Modeling, Execution and Analysis of Formalized Legal Norms 159

domains of software engineering. Thereby, the focus is on providing modeling

languages, i.e. notations, that are expressive enough to capture all relevant issues
of a particular domain while remaining simple enough to be used by end-users.
Well-known examples of those end-user oriented modeling languages are UML
(Unified Modeling Language [17]), BPMN (Business Process Modeling Nota-
tion [18]), CMMN (Case Management Modeling Notation [19]), or DMN (Deci-
sion Model and Notation [20]).
BPMN focuses on business modeling and specification [18]. Many attempts
have already been made to adapt the BPMN to enable automated compliance
checks regarding pre-defined executable rules [21]. The most recent standards
provided by the OMG (Object Management Group) have a more specific focus
on supporting modeling decision structures in work flows, e.g. business pro-
cesses or adaptive cases. Both, the CMMN and the DMN, provide end-users
with functionality to specify complex decision structures, such as decision tables
[20, Clause 8].
This brief sketch of the development shows how important the end-users
orientation and empowerment has been for the success of modeling notations.
It also shows, that the provision of executable semantics has always been an
important part of the standardizations efforts. Within the last years more and
more effort has been spent on formalizing work flows and decision structures.
This heavily increases the transparency of business processes and adaptive cases
and allows optimizations and leveraging of efficiency and effectiveness.

3 Model Based Formalization of Normative Regulations

3.1 The German Child Benefit Regulation

The German child benefit regulation is part of the tax law and can be formalized
using an ontological, i.e., model-based approach. The relevant articles from the
law are German tax income act §32, and §§62–78. §32 legally defines the term
“child” and what attributes are necessary for human beings to be considered
as children regarding the German tax law. We expressed the semantics of the
types, attributes and the relations in a UML class diagram (see Fig. 1).
The class diagram represents the user-defined semantic model. The result
of this model is still the result of a manual interpretation process, which can
be supported by analytics from the NLP framework, e.g., [22], to automatically
classify legal norms, such as obligations and permissions.
The semantic model consists of four different types, namely Taxpayer, Res-
idence, Child, and Employment. Each type has atomic attributes, indicated by
‘-’, and derived attributes, indicated by ‘/’. For the purpose of this model,
calculation of child benefit, it is sufficient for the Taxpayer to have only one
atomic attribute name of type string. In addition, the Taxpayer has two derived
attributes isQualified and sumChildbenefit. Both are defined through a MxL
expression, which is a strongly typed domain specific model based expression
language (see Sects. 3.2 and 4.1).
160 B. Waltl et al.

Child
-name: String
-dateOfBirth: Date
-childOfSpouse: Boolean
-grandchild: Boolean
Taxpayer -firstDegreeRelative: Boolean
-fosterchild: Boolean
-name: String -isDisabled: Boolean
claimsChildbenefit
/isQualified: Boolean 1 1...n /age: Number
/sumChildbenefit: Number /isQualifiedChild: Boolean

1
/amountForChild: Number

residesIn 1
isEmployed
1
1
Employment
Residence -name: String
-name: String -isEmployed: Boolean
-isJobseeking: Boolean
/isNationalTerritory: Boolean -inEducation: Boolean
-inInterimPeriod: Boolean
-inVoluntarySocialYear: Boolean

Fig. 1. A semantic model for the German child benefit claim.

3.2 Model Based Formalization

We show the appropriateness of our approach by a concrete example from the
German tax law, namely the determination if a taxpayer is eligible for retrieving
child benefit and the calculation of child benefit. Different conditions have to be
fulfilled, which are stated in German tax law.

Type Definitions. For the determination of the child benefit the German
tax law requires several different types, namely taxpayer, child, residence, and
employment.
t ∈ taxpayer (1)
cj ∈ child, j ∈ N (2)
r ∈ residence (3)
e ∈ employment (4)

Relations. The types have relations among each other, which have to be formal-
ized. Thereby, we can restrict our model to three different relations (see Fig. 1).
The relation between a taxpayer and its child (5), between a taxpayer and its
residence (6), and finally between a child and its employment (7).
claimsChildBenef it ⊆ taxpayer × child :
(5)
(t, cj ) ∈ claimsChildBenef it =⇒ kj is t s j th child
residesIn ⊆ taxpayer × residence :
(6)
(s, w) ∈ residesIn =⇒ taxpayer t lives in r
isEmployed ⊆ child × employment :
(7)
(cj , e) ∈ isEmployed =⇒ child cj is employed
 Modeling, Execution and Analysis of Formalized Legal Norms 161

Derived Attributes and Rules. Finally, the different types have attributes,
such as name, birth date, etc., that are required during the reasoning process.
Those attributes can either be atomic attributes or derived attributes. The first
kind of attributes describe those that inherently belong to a concrete type (e.g.,
name is of type string). The latter ones are those attributes that can be inferred
from other attributes. Consequently, derived attributes consist of the definition,
i.e. expression, containing the required information on how the attributes is
determined.
The following equations describe how the different attributes, which are
required during the determination of child benefit, are defined. The notation
for accessing an attribute of a type is the ‘.’ (dot). E.g., cj .dateOf Birth means
the dateOfBirth attribute from the j th child.

cj .age := N OW − cj .dateOf Birth (8)

cj .isQualif iedChild := cj .child§32 ∧
(cj .childOf Spouse
∨ cj .grandchild
∨ cj .f irstDegreeRelative
∨ cj .f osterchild) (9)
cj .child§32 := cj .age < 18
∨ cj .isDisabled
∨ cj .§32.4.1
∨ cj .§32.4.2 (10)
cj .§32.4.1 := (cj .age > 18 ∧ cj .age < 21)
∧ ¬e.isEmployed
∧ ¬e.isJobseeking (11)
cj .§32.4.2 := (cj .age > 18 ∧ cj .age < 25)
∧ (e.inEducation
∨ e.inInterimP eriod
∨ e.inV oluntarySocialY ear) (12)

The Eqs. (9)–(12) specify the different conditions that are defined by law
qualifying a child to be considered during the calculation for child benefit. The
claim can arise from different articles from the tax law. For example §32 states
that a child, which is younger than 18 or disabled is eligible. Beside, it is also
considered as child if the conditions in §32.4.1 or §32.4.2 are met.

C t := {cj ∈ child|(t, cj ) ∈ claimsChildbenef it

(13)
∧ cj .isQualif iedChild} f or t ∈ taxpayer
t.isQualif ied := r.isN ationalT erritory, (t, r) ∈ livesIn (14)
162 B. Waltl et al.

The Eqs. (13) and (14) ensure that a child is related to a taxpayer and that
the taxpayer lives within the national territory.

t.sumChildbenef it :=

cj ∈C t cj .amountF orChild if t.isQualif ied
(15)
0 if ¬t.isQualif ied
⎧
⎪
⎨190 if 1 ≤ j ≤ 2
cj .amountF orChild := 196 if 3 ≤ j ≤ 4 (16)
⎪
⎩
221 if 5 ≤ j

The remaining two Eqs. (15) and (16) determine the amount for the child
benefit based on the number of eligible children.

MxL - a Model Based Expression Language. The model based expression

language (MxL) as a type-safe domain specific language was developed in our
research group (see [23]) to support reasoning within a generic meta-model based
information system.
The meta model based information system incorporates the MxL for defining
executable semantics based on a semantic model. MxL users to apply simple (e.g.,
arithmetic) and higher-order functions (e.g., query operations), and to compose
them to complex and nested expressions. In addition it allows the access to
methods and operations implemented in Java and can easily be extended by
additional operators and functions.
An important property of MxL is its type-safety: It ensures that expressions
are valid regarding their static semantics and thus supports users in defining
consistent expressions with respect to the user-defined semantic model. Further-
more, MxL’s type-safety enables the system to resolve dependencies between
expressions, which in turn enables an automated adoption of expressions if ref-
erenced elements of the semantic model change.
In order to enhance the usability of MxL, we implemented helpful UI fea-
tures, e.g., syntax highlighting for better readability, auto-completion including
elements of the semantic model, and error localization in case of syntactic and
semantic errors.

4 The Model Based Legal Expert System

4.1 Executable Semantics of Legal Norms

To represent the content of legal norms in model based decision structures

two aspects need to be differentiated to fully capture the executable seman-
tics: Firstly, the structural semantics of legal norms (see Sect. 4.1). Secondly, the
behavioral semantics of legal norms (see Sect. 4.1).
 Modeling, Execution and Analysis of Formalized Legal Norms 163

Modeling Structural Semantics of Legal Norms. Modeling structural

properties of the norms’ content is necessary to capture mandatory legal con-
cepts with their attributes and relations. The meta model based information
system [23] empowers end-users to iteratively and collaboratively define seman-
tic models. End-users create and manage domain-specific types as well as corre-
sponding attributes and relations, e.g., a type Taxpayer with a relation residesIn,
or another type Child with the attribute dateOfBirth and the derived attribute
age. Since the semantic model is stored in a collaborative environment, i.e. ser-
vice oriented architecture for the model store, it can be revised by other users.
They can refine the model by adding additional types, attributes or relations.
They could also remove or add different kinds of constraints, e.g., cardinality
constraints (the number of minimal and/or maximal values an attribute has to
have), or type- and structure-related constraints, e.g., data types of attributes.
As illustrated in Fig. 1, the cardinality of the Taxpayer ’s relation Child is defined
to be 1...n, which allows a taxpayer to claim child benefit for different children.

Modeling Behavioral Semantics of Legal Norms. Besides the semantic

structure of legal concepts the behavioral aspects focus is on the interconnection
and the reasoning dependencies within structural components, namely types,
attributes, and relationships.

Fig. 2. Definition of the function §32.4.1. A short description, the input parameters,
the return type, and the MxL expression are provided.

The listing below shows the formalization of the normative structure deciding
on whether a child is eligible for child benefit or not. Thereby, the function uses
logical and arithmetical operations to combine numeric (e.g., greater-than) and
boolean attributes (e.g., or, and, not) of the child. It corresponds to the derived
attribute definition of Eq. 11 in Sect. 3 (Fig. 2).
(this.’#age’ > 18.0 and this.’#age’ < 21.0)
and
(not this.’hasEmployment’.isEmployed or
not this.’hasEmployment’.isJobseeking)
164 B. Waltl et al.

The decision structure belongs to a child instance, which can be accessed

using the this keyword. The semantics of the keyword is the same as in object
oriented programming languages, namely accessing the attributes or methods
from the same instance. Consequently, if the value of the derived attribute
cj .§32.4.1, cj ∈ child, j ∈ N is read, the decision structure above is evaluated
and the result calculated.
We can combine and nest the functions to define a function sumChildbenefit
(see below) determining how much a given taxpayer t receives. The function
reflects the Eqs. 13 (line 7), 15 (line 1 and 7), and 16 (line 4–6) of Sect. 3.
1 if not this.’isQualified’ then 0
2 else
3 let betrag = (children: Sequence) =>
4 if children.count() <= 2 then children.count() ∗ 190
5 else if children.count() = 3 then 2 ∗ 190 + 1 ∗ 196
6 else 2 ∗ 190 + 1 ∗ 196 + (children.count() − 3) ∗ 221
7 in betrag(find(Child).where(c => c.’isQualifiedChild’))
If a given taxpayer tp is not eligible, the function returns 0. Otherwise, the
function determines the number of children for which the taxpayer can claim the
benefit. Based on this result, the function calculates the amount that a taxpayer
will receive based on a formal calculation prescription derived from the law. This
small example demonstrates how partial expressions can be defined as separate
custom functions in order to formalize specific norms. It shows how to apply and
combine different kind of operations, e.g., conditionals (if-then-else) or logical
(and, or, not, etc.) and arithmetic (greater-than, summation, multiplication,
etc.) operations, in order to formalize behavioral aspects of normative texts.
In addition it shows the capabilities of MxL as a query language, that allows
complex queries to retrieve instances fulfilling a particular condition, such as all
children having the attribute isQualifiedChild set on true (see line 7 in listing
above).

4.2 Software Components of the Model Based Legal Expert System

Jandach [24] analyzed different notions of legal expert systems in 1993 with
a particular focus the concepts and characteristics that address LES for civil
law systems, more specifically the legal system in Germany. Several attempts
have been made to implement decision structures, arising from German legal
texts, into rule-based systems. However, hardly any attempt has been made to
formalize German laws using a model based, i.e., ontological approach, with a
reasoning engine that enables users to define expressions and infer knowledge
using propositional logic, first-order predicate logic, and arithmetical logic alike.
Based on Jandachs classification of the different components of legal expert
systems, namely knowledge base, inference engine, explanation component,
knowledge caption component, and dialog component, we designed a software
system, which’s components are shown in Fig. 3.
 Modeling, Execution and Analysis of Formalized Legal Norms 165

Fig. 3. System architecture consisting of a model storage, model execution components,

and interaction components.

The system’s components can be classified into three different groups, namely
a model store, a model execution component, and an interaction component.
Each group is implemented by multiple different software components.

Model Store. The model storage component contains the definition of the
model, i.e., ontological description, and the facts provided by the end-user.

Model. A model is described by its types with their attributes, i.e., scheme, and
the relations between types. Our implementation differentiates between two
types of attributes: atomic attributes and so call derived attributes. Atomic
attributes consist of concrete values and have a basic data type, such as number,
date, text, enumeration, boolean, and sequence. In contrary, derived attributes
are expressed as rules, formalized in a model based expression language (MxL),
which is a strongly-typed and functional domain specific language (DSL).
Facts. The instantiation of a model is done through the provision of facts. Those
facts are stored as explicit records in the model store. Each model instance
has a unique identifier and name which is used for unambiguous identification.
An instance does not need to assign a value to each attribute. The attributes
are optional and null-value (empty attributes) are allowed.

For our implementation we used a meta-model based information system to

implement the model storage component. This meta-model based information
system holds all the information about the scheme of the ontology, i.e., model,
as well as the instances (facts) of a concrete model. The system allows the
formalization of different models, that are logically separated into disjoint
workspaces (data rooms).
166 B. Waltl et al.

Model Execution Component. The model execution components are built

on top of the model storage and accesses the database of facts, i.e., instantia-
tion of types with concrete values for the available attributes, and the database
containing the information about the schema, i.e., types and attributes.

Inference Engine. The reasoning on the given facts considering the formalized
rules requires access to the database of facts and the storage holding the infor-
mation about the expressions required to determine the derived attributes. The
input parameters and the return values of those expressions are strongly typed.
The inference engine is developed in Java and retrieves data from the meta model
based information system. The MxL was designed using “Beaver - a LALR
Parser Generator”1 . The inference engine offers end-users to define semantics of
derived attributes in functional expressions, and allows expression of first and
second order logic as well as the definition of complex queries (projection, selec-
tion, and transformation).
Inference Analysis. Closely connected to the inference engine is the inference
analysis component. This component allows the inspection of complex expres-
sions. On the one hand it is possible to get the information about the abstract
syntax tree (AST) of an expression. It allows overviews on the provided and
derived facts in an object diagram like visualization (see Sect. 5.1). On the
other hand the component offers functionality to view complex data flows
based on the input parameters to inspect the resulting derived attribute (see
Sect. 5.2).

Interaction Component. The improvements of modern software systems with

regard to user experience can be considered as one of the main successes of soft-
ware engineering in the last decade. End-user development and human-centered
design thinking have become an established methodology in designing and imple-
menting software systems. Beside the provision of bare functionality, such as
logical reasoning, LES need to provide user interfaces that do not overexert
users. Instead it will be the challenge for the legal informatics domain to enable
end-users to use LES and leverage the full potential that they offer.
In our system the modeling component is separated from the knowledge
acquisition component and the explanation dialog:

Modeling Component. Users, e.g., legal data scientists or legal knowledge

engineers, are supported during the creation of the model with its attributes
and relations by an appropriate modeling component. The modeling compo-
nent offers functionality to add new types to a model and easily add attribute
definitions and derived attribute definitions (MxL). Thereby, the users are
supported with a graphical user interface that runs completely as web
application.

1
http://beaver.sourceforge.net/, last access on 03/01/17.
 Modeling, Execution and Analysis of Formalized Legal Norms 167

Knowledge Acquisition Component. The insertion of facts is done in a

separate component, which has been exclusively designed for this purpose.
Thereby, types and attributes, describing a case, can be added by the users.
Since the model attributes are typed a first type checking is performed within
the interface. It is not possible to insert not compatible types, such as text if
number is required. The knowledge acquisition components show the inferred
values of the derived attributes.
Explanation Dialog. To reconstruct the conclusion that was made by the sys-
tem, users need functionality providing them with information about the rea-
soning procedures. This explanation dialog visualizes the information from the
inference analysis component. Thereby, given a particular derived attribute
the user gets information about the underlying MxL expression and the
abstract syntax tree showing the different data values and operators that
contribute to the overall result.

4.3 Execution and Computation of Models in a Collaborative

Environment

The three interaction components of our system, namely modeling component,

explanation dialog and the knowledge acquisition component (see Fig. 3), are
developed as collaborative web application components. Each of them can be
used through a web browser. They access a shared model store containing the
model and the instances via a REST API. This enables users to create new
models, i.e., types, attributes, and relations, and share them with other users.
Additionally, it is possible to maintain and refine existing models that a user has
created prior or that another user created.
The creation of decision structures through the system components is sepa-
rated from the components handling the instantiation of models and their execu-
tion. The knowledge acquisition component, shown in Fig. 4, is structured in to
three views: the model view on the left, the form view in the middle, and the
document view on the right:
Model view. The model view is an interactive JavaScript component providing
an overview of the semantic model. It enables users to select and unselect
types, such as the ’Taxpayer’ (see Fig. 4), which is highlighted and to which
details are shown in the form view and document view. This view shows the
relations between types and the cardinalities of those relations. Note, that
multiple relations can exist between two types although this is not shown in
Fig. 4.
Form view. The area in the middle of the screen is the form view which is
vertically divided into 4 regions, namely the title of the selected type, a list
of attributes, a list of relations, and a list of derived attributes. This view
offers the user the functionality to graphically add new model instances and
provide them with proper facts (attributes and relations). The attributes
form the atomic properties that belong to a particular type and the references
indicate the relations to instances of another type, e.g., ‘Residence’ or ‘Child’.
168 B. Waltl et al.

The derived attributes list shows the evaluated MxL expressions. In addition
it offers the functionality to access the explanation dialog by clicking on the
question mark beside the input box of a derived attribute (see Sect. 5.1).
Since all attributes are strongly typed, the user interface prevents the input
of wrongly typed facts, e.g., number instead of boolean.
Document view. The document view allows access to information from a tex-
tual resource that was used during the creation of the model. This is con-
sidered be valuable since it provides additional information to the form view
based on the legislative or jurisdictional texts. Figure 4 shows article 1 entitled
“Steuerpflicht” (engl. tax liability) from the the German tax law.

The knowledge acquisition component is a central view of the system and allows
the creation of new model instances and the inspection of existing ones. The
modeling component is strictly separated from this view. It is not possible to
modify the model in the knowledge acquisition phase.

Fig. 4. The front-end of the knowledge acquisition component consisting of three

areas: (a) model overview (left), (b) a form-based input dialog distinguishing between
attributes, relations and derived attributes (middle), and (c) a set of textual documents
to provide additional information about the decision structure (right).

It is necessary to carefully distinguish between the interpretation and appli-

cation, e.g., subsumption, of a legal norm. Transferring the idea of interpretation
of norms to the model based reasoning approach, the interpretation can be con-
sidered as the modeling phase. During the modeling phase a legal expert or a
group of legal experts formalize the decision structures and specify the required
types, attributes, and relations. During a subsequent subsumption phase, legal
experts provide the system with facts that constituent themselves in instances of
types with concrete attributes and relations. Given this instantiation as input,
the model execution component deductively derives new information. It is not
 Modeling, Execution and Analysis of Formalized Legal Norms 169

necessary that every attribute of an instance is provided the execution compo-

nent can reason on partially filled types and attributes as well. It is designed to
determine as much new information as possible based on the given input.
Once the information of the reasoning structures are modeled and the
instances with their attributes and relations are provided, the system is capable
of automatically inspecting the reasoning structures regarding data flows and
dependencies, which is going to be discussed in the next chapter.

5 Analysis of Decision Structures for End-Users

5.1 Attribute-Based Dependency Trees
Derived attributes, which semantics are specified using the model-based expres-
sion language, can potentially consist of complex and nested functional expres-
sions. Those expressions are evaluated before the attribute is accessed in the
model store. This read event triggers the evaluation and the execution of the
expressions.
A dependency tree for a given derived attribute can automatically be gener-
ated by the inference analysis component and forwarded to the dialog component
which renders the resulting tree. Thereby, the leaves of the given tree denote
either constants, atomic variables, such as ‘isEmployed’ or ‘isJobseeking’ (both
boolean), or other derived attributes, such as ‘#age’. The nodes within the tree
constitute the various logical and arithmetical operations, such as ‘GreaterThan’
or ‘And’, that evaluated during the execution.
The dependency tree view is widely used in semantic program analysis. It
enables end-users to understand and reconstruct how the different input param-
eters contribute to the final result. Consequently, this view fosters the compre-
hensibility and traceability of a complex decisions structures.

5.2 Interactive Data Flow Graphs

In addition to means for collaborative modeling, model based reasoning, and
inspection of attribute-based dependency trees the system offers a holistic per-
spective to view and explore the data flow throughout a semantic model. This
data flows can again be represented as directed graphs, where the nodes represent
attributes or types and the edges represent the usage or contribution within a
subsequent type or derived attribute. Thereby, the inference analysis component
inspects the data flow during the execution of expressions and passes this infor-
mation to the dialog component, which renders and visualizes the interactive
graphs.
For instance, the data flow graph allows the analysis of contributing variables
to the derived attribute ‘isQualifiedChild’ of a child (Eqs. 9–12 in Sect. 3.2).
Thereby, it can be seen which attributes influence each other and how it is
determined whether a child is eligible for retrieving a benefit or not. It also
allows the inspection, which facts are given by a particular instance and how a
change in the facts would influence the different attributes resulting into claims
that someone has or has not.
170 B. Waltl et al.

6 Conclusion
This paper is a contribution enabling end-users to create executable models rep-
resenting the semantics of statutory texts, i.e. laws. Based on the theory and prior
approaches in the domain of artificial intelligence, in particular legal expert sys-
tems and ontologies, we designed an implemented a model based expert system
focusing on the end-users perspective. We divided the system into three different
components: a model store, a model execution component, and an interaction
component.
The model store is a meta-model based information systems persisting the
model, i.e. schema, consisting of types, attributes, derived attributes, relations
and the facts, i.e. instances of the model. The model execution component is a
deductive reasoning engine providing a domain specific language enabling end-
users to create executable rules defining the derived attributes. The interaction
component is a web based application fostering collaborative access to the model
store to create and maintain models and to provide facts. In addition the system
has components to analyze the dependency tree of derived attributes and the
data flow within a model.

References
1. Bench-Capon, T., et al.: A history of AI and Law in 50 papers: 25 years of the
international conference on AI and Law. Artif. Intell. Law 20, 215–319 (2012)
2. Sartor, G.: Legal Reasoning: A Cognitive Approach to the Law, Ser. A Treatise of
Legal Philosophy and General Jurisprudence. Springer, Dordrecht (2005)
3. Rissland, E.L., Ashley, K.D., Loui, R.P.: AI and law: a fruitful synergy. Artif. Intell.
150(1–2), 1–15 (2003)
4. van Engers, T.M., van Doesburg, R.: First steps towards a formal analysis of law.
In: Proceedings of eKNOW (2015)
5. Francesconi, E. (ed.): Semantic Processing of Legal Texts: Where the Language of
Law Meets the Law of Language. Springer, Heidelberg (2010)
6. Liao, S.-H.: Expert system methodologies and applications–a decade review from
1995 to 2004. Expert Syst. Appl. 28(1), 93–103 (2005)
7. Prakken, H., Sartor, G.: Law and logic: a review from an argumentation perspec-
tive. Artif. Intell. 227, 214–245 (2015)
8. Ashley, K.D., Rissland, E.L.: A case-based approach to modeling legal expertise.
IEEE Expert 3(3), 70–77 (1988)
9. Aikenhead, M.: Uses and abuses of neural networks in law. Santa Clara Comput.
High Tech. LJ 12, 31 (1996)
10. Gerathewohl, P.: Erschließung unbestimmter Rechtsbegriffe mit Hilfe des Comput-
ers: Ein Versuch am Beispiel der angemessenen Wartezeit bei §142 StGB. Disser-
tation, Eberhard-Karls-Universität, Tübingen (1987)
11. Timmer, S.T., Meyer, J.-J.C., Prakken, H., Renooij, S., Verheij, B.: A structure-
guided approach to capturing Bayesian reasoning about legal evidence in argumen-
tation. In: Proceedings of the 15th International Conference on Artificial Intelli-
gence and Law, pp. 109–118 (2015)
12. Francesconi, E. (ed.): Proceedings of LOAIT 2010 -: IV Workshop on Legal Ontolo-
gies and Artificial Intelligence Techniques
 Modeling, Execution and Analysis of Formalized Legal Norms 171

13. Wyner, A.: An ontology in OWL for legal case-based reasoning. Artif. Intell. Law
16(4), 361–387 (2008)
14. Casanovas, P., Biasiotti, M. A., Francesconi, E., Sagri, M.T. (eds.): Proceedings
of LOAIT 2007: II Workshop on Legal Ontologies and Artificial Intelligence Tech-
niques (2007)
15. Ramakrishna, S., Paschke, A.: A process for knowledge transformation and knowl-
edge representation of patent law. In: Bikakis, A., Fodor, P., Roman, D. (eds.)
RuleML 2014. LNCS, vol. 8620, pp. 311–328. Springer, Cham (2014). https://doi.
org/10.1007/978-3-319-09870-8 23
16. Islam, M.B., Governatori, G.: Ruleoms: a rule-based online management system.
In: Proceedings of the 15th International Conference on Artificial Intelligence and
Law. ACM, pp. 187–191 (2015)
17. Object Management Group: Unified Modeling Language (UML) 2.4.1 Infrastruc-
ture. http://www.omg.org/spec/UML/2.4.1/
18. Object Management Group: Business Process Model and Notation (BPMN), Ver-
sion 2.0
19. Object Management Group: Case Management Model And Notation Version 1.0
20. Object Management Group: Decision Model and Notation Version 1.0
21. Governatori, G. (ed.): Law, logic and business processes. In: 2010 Third Interna-
tional Workshop on Requirements Engineering and Law (RELAW) (2010)
22. Waltl, B., Matthes, F., Waltl, T., Grass, T.: LEXIA: a data science environment
for Semantic analysis of German legal texts. Jusletter IT (2016)
23. Reschenhofer, T., Monahov, I., Matthes, F.: Type-safety in EA model analysis. In:
IEEE EDOCW (2014)
24. Jandach, T.: Juristische Expertensysteme: Methodische Grundlagen ihrer Entwick-
lung. Springer, Heidelberg (1993). https://doi.org/10.1007/978-3-642-84978-7
 Causal Models of Legal Cases

Ruta Liepina1(B) , Giovanni Sartor1,2 , and Adam Wyner3

1
Law Department, European University Institute, Florence, Italy
Ruta.Liepina@EUI.eu
2
CIRSFID, University of Bologna, Bologna, Italy
3
School of Law and Department of Computer Science, Swansea University,
Swansea, UK

Abstract. Legal causation is a complex aspect of legal reasoning. Due to

its significant role in the attribution of legal responsibility, it is important
that there is a clear understanding of the requirements for establishing
and reasoning with causal links. This paper presents preliminary results
of modelling causal arguments based on the legal decisions with partic-
ular focus on physical causation. We introduce a semi-formal framework
for reasoning with causation that uses strict and defeasible rules for mod-
elling factual causation arguments in legal cases. We further discuss the
complex relation between formal, common sense, norm and policy based
considerations of causation in legal decision making with particular focus
on their role in comparing alternative causal explanations.

Keywords: Causation · Law · Evidence · Legal reasoning

1 Introduction
This study is inspired by the theoretical challenges of establishing causation.1
We focus on cases addressed by the US Vaccine Injury Court [17] and in par-
ticular the Althen’s decision [1]. In vaccine cases a Special Master has the task
of ‘determining the types of proceedings necessary for presenting the relevant
evidence and ultimately weighing the evidence in rendering a final, enforceable
decision’ [2]. While the Special Master has the training and experience in address-
ing causation in vaccine cases, uncertainties remain, concerning not only specific
decisions on causality, but also the criteria to be used in such decisions. In
fact, in the Althen’s decision, the Special Master affirmed that the criteria for
establishing a satisfactory causal connection is ‘an unresolved legal issue’; and
moreover, he emphasised that ‘without articulate standards providing guidance
[on causation], the experts bring their own beliefs and biases into the courtroom’.

1
The current article is version of our semi-formal causal argumentation framework,
which was presented at AICOL Workshop, JURIX 2017, Luxembourg, Luxembourg.
Another version was presented at the Evidence and Decision Making Workshop in
ICAIL 2017, London, United Kingdom; that work is being reworked for publication.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 172–186, 2018.
https://doi.org/10.1007/978-3-030-00178-0_11
 Causal Models of Legal Cases 173

With causation being an active area of research and having practical impli-
cations, this paper discusses the connection between the cause-in-fact and legal
causation as observed in the case law. Our study aims to provide analysis of the
arguments establishing the cause-in-fact. This paper contributes to the current
state of art by providing a logic-based approach designed for causal argument
analysis in the law. While our study is at its early stages, it has already shown
that it is possible to capture important aspects of causal reasoning in the law.
Despite many more challenges that have to be addressed before causal issues
can be fully understood in the law, this is a step towards argument modelling
and assessment with the potential to support practical analysis of civil law cases
(especially those relying on expert testimonies).
We present our analysis in three main sections. In Sect. 2, we begin by intro-
ducing the relevant theories of causation and the facts of the case study. We
focus on the legal and formal approaches of causation. These use various logic
and common sense reasoning tools and have the potential to be applied in law. In
Sect. 3, we introduce a semi-formal framework for modelling causal arguments in
legal cases with notions of causal and evidential links as well as strict and defea-
sible rules. We present an application of our framework through a case analysis
of Althen’s vaccine injury case. We model two conflicting causal arguments as
presented by the medical experts in the case. We comment on the use of causal
language and logical structures present in the expert testimonies. In Sect. 4, we
further discuss the complex relationship between formal, common sense, norm
and policy based considerations of causation in legal decision making, with par-
ticular focus on their role in comparing alternative causal explanations. In Sect. 5,
we conclude with discussion of future developments of our framework, including
norm based extensions.

2 Background
Our interdisciplinary study aims to bridge the formal theories of causation and
the needs of the practical field of law. As many other areas that concern them-
selves with causal issues, law has adapted various domain specific measures to
deal with causality [12]. In this section, we introduce two current views on causa-
tion: Hart and Honoré’s NESS [9,19], which represents the legal perspective, and
Halpern and Pearl’s [8] formal theory of ‘actual causation’. We also introduce
the facts and legal norms of the working case example.

2.1 Theories of Causation

Causation is a theoretically rich field [3], but not all discussions are relevant
in legal analysis. Here, we discuss two groups of approaches of causal analysis,
starting with Hart and Honoré’s [9] theory of necessary and sufficient conditions
in common sense analysis in law. We then present a formal theory of causation,
Halpern and Pearl’s [8] actual causation.
174 R. Liepina et al.

Legal Causation: Hart and Honoré’s Approach. With ‘Causation in the

Law’, Hart and Honoré [9] shaped the discussions of causation in the field of legal
theory in at least two major ways: (1) through the distinction between the cause-
in-fact and legal causation in legal responsibility attribution; and (2) through
defining a more restricted version of the existing counterfactual approaches in the
law2 based on the conditionality of necessary and sufficient conditions [9,12,19].
Cause-in-fact in law is established based on the factual information available in
the case. However, only some of the causal links are sufficient for attributing
legal responsibility in the case. These legal causes are governed by the rules and
policy considerations. Following the definition of legal responsibility attribution
in [12], it can be satisfied on the following three grounds: ‘(a) the conduct of
the person; (b) the causal connection between the conduct of the person and
the given harm; and (c) the fault legally implied by the conduct of the per-
son’. While (a) and (c) are given, it is the task of the judge and the jury to
reason about the causal connection defined in (b). A number of things, such
as multiple causes or absence of action under the duty of care can increase the
complexity of establishing a causal link and consequently complicate the task of
legal responsibility attribution. An example of over-determination in the medi-
cal domain can occur when there is more than one contributor to the patient’s
injuries. Moreover, there can be issues with foreseeing the effects of how certain
types of substances react with medical preconditions. While it might be possible
to determine the physical causes in such cases with some level of accuracy, for
policy reasons the judge in the case might wish to lower the standards of the
proof to allow for the victim to be compensated based on non-causal considera-
tions, such as fairness, or suffering of the patient. It is important to distinguish
between the different types of causation in order to design a framework that
supports legal reasoning. The second contribution by Hart and Honoré is their
discussions about the necessary and sufficient conditions for the outcomes. The
supporters [19] of this view argue that in ‘a specific situation a causally relevant
condition is a necessary element of a set of conditions jointly sufficient for the
harmful outcome’, or the so-called NESS test, where the cause is a necessary
element of a sufficient set of conditions. The NESS approach allows one to pick
out the cause with more precision compared to the sine qua non approach. In
order to compare alternative causal models, the NESS approach appeals to the
idea of generalisations about causal relations [10], where causal models would be
based on instantiations of such generalisations to compare alternative models.

Causation in AI and Law: Halpern and Pearl’s Approach. Causation

has been described as one of the essential bonds of the universe as perceived and
understood by humans [3]. Indeed, it seems inconceivable that everyday or spe-
cialised reasoning could be coherent without the causal elements. In the sub-field
of artificial intelligence and law, causation holds a central role in understanding
and modelling legal reasoning and legal responsibility attribution [12]. This mat-
2
The most common approach in law being sine qua non: ‘but for the action, the result
would not have happened’.
 Causal Models of Legal Cases 175

ter has been approached from various perspectives, including logic [11,13,16],
statistics [5], argumentation [17,18], and new task specific approaches [6,14].
These approaches are not exclusive, and we argue that a combination of these
and legal approaches introduced above provide a more comprehensive solution
for causal analysis in the law. Our framework was inspired by several of the
current state of art approaches, but especially by the considerations of Halpern
and Pearl’s ‘actual causation’.
Halpern and Pearl’s [7,8] ‘actual causation’ is a formal approach to causal
analysis that allows for comparison of alternative explanations based on normal-
ity, typicality and default notions. What distinguishes this approach from others
is the method of modelling causal relations through interventions. Halpern and
Pearl try to avoid the ambiguity in determining the cause and effect by mod-
elling causal events through structural equations. Actual causation can model
more complex causal relations and is a good competitor with the NESS test.
Furthermore, Halpern and Hitchock have developed an extension of the theory
that allows for alternative causal explanations to be compared [8], based on the
notions of normality, defaults and typicality. The idea behind this extension is
to compare the alternative causal explanations based on the closeness to what
has been defined as the normal state of the relevant events. For instance, when
multiple agents could have prevented a harmful event, causal responsibility could
be attributed to the omission of the agent who had the obligation to prevent it.
The authors claim that it can be based on various criteria, including statistical
data, moral norms, and prescribed norms. The latter are especially relevant to
considerations in law.

Towards a Semi-formal Approach. The approaches we have just presented

show the complexity of causal attribution. Any adequate model of causality has
to take into account the issues addressed by such authors. In particular the
model by Halpern and Pearl appears to be the most promising, and we aim to
expand our approach so that it includes the ideas developed by these authors.
However, we think that for modelling many legal cases, a simpler approach
is sufficient. Even if the model will not yet capture some causal issues (overde-
termination, pre-emption, etc.), it may help us to understand cases that do not
involve such subtleties. We need a sufficient representation that works with the
presentations of information in the cases. This is the purpose of the model that
will be presented in the next sections.
Below, we introduce the working example of the Althen’s case. We then pro-
vide our framework for capturing the causal reasoning, which is later applied to
the Althen’s case. Finally, we contrast the various approaches to causal models
in Sect. 4, showing how the formalism can aid legal discussion and vice versa.

2.2 The Althen’s Case

The Althen’s Special Master’s decision [1] is our source for evidentiary and legal
issues in relation to causation. We focus on how the cause-in-fact is established
176 R. Liepina et al.

based on the conflicting testimonies from the expert witnesses explaining the
symptoms observed on Mrs. Althen, and how the special master reasons about
these causal arguments to attribute legal responsibility in the case.
Facts of the Case: The case concerns Mrs. Althen (petitioner) and her worsening
health conditions after receiving a tetanus toxoid (TTV) vaccine. Prior to the
vaccines, petitioner was reasonably healthy. After roughly two weeks from the
vaccination date, petitioner started reporting various symptoms ranging from
blurred vision, to steady headache and temporal loss of vision. The complaints
and hospital visits continued for the following three years. Petitioner underwent
various types of treatments and extensive medical testing (MRI, EEG, blood
tests) that showed inconclusive results for acute-disseminated encephalomyelitis
(ADEM), multiple sclerosis, and vasculitis. Petitioner subsequently applied to
the Vaccine Court for compensation, which requires that petitioner establishes
a causal link between the TTV and ADEM.
Legal Criteria: For a causal link to be successfully proven in the US Vaccine
Court, the vaccine either has to be listed on the approved ‘Vaccine Injury Table’
or the petitioner has to satisfy the causation-in-fact conditions set out by the
court (off-table vaccine injury). This case involves an off-table vaccine injury.
This case involves an off-table vaccine injury. The claimant’s burden (also known
as the Stevens test) can be summarised as follows: “the claimant has to show
by preponderant evidence that the vaccination brought about her injury by
providing a medical theory causally connecting the vaccination and the injury, a
logical sequence of cause and effect showing that the vaccination was the reason
for the injury, and a showing of a proximate temporal relationship between
vaccination and injury; if the claimant satisfies this burden, she is entitled to
recover unless the government shows, also by a preponderance of evidence, that
the injury was in fact caused by factors unrelated to the vaccine”.
In this paper, we mostly focus on the criterion that requires a ‘logical sequence
of cause and effect’. We claim that our semi-formal model is able to support
expert witness testimonies in meeting this criterion.
Expert Witness Testimonies: There were three expert witnesses assigned to this
case, of which we discuss two competing testimonies. Dr. Smith for petitioner
arguing for the positive causal link between TTV and ADEM, and Dr. Safran
for respondent rejecting the causal link.
The experts agreed that it is theoretically possible that the TTV can cause
ADEM. However, the main disagreement was whether in this particular scenario
the TTV which instigated the reaction, where healthy cells were attacked and
demyelinated. Dr. Smith supported his opinion by referring to the medical theory
of cell degeneracy, which shows cell modification through evolution and can
explain T cell reaction after a vaccine. The theory of cell degeneracy is based
on the premise that TTV can attack and destroy both bad (antigen) and good
(myelin) cells, as the antigen and myelin are sufficiently similar. Dr. Safran,
while accepting the general causal theory of degeneracy, denies the link in this
particular case on two grounds. Firstly, he argues that antigen and myelin cells
 Causal Models of Legal Cases 177

are not proven to be sufficiently similar. Secondly, the symptoms observed on the
patient can be better explained by multiple sclerosis. Based on this information
that has been abstracted from the case, we will model the causal explanations
given by the experts. We decided to focus on the expert testimonies due to
their core value in decision making in the vaccine courts. Furthermore, these
testimonies provide detailed information on the causal claims made in the case,
and the special master carefully examines them and weights them for the legal
responsibility attribution.

3 Semi-formal Framework and Case Analysis

In this section, we present a semi-formal framework for assessing causal links.

Then we use this framework to capture the views expressed by Dr. Smith and Dr.
Safran on the causal relation between TTV and ADEM. Our analysis is meant to
accommodate the causal relations that are presented as the core disagreements
between the parties and shed some light on the structure of the arguments pro-
vided by the expert witnesses. Other matters are out of scope at this moment,
in particular, the dynamics of dialogue, belief change, and time. Rather, it is a
static model that lays out all the information available. It makes explicit some
of the otherwise implicit assumptions that are highlighted in the course of the
presentation of the expert testimony and by way of attack.

3.1 Semi-formal Framework

Our semi-formal framework is a propositional language to represent and reason

with basic facts/events and causal relations between them. Our model has basic
knowledge structures and rule schemes using the structures.

Basic Knowledge Structures. There are several basic knowledge structures:

factual propositions; predicates for similarity and evidentiality; causal links; and
inference rules.

Factual Propositions. A factual proposition has the form

±H(X)

where X is a positive or negative literal. H(X) states that the literal X holds
and ¬H(X) that this was not the case. For instance H(T T V ) and H(¬T T V ) are
such literals, affirming respectively, that it holds that tetanus toxoid vaccination
was performed, and that it holds that this was not the case. We assume that
H(X) and H(¬X) are incompatible, and that H(¬X) is indeed equivalent to (it
strictly implies and is implied by) ¬H(X). For instance H(¬T T V ) is equivalent
to ¬H(T T V ).
178 R. Liepina et al.

Similarity and Evidentiality. The case scenario also includes reasoning with sim-
ilarity and evidentiality, which we introduce as a relation and a predicate, respec-
tively. For similarity, we have
Sim(X, W )
which is understood as the propositional content of X is similar to the proposi-
tional content of W (and vice versa). For evidentiality, we have

EV (X)

which is understood to represent the existence of evidence for the propositional

content of X.

Causal Links. Causal links are represented as

C(L, X, Y )

where L takes values 1 or 2, and X and Y are factual propositions. L = 1 indi-

cates that cause X always (necessarily) produces effect Y , and L = 2 indicates
that cause X usually (normally) produces effect Y . The causal links and levels
of causal links loosely correspond to the language used by the reasoners.3 For
instance, C(1, T T V, T CellAct) is an expression that T T V and T CellActivation
have the always causal relation, as the witnesses present it.

Inference Rules. We use strict and defeasible inference rules represented as

X1 . . . Xn → Y

for strict rules, and

X1 . . . Xn ⇒ Y
for defeasible rules, where X1 , . . . Xn , Y are factual propositions.
We assume that strict rules allow for contrapositive reasoning. For our pur-
poses it is sufficient the contrapositive inference where, given all antecedents of
a defeasible rule except X, and the negation of the rule’s consequent, we can
infer X’s negation. Thus we assume that a strict rule stands for the set of all of
its contrapositions. A contrapositive of a strict rule

X1 . . . Xn → Y

is a rule
X1 . . . Xi−1 , ¬Y, Xi+1 , . . . Xn → ¬Xi
3
At the initial stages of the study, we annotated the decision identifying causal and
accompanying hedging expressions. After identifying the main causal links in the
case, we ranked the various expressions in two levels of strength. For instance, ‘a
probable causal relation between tetanus toxoid and two injuries’, ‘it is more prob-
ably than not the case that tetanus toxoid can cause the injuries suffered here’ are
examples of level 2 (usually causal) support.
 Causal Models of Legal Cases 179

For instance, if a knowledge base contains the rule

H(X) ∧ C(1, X, Y ) → H(Y )
it also contains a rule
¬H(Y ) ∧ C(1, X, Y ) → ¬H(X)
Accordingly, if Y does not hold and X strongly causes Y , we can infer that X
does not hold, for otherwise Y should hold.
To capture the reasoning of the experts in our cases, we need to supplement
the general framework above with some schemes for rules.

General Rule Schemes. There are general rule schemes for causal rules, sim-
ilarity, and evidence (including abduction).

Causal Rules. Rule schemes R1 and R2 capture the inferential relevance of causal
relations
[R1.] H(X) ∧ C(1, X, Y ) → H(Y )
captures the inferential meaning of necessary causal links. If the precondition
holds and the causal link is level 1, we strictly conclude that the effect holds.
[R2.] H(X) ∧ C(2, X, Y ) ⇒ H(Y )
captures the inferential meaning of defeasible causal links. If the precondition
holds and the causal link is level 2, we defeasibly conclude that the effect holds,
but can have exceptions.

Similarity. The case scenario also includes reasoning with similarity, in partic-
ular, it is assumed that similar facts cause the same effect. This is captured by
the following scheme for defeasible conditionals:
[R3.] Sim(X, W ) ∧ C(L, X, Y ) ⇒ C(L, W, Y )

Evidence. Finally, our experts reason with evidence. First, it is assumed that
having evidence for a proposition defeasibly implies that this proposition holds:
[R4.] EV (X) ⇒ H(X)
Evidence plays two roles in our domain. First, it may establish facts that
contradict the conclusion of a causal inference. This will defeat the causal argu-
ment leading to the conclusion contradicted by the evidence. If that argument
culminates with the effect of a level 1 causal relation, defeat may extend to the
precondition of that causal relation via contraposition.
Second, evidence may provide abductive support for establishing the
antecedent of a level 1 causal relation. In fact, it seems that our experts assume
that given a necessary causal relation and its effect, we can abductively infer the
precondition. We capture this aspect of their reasoning through the following
pattern defeasible conditionals:
180 R. Liepina et al.

[R5.] H(Y ) ∧ C(1, X, Y ) ⇒ H(X)

meaning, that we can defeasibly infer the precondition X when we have estab-
lished its effect. For instance, given evidence of certain symptoms, we can defea-
sibly infer that these symptoms hold by R4; moreover, if such symptoms are
strictly caused by an illness, we can defeasibly infer that the illness holds by R5.

3.2 Models in the Language

In this section we apply the semi-formal framework to represent the relations as
described in the case by the expert witnesses.
We shall use the following atoms, representing the circumstances that are
relevant to our case:

Factual Atoms
T T V - tetanus toxoid vaccination being injected in the patient
ADEM - acute-disseminated encephalomyelitis illness
M S - multiple sclerosis
T CellAct - the chemical process of tetanus toxoid vaccination activating T
cells
AntigDestr - antigen cells destroyed (T cells should target this, wanted effect)
M lnDest - myelin cells destroyed (T cells should not target this, unwanted
effect)
Symp(M ono) - monophasic symptoms occur. This means that the symp-
toms that occurred just once. −Symp(M ono) means here that the symptoms
occurred on multiple occasions.
We use the following causal links to represent the causal claims of the expert
witnesses:
C1. Tetanus toxoid vaccination always causes T-cell activation: C(1, T T V,
T CellAct)
C2. T-cell activation always causes antigen destruction: C(1, T CellAct,
AntigDestr)
C3. T-cell activation usually causes myelin destruction: C(2, T CellAct,
M lnDestr)
C4. Myelin destruction usually causes ADEM: C(2, M lnDestr, ADEM )
C5. ADEM usually causes monophasic symptoms: C(2, ADEM, Symp(M ono))
C6. Multiple sclerosis usually causes recurrent symptoms: C(2, M S, −Symp
(M ono))

3.3 The Arguments of the Expert Witnesses in Althen’s

We now model the arguments of Dr. Smith and Dr. Safran in Althen’s.

Dr. Smith. Given the language and rules above, we model the reasoning by
Dr. Smith. He argues for a causal link between the TT vaccine and ADEM.
 Causal Models of Legal Cases 181

The argument starts from the explicit assumption of T T V (given as a fact) and
uses the strict and defeasible rules to model the reasoning process and justifica-
tion of linking the vaccination, symptoms and illness. In particular, Dr. Smith’s
argument is based on the explicit assumption Sim(AntigDestr, M lnDestr) and
implicit assumption C(2, ADEM, Symp(M ono)) (an assumption exposed by
another witness’ counter-argument). We take these assumptions as assertions in
Dr. Smith’s model. We give the assumptions, rules, and inferences in sequence,
labelling the reasoning steps by agent.

Dr. Smith’s Model: The rules for Dr. Smith’s argument, based on the rules
proposed above:

1. H(T T V )
2. C(1, TTV, TCellAct)
3. H(T T V ) ∧ C(1, T T V, T CellAct) → H(T CellAct)
4. H(T CellAct) (from 1, 2, and 3)
5. C(1, T CellAct, AntigDestr)
6. Sim(AntigDest, M lnDest) (similarity assumption by Smith)
7. Sim(AntigDest, M lnDest)∧C(1, T CellAct, AntigDestr) ⇒ C(2, T CellAct,
M lnDestr) (an instance of the rule scheme that similar effects are caused
by the same cause)
8. C(2, T CellAct, M lnDestr) (from 5, 6, 7)
9. H(M lnDestr)
10. C(2, M lnDestr, ADEM )
11. H(M lnDestr) ∧ C(2, M lnDestr, ADEM ) ⇒ H(ADEM )
12. H(ADEM )
13. C(1, ADEM, Symp(Mono))
14. H(ADEM ) ∧ C(1, ADEM, Symp(M ono)) → H(Symp(M ono))
15. H(Symp(M ono))

Smith can also build an additional argument for ADEM, this time based on
the evidence of the symptoms.

1. EV (Symp(M ono))
2. EV (Symp(M ono)) ⇒ H(Symp(M ono))
3. H(Symp(M ono))
4. C(1, ADEM, Symp(M ono))
5. H(Symp(M ono)) ∧ C(1, ADEM, Symp(M ono)) ⇒ H(ADEM ) (instance of
the abduction rule)
6. H(ADEM )

Dr. Safran. Dr. Safran’s objections against the causal link between the TTV
and ADEM can be summarised into two main arguments. Firstly, he argues
against Dr. Smith’s claim that myelin and antigen cell destruction are sufficiently
similar for the vaccine to make an error in distinguishing between these processes.
182 R. Liepina et al.

This argument consists in the single claim

1. ¬Sim(AntigDest, M lnDest)

Secondly, Dr. Safran argues against the conclusions based on the monophasic
symptoms, he claims that the evidence shows that the symptoms observed on
the patient are clearly reoccurring. Therefore, its condition cannot be ADEM ,
since ADEM is described as a monophasic disease in the medical literature.
His argument against ADEM is is the following:

1. EV (C(1, ADEM, Symp(M ono)))

2. EV (C(1, ADEM, Symp(M ono))) ⇒ C(1, ADEM, Symp(M ono))
3. C(1, ADEM, Symp(M ono))
4. ¬H(Symp(M ono))
5. ¬H(Symp(M ono)) ∧ C(1, ADEM, Symp(M ono)) → ¬H(ADEM ) (by
transposition of the rule H(ADEM ) ∧ C(1, ADM, Symp(M ono)) →
H(Symp(M ono))
6. ¬H(ADEM )

Furthermore, Dr. Safran provides an alternative explanation for the illness

stating that recurrent symptoms can be linked with multiple sclerosis, and there-
fore, he defeasibly concludes for multiple sclerosis.

1. C(2, M S, ¬Symp(M ono))

2. H(¬Symp(M ono))
3. H(M S) ∧ C(2, M S, ¬Symp(M ono)) ⇒ H(¬Symp(M ono))
4. H(M S)

Multiple sclerosis indeed provides an explanation of the evidence, as he sees

it, i.e., as including non-monophasic (recurrent) symptoms H(¬Symp(M ono)).

3.4 Preliminary Assessment

Preliminary results show that our semi-formal framework is able to capture the
core points of witness arguments in the causal analysis in the Althen’s case. In
addition to modelling the causal relations and their relative strengths, we also
intend to integrate a level of evidential reasoning from the evidence to a sup-
ported conclusion. This is especially helpful in legal reasoning where the parties
are expected to justify their arguments by showing how the evidence links to their
claims. We also observed that there is a mix of everyday and legal causal expres-
sions with causal language often being accompanied by hedging expressions. We
observed that witnesses used the language of uncertainty, which required our
framework to integrate degrees of belief to portray the nature of discussions
in the court. This suggests a possible development using statistical approaches
to further improve the causal models to reflect causal reasoning present in the
courtrooms.
 Causal Models of Legal Cases 183

Legally relevant causal arguments are often at the heart of legal disputes,
therefore, investigating and modelling these links from various perspectives
enables more thorough analysis of the reasoning involved. Our aim was to cap-
ture how causal links are established in the Vaccine Court assessing expert wit-
ness testimonies. The advantage of our approach is shown by the fact that the
arguments modelled for causal analysis have been scrutinised, and an additional
layer of justification based on defeasible rules has been added. At the moment,
there is a lack of consistency in assessing causal links in courts according to the
guidelines. We believe that with better understanding of the causal links and
their justifications, these guidelines could be improved for greater consistency
and predictability in the courts. What we have observed in this particular case
is the continuous reference to the normative guidelines without fully explaining
what is meant by ‘medical causation’, ‘logical sequence’, etc. Our semi-formal
analysis is intended to explicitly represent causal and evidential reasoning for
these concepts.
While we agree that law is a practical field and does not necessarily have
to engage in a philosophical debate on causation, it is important that such an
essential component of decision making as causation is not left undefined. It is
of especial importance when relying on expert witness testimony where personal
bias can play a role [18].

4 Causation and Burden of Proof

Thus far, we have mainly focused on establishing and modelling the cause-in-fact.
However, the mere link of physical events and effects constitutes only a compo-
nent of the task of attributing legal responsibility and, more generally, legal rea-
soning. The work of a decision maker involves assessing the causal explanations
presented by the parties and experts, then deciding which of the explanations
is the most appropriate according to a complex set of criteria, including the
rules set out by the legislators and previous courts. In this section, we examine
how formal theories can support the choice between alternative causal models,
and we propose extensions that are needed to accommodate their use in legal
reasoning.
Two approaches of interest here are the generalisation theory by Hart and
Honoré [9], and normality theory by Halpern and Pearl [8]. According to the
generalisation theory, one can establish ‘particular causal links as instances of
generalisations’ of the cause and effect [10]. This theory heavily relies on com-
mon sense understanding of causation. It has the advantage of simplifying the
discussion about causation in cases where the choice of the alternative expla-
nations is based on subjective assessment and provides limited rationality, as
it is commonly observed in everyday reasoning. It is also useful in cases where
there is an agreement as to what the generalisation of the causal link is. For
instance, taking a billiard ball, it can be agreed that we can generalise about the
movement and impact of the billiard ball on the table and the effects it might
cause in a particular context. While this approach is on many levels intuitive and
184 R. Liepina et al.

reflects many instances of human reasoning, it has two serious flaws. Firstly, it
emphasises human reasoning about causation as ‘arbitrary and irrational’, which
is not necessarily the case and is not a satisfactory justification in the legal dis-
course [19]. And secondly, there is a very limited set of generalisations people
can agree on, where most of them do not help resolving disagreements. On the
other side of the spectrum, there are approaches that emphasise the rational
and quantitative nature of causal decisions by employing statistical methods of
reasoning with causation [5]. These provide the advantage of showing rational
means of choosing particular causal explanation over another, but as observed
in the Althen’s case, experts and decision makers often lack the tools and infor-
mation to make such quantitative claims.
To some extent, our approach is similar to the generalisation theory, since
we say nothing about how the relation is established, which is the crux of much
philosophical discussion. Yet, our approach is significantly less arbitrary and irra-
tional in that we provide plausible and useful causal reasoning rules; moreover,
in a scientific setting, there are meaningful generalisations that are agreed upon
(or at least disagreed on, which may lead to further empirical testing). While
establishing the causal relation is a significant matter, this is beyond what we
address here, which is to represent and clarify the causal arguments as they are
provided in a legal case.
Based on our observations in the case study and the work of Walker and his
team in the Vaccine Injury Project [2], it can be said that there is a combination
of common sense and formal reasoning based on a diverse pool of evidence. One
of the challenges in the cases, where the Special Masters rely heavily on expert
input to determine cases, is to balance the standards of proof between what is
accepted in various domains involved [17], i.e., medical proof and legal burden
of proof [4,15]. Legal proof in the vaccine cases does not amount to the proof as
considered in science. There are also other considerations that the judges have
to take into account that might not always fit into strictly rational frameworks.
To better reflect legal reasoning about causation, we believe that there is a need
for a norm-based analysis.
Halpern and Pearl [8] have proposed the use of normality, typicality and
default measures in comparing causal explanations. Each of the terms is vaguely
defined to reflect the notions of strength in the ways these can be supported by
evidence. To give a flavour of these, some aspects of normality can be covered
by social norms (descriptive) and legal norms (prescriptive), while typicality can
be supported by statistical evidence or various templates of causal observations
in life. These provide a formal way of comparing the alternative explanations
based on how close such scenarios are to the ‘normal world’. This approach has
the advantage of reflecting how human reasoners build counterfactual scenarios.
The shortcoming, as with the generalisation theory, is the loose definitions of
the terms and criteria of setting norms.
Our idea is to apply these theories of generalisation and normality to a set
of cases to establish a pattern of norms used to prove a causal link for legal
responsibility attribution. Our hypothesis is that by surveying a set of class
 Causal Models of Legal Cases 185

action vaccine cases, we would be able to identify the causal arguments and
reasoning patters that establish the ‘normal world’. Based on such patterns, the
reasoning and justification behind causal links are uncovered, and new argu-
ments can be built allowing that case outcomes might be predicted in the same
restricted domain. Furthermore, we intend to utilise the semi-formal framework
in modelling cause-in-fact arguments in other legal cases.

5 Conclusions and Future Work

The overall aim of this line of research is to develop an approach that identifies,
formally models, and supports reasoning with causation in a manner that is
relevant to the legal domain. In this paper, we have introduced a semi-formal
approach for modelling causal arguments in law based on strict, defeasible and
causal relations. As it can be seen from our preliminary results, we have mainly
focused on simple causal relations that are expressed through complex arguments
in the court. The proposed rules can be reused to model similar cases in the
vaccine court and have the potential to model a large variety of causal issues.
As part of the future work, we intend to use the methods and insights devel-
oped here to tackle more challenging causal issues. Furthermore, we plan to
integrate preference relations and argument assessment in our framework to
accommodate some of the decision making tasks based on causal relations. Such
an endeavour would employ both qualitative and quantitative analytical tools
in determining liability attribution based causal expressions used by the reason-
ers. There is a lot more that needs to be done to understand and automate the
arguments and reasoning based on causal models, but we agree with the other
specialists in the field [12], that it is work worth pursuing to add some of the
missing pieces in the puzzle of legal reasoning and argumentation.

References
1. Althen v Secretary of HHS, The Court of Federal Claims, Golkiewicz, Chief Special
Master, 2003 WL 21439669 (2003)
2. Research Laboratory for Law, Logic & Technology, Vaccine Injury Project. http://
www.lltlab.org/projects/data-projects/vaccineinjury-project/. Accessed 10 Nov
2017
3. Beebee, H., Hitchcock, C., Menzies, P.: The Oxford Handbook of Causation. Oxford
University Press, Oxford (2009)
4. Bex, F., Walton, D.: Burdens and standards of proof for inference to the best
explanation: three case studies. Law Probab. Risk 11(2–3), 113–133 (2012)
5. Chockler, H., Fenton, N., Keppens, K., Lagnado, D.: Causal analysis for attributing
responsibility in legal cases. In: Proceedings of the 15th International Conference
on Artificial Intelligence and Law, pp. 33–42. ACM (2015)
6. Giunchiglia, E., Lee, J., Lifschitz, V., McCain, N., Turner, H.: Nonmonotonic causal
theories. Artif. Intell. 153(1–2), 49–104 (2004)
7. Halpern, J.Y.: Actual Causality. MIT Press, Cambridge (2016)
186 R. Liepina et al.

8. Halpern, J.Y., Hitchcock, C.: Actual causation and the art of modeling. In: Dechter,
R., Geffner, H., Halpern, J. (eds.) Heuristics, Probability, and Causality: A Tribute
to Judea Pearl, pp. 383–406. College Publications, London (2010)
9. Hart, H.L.A., Honoré, T.: Causation in the Law. Oxford University Press, Oxford
(1985)
10. Honoré, T.: Causation in the law. In: Zalta, E.N. (ed.) The Stanford Encyclopedia
of Philosophy. Winter 2010 Edition (2010)
11. Kowalski, R., Sergot, M.: A logic-based calculus of events. In: Schmidt, J.W.,
Thanos, C. (eds.) Foundations of Knowledge Base Management, pp. 23–55.
Springer, Heidelberg (1989). https://doi.org/10.1007/978-3-642-83397-7 2
12. Lehmann, J., Breuker, J., Brouwer, B.: Causation in AI and law. Artif. Intell. Law
12(4), 279–315 (2004)
13. Lehmann, J., Gangemi, A.: An ontology of physical causation as a basis for assess-
ing causation in fact and attributing legal responsibility. Artif. Intell. Law 15(3),
301 (2007)
14. Mueller, E.: Commonsense Reasoning: An Event Calculus Based Approach. Mor-
gan Kaufmann, Burlington (2014)
15. Prakken, H., Sartor, G.: A logical analysis of burdens of proof. In: Kaptein, H.,
Prakken, H., Verheij, B. (eds.) Legal Evidence and Proof: Statistics, Stories, Logic.
Ashgate, Aldershot (2009)
16. Turner, H.: A logic of universal causation. Artif. Intell. 113(1–2), 87–123 (1999)
17. Walker, V., Vazirova, K., Sanford, C.: Annotating patterns of reasoning about
medical theories of causation in vaccine cases: toward a type system for arguments.
In: ArgMining@ ACL, pp. 1–10 (2014)
18. Walton, D.: Argumentation Methods for Artificial Intelligence in Law. Springer,
Heidelberg (2005). https://doi.org/10.1007/3-540-27881-8
19. Wright, R.W.: The NESS account of natural causation: a response to criticisms
(2011)
 Developing Rule-Based Expert System
for People with Disabilities – The Case
of Succession Law

Michał Araszkiewicz1(&) and Maciej Kłodawski2

1
Department of Legal Theory, Faculty of Law and Administration,
Jagiellonian University, Kraków, Poland
michal.araszkiewicz@uj.edu.pl
2
Department of Legal Theory, Faculty of Law and Administration,
University of Zielona Góra, Zielona Góra, Poland
m.klodawski@wpa.uz.zgora.pl

Abstract. This paper presents the features of a moderately simple legal expert
system devoted to solving the most frequent legal problems of disabled persons
in Poland. The authors focused on the structure of legal expert system and
methodology used for the sake of its development. The succession law of Poland
has been selected in the paper as the illustrative domain, because the modelling
of the succession procedures delivers sufficient material to reveal the most
important issues concerning project of the legal expert system.

Keywords: Legal expert system
Legal problems of disabled persons

Rule-based reasoning
Succession law

1 Introduction

This paper presents a selection of results of the project “New applications in the
judiciary against the exclusion of people with disabilities”1. The authors of this paper
were members of an interdisciplinary team encompassing 7 people, among them
lawyers, a doctor of arts, an expert on integration of people with disabilities and a
computer scientist. The aim of the project was to develop a set of documents (inter-
active court forms, contract templates, instructions) as well as informative infographics
and finally a modest legal expert system that would enhance access to justice for people
with disabilities. The project encompasses not only the domains of law that are
specifically interesting for people with disabilities, but also the most common domains
such as succession law or accommodation law. It was assumed that people with dis-
abilities are particularly vulnerable to potential problems related to those basic fields.
The project assumed the use of simplified plain language (rather than formalistic legal
language) as well as attractive graphical form. Its results should be assessed as

1
Financed from the EEA Funds in the frame of the program and realized in years 2015–2016 by
INPRIS – Institute for Law and Society, a NGO concerned, inter alia, with legal policy, new
technologies and access to justice (the website of the organization is accessible at http://www.inpris.
pl/en/home/).

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 187–201, 2018.
https://doi.org/10.1007/978-3-030-00178-0_12
188 M. Araszkiewicz and M. Kłodawski

interesting from the point of view of usage of simple information technology tools in
legal aid systems. It is worth noting that in Poland the system of legal aid is on the early
stage of development.
The present contribution is therefore focused on representation of actual reasoning
of lawyers (in particular, asking proper questions to the clients/users and adequately
reacting to the information provided by them) rather than on application of a more
advanced technology for the sake of generation of valuable output in a procedure that
does not resemble actual legal reasoning. Therefore, this paper is closer to the field of
computational legal theory rather to the domain of applied legal AI [13]. Apart from its
practical value, the expert system provides insights concerning the structure of norms
of the modeled domains, the optimal procedure of information gathering with regard to
the issue in question, as well as to the content of the investigated domain itself. Due to
the technological limitations present in the project from its initial steps to the very end,
and therefore lesser relevance of such issues such as combinatorial explosion or han-
dling inconsistencies (all potential inconsistencies were eliminated manually), the
analysis of computational features is not included here.
The authors of the present paper, as experts in law, were responsible for preparation
of the legal content of the interactive documents, the infographics and for the content
and the logical structure of legal expert system which was then implemented by the
programmer - Michał Szota - in the Angular framework in JavaScript. This paper
focuses on the structure of legal expert system and methodology used in its develop-
ment. The order of investigations is as follows. In Sect. 2 we discuss the characteristics
of the developed legal expert system on the background of the state of art. Section 3
discusses the illustrative domain chosen for the aims of the present paper: the suc-
cession law of Poland. Although the expert system encompasses also other selected
topics (such as eviction and social benefits procedures), during our works we found the
modelling of the succession procedures the most challenging task. Section 4 shows an
example of adoption of the chosen methodology. Section 5 presents the application in
action. Section 6 presents a discussion and directions of further research.

2 The Basic Assumptions of the Legal Expert System

The works on the legal expert system in the project were based on the following
assumptions.
• The developed system is purely a rule-based one. No Case-Based Reasoning,
probabilistic or argumentative components were planned to be included in the
system – therefore it was a challenging task to select domains and problem fields
that may be adequately represented without resort to richer knowledge represen-
tation tools.
• The system is designed to deal with recent fact situations; the content of legal
system was fixed for the date of March the 1st, 2016. The system models concrete
versions of normative acts and does not deal with any intertemporal problems, as
opposed to the system developed for instance by Yoshino [8].
 Developing Rule-Based Expert System for People with Disabilities 189

• The functioning of the system is typical for classical legal expert systems: first
defining the issue to be decided and then determining an answer to the issue by
providing answers by the user to the questions asked by the system.
• The questions asked by the system should be either “yes or no” questions (where an
answer is dictated by application of a well-defined concept), or a question about
quantities of certain objects.
• The system assumes negation as failure: if it is not known to the user that a state of
affairs S holds, it is concluded that S does not hold.
• The terms used in “yes or no” questions should be extensionally unequivocal, i.e. it
should be obvious even for a lay person whether the state of affairs designated by
the term holds in the world or not. In case of any interpretative problem the user of
the system is referred to consult a lawyer.
Although it is clear that adoption of the abovementioned assumptions has to lead to
development of a product of rather limited application, they enabled us to focus
exclusively on modelling of legal information representable by rules. As Bench-Capon
rightly notes [3], rule-based legal expert systems give the user the information what
questions should be asked in order to solve a legal problem, rather than an actual
solution to such concrete problem. It is trivial to note that much of the work in the field
of AI and Law has been developed since the 80s precisely to enable the legal reasoning
models to deal with issues of vagueness, open texture and context-sensitivity, to
mention classifying legal problem into simple and hard cases [4], the use of CBR
structures in argumentation [1, 2], joining rules with CBR in hybrid systems [6] and
recently the use of argumentation schemes theory to legal interpretation [7]. As the
present expert system was assumed not to make use of any of these components, our
theoretical aim was to re-explore the expressiveness of rule-based expert systems and
methods to efficiently develop them from the raw legal text. Taking into account the
results of the famous modelling of British Nationality Act [5] we were also interested in
identification of potential gaps in the system of law. Importantly, due to the temporal
and financial limitations, it was not foreseen in the project to develop a domain-
independent shell of the system or its (onto)logical architecture. Therefore, the
developed and implemented application is able only to represent the knowledge con-
tained therein, is not able to learn and each extension of its knowledge base requires
manual work. However, on the other hand, the application correctly resolves the legal
problems which may be posed by the users.

3 Description of the Domain

Although the legal expert system consists of 6 components (applications), here, due to
limitations of space, we choose the domain of succession law as illustrative material,
therefore limiting our examples to the functioning of two applications: one modelling
statutory succession and another one dealing with testamentary succession, both under
Polish law.
190 M. Araszkiewicz and M. Kłodawski

In Polish law statutory provisions related to succession law are gathered mostly in
the Civil Code2 (hereinafter “PCC”), enacted in 1964. The legislator constantly keeps
classic division into testamentary succession and statutory succession, derived from
Roman law. In the structure of the PCC these characteristic for European private law’s
types of succession are separated only partially, what means that they are connected not
only by many general inheritance provisions and rules common both for testamentary
and statutory succession (e.g. unworthiness regulated in Art. 928 of the PCC, which
means that an heir may be adjudged unworthy by a court of law if, for instance, he has
intentionally committed a serious crime against the decedent), but also by tangled, at
least prima facie, rules (e.g. rule regulated in Art. 967 § 1 of the PCC, which states that
if a person entitled to be the testamentary heir does not want to or may not be an heir, a
statutory heir who the share allocated to this person fell to shall be obliged to perform
ordinary legacies, instructions and other dispositions of the decedent which encumber
this person, unless the decedent decides otherwise). Other rules, specific strictly for
given type of succession (e.g. related to testamentary succession accrual, regulated in
Art. 965 of the PCC, which is explained extensively in further part of present paper),
refer to some provisions in “go there and back” way, creating sometimes necessity of
iteration in algorithm.
The rule common for both types of succession in Polish law is that, as stated in Art.
924 of the PCC, the inheritance shall be opened upon the death of the decedent. In
statutory succession that means no statements are needed to make someone an heir.
Almost everyone (exception concerns unworthy persons), who simply is alive and
belongs to group of relatives entitled by law to inherit, may acquire the estate.
The crucial difference between statutory and testamentary succession under Polish
law is dependence on action, at least one, of the decedent. Testamentary succession
requires the declaration of intent of the decedent. Although testamentary succession
may seem more complex than statutory succession in Polish law, this finding is illu-
sory. Statutory succession becomes intricate for instance when family of the decedent
contains numerous relatives of the decedent. Then amount of calculations, fixed shares
in the estate and relations of rules determining the shares of heirs in the estate – some of
which may be ascendants, descendants or siblings – arise. Also factors from beyond the
law, for example fact, that some of descendants are not alive, must be taken into
account.

4 Methodology

The development of any of the components of the legal expert system consisted of the
following steps.
1. Definition of the legal issue to be decided by the system and its sub-issues.
2. Identification of statutory provisions relevant for the modelling.

2
Journal of Laws from 2016, item 380, with further amendments (http://dziennikustaw.gov.pl/DU/
2016/380/1).
 Developing Rule-Based Expert System for People with Disabilities 191

3. Initial transformation of the statutory provisions into rules expressed in any lan-
guage rich enough to express natural numbers, separately for each of the sub-issues.
4. Determination of a list of sufficient conditions for the negative answer to the legal
issue in question.
5. Determination of a list of the remaining legal issues to be taken into account to
provide a final answer to the legal issue in question.
6. Development of an exhaustive list of “yes or no” or quantitative questions such that:
(a) Providing answers to all questions from the list yields an unequivocal and
legally adequate answer to the legal issue in question;
(b) It is determined first whether any of the sufficient conditions for the negative
answer hold;
(c) The sequence of questions is the shortest one possible;
(d) One question asks for one piece of information (here understood as a simple
proposition with no connectives), unless it is possible to ask a complex question
without significant risk of its misinterpretation by the user.
7. Development of a list of rules dictating the system what it should do in reaction to a
given answer to a particular question, where the options are as follows:
(a) Present an information “please use the algorithm X” if according to the initial
information given by the user he should not use the present application, but
another one;
(b) Present an information “explanation” if the question concerns legal term or
complex factual issue and it is assumed that the user will handle the question
after acquiring and understanding the explanation;
(c) Present an information “please consult a lawyer” if the question uses an open-
textured term or if the degree of complexity of legal issue initially described by
the user is too high;
(d) Present a screen “you do not inherit estate after the deceased person” if any of
the sufficient conditions for such conclusion is satisfied;
(e) Go to the next question in the sequence;
(f) Provide a final answer to a legal issue in question, for instance calculate a share
of a person by application of a certain formula, or simply give the answer, if no
calculations are necessary.
Let us show how this methodology was employed with regard to the legal issue of
statutory succession.
The legal issue is defined as follows: (1) is the user of the system a statutory
successor of the deceased person and (2) if an answer to the question (1) is positive,
what is the share of the user in the estate? Let us note that due to the fact of existence of
different groups of precedence among the set of statutory successors under Polish law,
this legal issue is divided into sub-issues concerning inheritance of persons belonging
to different groups. In the system, we have distinguished the following categories of
potential users of the “Statutory Succession” application: Spouse, Child, Grandchild or
Grand-grandchild (for practical reasons, this category has not been extended further),
Sibling, Siblings’ children or grandchildren, Parent, Grandparent, Children of the
deceased spouse.
192 M. Araszkiewicz and M. Kłodawski

The set of statutory provisions for the modelling was chosen on the basis of internal
systematization of the PCC, where provisions ranging from Art. 931 to the Art.
940 together from a Title II entitled “Statutory Succession”.
In the third step, statutory provisions expressed in natural language were trans-
formed into rules expressed in a simplified first-order language. Let us consider the
following provision of the PCC (the translations to English are taken from the Legalis
system published by C.H. Beck, 2017 and Lex system published by Wolters Kluwer,
2017):
Art. 931 of the PCC
§ 1. The children of the deceased and his spouse shall, by virtue of statutory law, be appointed
to inherit first; they shall inherit in equal parts. However, the part of the spouse cannot be
lesser than one fourth of the entire estate.
§ 2. If a child of the deceased did not survive opening the inheritance, that share of the estate
which would have been his shall pass to his children in equal parts. This provision shall apply
respectively to more distant descendants.

Taking into account the existence of groups of precedence in the Polish law of
statutory succession, one may easily note that the quoted provision comprises two types
of information: (1) it defines the first group of precedence and (2) it provides a formula
for calculation of the share of the successor. It was noted earlier that such recon-
structions are done separately for each sub-issue. For instance, if we are concerned in
developing a set of questions and rules for the calculation of a share of a spouse, the
following rules are reconstructed from the quoted provision.
Sub-issue: SPOUSE
Formula 1. IF [the number n of the inheriting children shares  2] THEN
[spouse’s share = 1/n]
Formula 2. IF [the number n of the inheriting children shares > 2] THEN
[spouse’s share = 1/4]
[inheriting child] = [a child of the deceased person alive during the time of death of the
deceased person] and [not excluded from succession]
[inheriting children share] = [a share of an inheriting child] or [a share of inheriting children of
an inheriting child]

Let us note that the latter definition allows for recursive nesting of further inheriting
children shares in the previously considered inheriting children shares, but it does not
have an effect on calculation of the spouse’s share.
The determination of sufficient conditions for negative answer to the legal sub-issue
in question is relatively simple; apart from the general conditions for exclusion from
succession which are applicable to any successor, there are two specific conditions
excluding the spouse from succession:
• Being in the state of separation with the deceased person during the death of the
latter,
• The deceased person, before the time of death, filed a lawsuit for divorce or sep-
aration against the spouse, based on the fault of the spouse, and this lawsuit was
justified.
A circumstance modifying the formulas for calculation of the spouse’s share is the
absence of inheriting children shares. In such situation, the spouse inherits in
 Developing Rule-Based Expert System for People with Disabilities 193

concurrence with parents of the deceased. We will provide the formulas for calculation
of the spouse’s share below together with the set of questions.
The sequence of questions, with rules determining the reactions of the system to the
answers given by the user, is as follows. It should be noted that the following pre-
sentation is not the code of the application (which was developed in Java) but the
representation of the sequence of questions which encompasses both substantial rules
(on the merit of law) and procedural steps (such as “go to” instructions). The proce-
dural steps are introduced to give the Reader the sense of sequence of questions. The
sequence is important because the sufficient conditions of negative answer to the legal
question are investigated. The procedural steps are marked by capital letters. Let us
stress that this exposition serves as faithful representation of actual steps that are taken
by a competent lawyer dealing with this domain (where in reality of course procedural
steps are made implicitly) and it does not represent the actual code of the Java program.
1. Is it the case that in the time of death of the deceased person you were in separation
with that person? Y/N
a. IF Y, THEN END.
b. IF N, THEN GO TO 2.
2. Is it the case that before death, the deceased person filed a lawsuit against you for
divorce or separation based on your fault? Y/N
a. IF Y, THEN END; PLEASE CONSULT A LAWYER (reason: the said lawsuit
must be justified, which is a context-sensitive term).
b. IF N, THEN GO TO 3.
3. Did the deceased person have children who were alive during the time of death of
the deceased person, or who had children who were alive during the time of death of
the deceased person?
a. IF Y, THEN PROVIDE A NUMBER. APPLY Formula 1 or Formula 2.
b. IF N, THEN GO TO 4.
4. Is it the case that both parents of the deceased person were alive during the deceased
person’s death? Y/N
a. IF Y, THEN spouse’s share = ½.
b. IF N, THEN GO TO 5.
5. Is it the case that one of the parents of the deceased person was alive during the
deceased person’s death? Y/N
a. IF Y, THEN spouse’s share = ½.
b. IF N, THEN spouse’s share = 1.
Let us note that in the sub-issue described above, only one legal problem was
classified as requiring a consultation with a lawyer.
Let us now present adoption of the methodology with regard to an issue of testa-
mentary succession: accrual. On the contrary to the legal issue of statutory succession,
the legal issue of testamentary succession is defined as follows:
(1) Is the testament valid?
(2) Has the testament been revoked?
(3) Are there any other persons specified in the testament, except for the user of the
system?
194 M. Araszkiewicz and M. Kłodawski

(4) Is the user of the system a person appointed as a heir in the testament?
(5) If an answer to the question (4) is positive, what is the share of the user in the
estate?
It must be noted that enlarged, in comparison to statutory succession, group of issues is
a consequence of fact that the user is entitled to inherit not directly by law, but by the
will of the decedent expressed in the testament. Therefore an examination which
concerns both legal (validity) and factual (revocation) state of the testament must be a
part of algorithmic way to solution.
Moreover, the will might concern different issues, for instance (I) a role of each
person specified in the testament by the decedent, (II) amount of persons specified in
the testament, (III) a part of the estate being object of the testament. The issue (I) is
most complex as it comprises 5 different types of the decedent’s will among the
distinguished in testamentary succession under Polish law:
(i) to make specified person a heir, namely to appoint to the estate – Art. 959 of the
PCC and the following in Section II of Title III in Book IV of the PCC,
(ii) to make a specified person a legatee (ordinary legacy), namely to oblige a
statutory or testamentary heir to render specific property-related performance for
the benefit of the specified person – Art. 968 § 1 of the PCC,
(iii) to make a specified person a sublegatee (sublegacy), namely to encumber a
legatee with the ordinary legacy – Art. 968 § 2 of the PCC,
(iv) to make a specified person a bequeathed (specific bequest), namely to decide in a
testament drawn up in the form of a notarial deed that a specified person shall
acquire the object of the bequest upon the opening of the inheritance – Art. 9811
§ 1 of the PCC,
(v) to encumber specified person with the instruction (instruction), namely to
impose on an heir or a legatee the obligation of specific acting or refraining from
acting without making anyone a creditor – Art. 982 of the PCC.
Due to assumed limitations of the user we decided to project algorithm as part of the
expert system which can solve only situation of heirs (i). In remaining cases the user of
the system is referred to consult a lawyer, because thorough interpretation of the
testament and distinguishing of types of the decedent’s will are required. We also
decided to exclude the (III) question from algorithm, as the assumed aim of algorithm
of testamentary succession, similarly to algorithm of statutory succession, is calculate
the share of the user in the estate and the answer to the (III) does not have an impact on
this answer in any manner. Furthermore it must be mentioned that questions (3) and
(4) are followed in algorithm by series of questions related to specific testamentary
succession institutions like substitution or accrual.
The set of statutory provisions for the modelling was chosen on the basis of internal
systematization of the PCC, where provisions ranging from Art. 941 to the Art.
967 together from Section I (“Testament”) and Section II (“Appointment of Heir”),
both being a part of Title III entitled “Disposition in Case of Death”.
As regards the third step, the user, after selecting “Testamentary Succession”,
always follows – on the contrary to statutory succession – one path which obviously
may lead him to different solutions.
 Developing Rule-Based Expert System for People with Disabilities 195

Let us present the sequence of questions, with rules determining the reactions of the
system to the answers given by the user, in testamentary succession. It must be noted
that algorithm starts with questions about unworthiness, subsequently is followed by
questions (1) and (2) described above and only successful result of those steps leads to
the questions presented below. Also essential are two hereunder provisions:
Art. 960 of the PCC
If the decedent appointed several heirs to the estate or to a specified part of the estate without
determining their shares in the estate, they shall inherit in equal parts

Art. 965 of the PCC

If the decedent has appointed several testamentary heirs and one of them does not want to or
may not be an heir, the share allocated to him shall fall to the remaining testamentary heirs in
proportion to the shares falling to them (accrual), unless the decedent decides otherwise.

The sequence of questions, with rules determining the reactions of the system to the
answers given by the user, is as follows. It should be noted again that the following
presentation is not the code of the application, but a representation of a set of questions,
encompassing both substantial rules and procedural steps.
1. Are there any other persons specified in the testament, except for the user of the
system? Y/N
a. IF Y, THEN GO TO 1A;
b. IF N, THEN GO TO 1B.
(explanation: step 1B. and the following B. steps are not related with the accrual,
so they are omitted)
1A. Are all of persons specified in the testament heirs? Y/N
a. IF Y, THEN GO TO 2A;
b. IF N, THEN END; PLEASE CONSULT A LAWYER
(reason: ordinary legacy, sublegacy, specific bequest and instruction, are difficult
institutions of testamentary succession, so thorough legal interpretation of the
testament is required)
2A. Input a total amount of heirs in the testament (including the user) NUMBER(H)
(explanation: number must be > 1, which is logical consequence of decision made
in step 1.)
a. IF NUMBER(H) = 1, THEN WAIT FOR NUMBER > 1;
b. IF NUMBER(H) > 1, THEN GO TO 3A.
3A. Did the decedent specify in the testament the user’s share in the estate? Y/N
a. IF Y, THEN INPUT USER’S SHARE [S(1)] AND GO TO 4A;
b. IF N, THEN LEAVE UNKNOWN [U(1)] AND GO TO 4A.
(explanation: algorithm memorises [S(1)] and [U(1)] values)
4A. Did the decedent specify in the testament the other heir’s share in the estate? Y/N
(explanation: the question 4A. iterates H − 1 times and algorithms retains to
memory each obtained value)
a. IF Y, THEN INPUT HEIR’S SHARE [S(2)] AND GO TO 5A.
b. IF N, THEN LEAVE UNKNOWN [U(2)] AND GO TO 5A.
(explanation: algorithm memorises [S(2)] and [U(2)] values)
196 M. Araszkiewicz and M. Kłodawski

5A. Does the other heir specified by the decedent want to or may become an heir? Y/N
(information: the user obtains a visually assisted explanation presented on the
additional screen)
a. IF Y, THEN
i. IF Y IN 3A. AND IF Y OR N IN 4A., THEN user’s share = [S(1)]
ii. IF N IN 3A. AND IF Y IN 4A., THEN user’s share = 1 − [S(2)]
iii. IF N IN 3A. AND IF N IN 4A., THEN user’s share = 1/H (in presented
example: ½)
(explanation: according to Art. 960 of the PCC; steps i., ii. and iii. are not
related with the accrual institution, so we present them only as a sample
calculation)
b. IF N, THEN GO TO 6A.
6A. Did the decedent specify whom fall share in the estate allocated to the heir, who
does not want to or may not be an heir? Y/N
a. IF Y, THEN GO TO 1C.
(explanation: step 1C. and the following C. steps. are not related with the
accrual, so they are omitted)
b. IF N, THEN GO TO 7A.
7A. Has the decedent excluded accrual towards the heir who does not want to or may
not be an heir?
(information: the user obtains a visually assisted explanation presented on the
additional screen)
a. IF Y, THEN
i. IF Y IN 3A. AND IF Y OR N IN 4A., user’s share = [S(1)]
ii. IF N IN 3A. AND N IN 4A., THEN user’s share = 1/H of the estate (in
presented example: ½)
(explanation: according to Art. 960 of the PCC)
iii. IF N IN 3A. AND Y IN 4A., THEN user’s share = 1 − [S(2)]
b. IF N, THEN
i. IF Y IN 3A. AND Y IN 4A., THEN user’s share = [S(1)] + [S(1)] * [S
(2)]
ii. IF Y IN 3A. AND N IN 4A., THEN user’s share = [S(1)] + [S(1)] *
{1 − [U(2)]}
iii. IF N IN 3A. AND Y IN 4A., THEN user’s share = {1 − [U
(1)]} + {1 − [U(1)]} * [S(2)]
iv. IF N IN 3A. AND N IN 4A., THEN user’s share = 1 − [U
(1)]} + {1 − [U(1)]} * {1 − [U(2)]}
It must be noted that when [S(1)] + [S(…)] + [S(H)] > 1, algorithm informs the
user, that sum of all shares (expressed as fractions) must be no greater than 1 (as 1
represents the maximum possible value to be divided in the testament), and if the user
is certain about [S(1)] + [S(…)] + [S(H)] > 1, he is advised by algorithm to consult a
lawyer. The similar situation occurs in opposite case, i.e. [S(1)] + [S(…)] + [S(H)] < 1
– (1) the decedent probably made a mistake in determining the shares of the various
heirs or (2) the decedent intentionally described the shares, which do not cover the
 Developing Rule-Based Expert System for People with Disabilities 197

entire estate. In both cases exists a discrepancy between the amount of shares and the
size of the entire estate and the user is advised by algorithm to consult a lawyer.

5 Expert System in Action

In this section we present selected screens captured from the Javascript interactive
application available at http://www.inpris.pl/infografika/2016/index.php. The entire
application is accessible only in Polish language version, so we shortly summarize
content of each screen below each Figs. 1, 2, 3, 4, 5, 6 and 7. All graphics being part of

Fig. 1. Initial screen of the interactive application (from the left, upper row “debts, litigation and
its costs”, “inheritance and estates”, “accommodation”; from the left, lower row “consumer
rights”, “social benefits”, “employment”).

Fig. 2. Sub-menu “inheritance and estates”. “Statutory succession” (upper centre) and
“testamentary succession” (upper right) algorithms are separated. Also present (lower row) are
“legitim” algorithm (on the left) and “models documents in the polish succession law”.
198 M. Araszkiewicz and M. Kłodawski

Fig. 3. Step 1A (question “Are all of persons specified in the testament heirs?”, possible
answers “yes” or “no”) of the accrual case presented in Sect. 4. blue button leads to
“informations”. (Color figure online)

Fig. 4. Explanation available after clicking on blue button presented in Fig. 3. This screen
explains differences between heirs, legatees and instructed persons. (Color figure online)

the application have been designed by prof. Justyna Lauer from Academy of Fine Arts
in Katowice (Akademia Sztuk Pięknych w Katowicach). We asked for and were
granted permission from prof. Lauer, as well from programmer Michał Szota, to the
presentation of screen captures of the application in this paper. The graphics combine
text and pictures and aim to enhance the understanding of the presented content. The
language used in the application is a simplified version of statutory language.
 Developing Rule-Based Expert System for People with Disabilities 199

Fig. 5. Step 2A of the accrual case presented in Sect. 4. The user must input amount of all heirs,
including the user.

Fig. 6. Steps 3A–7A of the accrual case (questions aforementioned in Sect. 4).

Fig. 7. Solution of the case presented in Sect. 4 (a sample consistent with variant 7A. b. ii.).
200 M. Araszkiewicz and M. Kłodawski

6 Conclusions

The project described above shows, in our opinion, that classical rule-based expert
systems may still be assessed as useful tools in enhancing access to justice, even
though their expressive power is limited and even though they are not backed by a set
of (onto)logical assumptions.
The developed system enables us to rise questions concerning the benefits stem-
ming from similar simple projects. In our opinion, such implementations are valuable
even though their theoretical import is limited. We may identify at least three main
advantages of similar projects. The first advantage is quite obvious: the application
enables a broader circle of people to access the content of legal rules which are
important for they in their day-to-day activities. Even though similar applications, for
obvious reasons, cannot provide for all the questions the user might be interested to
ask, they still provide answers to many questions and they rise the user’s awareness of
problems which require an advice from a professional lawyer. Therefore, similar
systems, development of which is not very expensive, may play an important role in
legal aid systems. Second, adoption of such methodological attitude reduces many of
the problems with implementation of the content of application, for the programmer is
not constrained by a large number of theoretical assumptions (concerning, for instance,
the assumed ontology, definitions of objects etc.). This feature is an advantage only if
an application is not large and if it is not primarily developed to encompass all future
amendments to the law. The application is, therefore, evaluated on the basis of sub-
stantial rightness of the provided answers and preservation of correct sequence of
questions. This evaluation has been done manually by a group of lawyers. Third, the
development of the system in question provides some important theoretical insights
concerning complexity of the regulation in question, enabling us to list and count the
questions for which an unequivocal answer may be yielded (easy questions), and
questions which involve an advice from a lawyer (hard questions). Let us recall that in
case of the statutory inheritance of a spouse, 4 (out of 5) questions were assessed as
easy ones. Similar result has been obtained for the determination of a user’s share in
case of accrual, however, this involved putting more information into the set of pro-
cedural rules.
The presented examples of modules of the system show that the modest set of tools
is sufficient to model quite complex issues, provided that they involve well-defined
concepts or numbers. In the future we are intending to develop the system further, to
include a CBR component and intertemporal issues as well as a module related to the
problem of legal interpretation, including doctrinal interpretation of the concepts
contained in the application. We are particularly concerned with the latter issue, for
enhancement of rule-based systems with a base of cases has been a known approach
from the early days [6]. The introduction of an argumentative knowledge concerning
legal interpretation may lead to a more fine-grained elaboration of hard questions.
A rule-based implementation of the argumentative schemes used for statutory inter-
pretation [7, 9–12] should enhance the system’s usefulness not only for its primary
users, but for professional lawyers, too. This component of the system would prepare
the user for a meeting with a lawyer by providing an information what kind of
 Developing Rule-Based Expert System for People with Disabilities 201

information (for instance, on conflicting interpretations) should be expected with

respect to a given question.
The scope of the project did not encompass systematized evaluation of the appli-
cation by potential users. The functionalities of the system were assessed and validated
by professional lawyers. It is intended that such evaluation is performed in the later
stages of the project.

References
1. Aleven, V.: Teaching case-based argumentation through a model and examples. Ph.D.
Dissertation, University of Pittsburgh, Graduate Program in Intelligent Systems (1997)
2. Ashley, K.: Modeling Legal Argument: Reasoning with Cases and Hypotheticals. MIT
Press, Cambridge (1990)
3. Bench-Capon, T.: What makes a system a legal expert? In: Schäfer, B. (ed.) Twenty-Fifth
Annual Conference Legal Knowledge and Information Systems, JURIX 2012, pp. 11–20.
IOS Press, Amsterdam (2012)
4. Gardner, A.L.: An Artificial Intelligence Approach to Legal Reasoning. MIT Press,
Cambridge (1987)
5. Sergot, M., Sadri, F., Kowalski, R., Kriwaczek, F., Hammond, P., Cory, H.T.: The British
nationality act as a logic program. Commun. ACM 29, 370–386 (1986)
6. Skalak, D., Rissland, E.: Arguments and cases: an inevitable intertwining. Artif. Intell. Law
1, 3–44 (1992)
7. Walton, D., Sartor, G., Macagno, F.: Contested cases of statutory interpretation. Artif. Intell.
Law 24, 51–91 (2016)
8. Yoshino, H., et al.: Legal expert system—LES-2. In: Wada, E. (ed.) LP 1986. LNCS, vol.
264, pp. 34–45. Springer, Heidelberg (1987). https://doi.org/10.1007/3-540-18024-9_20
9. Żurek, T., Araszkiewicz, M.: Modelling teleological interpretation. In: Verheij, B.,
Francesconi, E., Gardner, A.L. (eds.) Proceedings of the Fourteenth Conference on Artificial
Intelligence and Law, ICAIL 2013, pp. 160–168. ACM, New York (2013)
10. Sartor, G., Walton, D., Macagno, F., Rotolo, A.: Argumentation schemes for statutory
interpretation: a logical analysis. In: Hoekstra, R. (ed.) Twenty-Seventh Annual Conference
Legal Knowledge and Information Systems, JURIX 2014, pp. 11–20. IOS Press, Amsterdam
(2014)
11. Macagno, F., Sartor, G., Walton, D.: Argumentation schemes for statutory interpretation. In:
Araszkiewicz, M., Myška, M., Smejkalová, T., Šavelka, J., Škop, M. (eds.) International
Conference on Alternative Methods of Argumentation in Law, ARGUMENTATION 2012,
pp. 31–44. Masaryk University, Brno (2012)
12. Araszkiewicz, M.: Towards systematic research on statutory interpretation in AI and law. In:
Ashley, K. (ed.) Twenty-Sixth Annual Conference Legal Knowledge and Information
Systems, JURIX 2013, pp. 15–24. IOS Press, Amsterdam (2013)
13. Schäfer, B.: Formal models of statutory interpretation in multilingual legal systems. Statut.
Law Rev. 38(3), 310–328 (2017)
Legal Vocabularies and Natural
Language Processing
 EuroVoc-Based Summarization
of European Case Law

Florian Schmedding1(B) , Peter Klügl1 , David Baehrens1 , Christian Simon2 ,

Kai Simon3 , and Katrin Tomanek4
1
Averbis GmbH, Tennenbacher Str. 11, 79106 Freiburg, Germany
{florian.schmedding,peter.kluegl,david.baehrens}@averbis.com
2
INTER CHALET Ferienhaus-Gesellschaft mbH,
Heinrich-von-Stephan-Str. 25, 79100 Freiburg, Germany
c.simon@interchalet.com
3
European Patent Office, Rijswijk, The Netherlands
4
VigLink Inc., 333 Bush Street, San Francisco, CA 94110, USA
katrin.tomanek@gmx.de

Abstract. This work reports on the ongoing development of a multi-

lingual pipeline for the summarization of European case law. We apply
the TextRank algorithm on concepts of the EuroVoc thesaurus in order
to extract summarizing keywords and sentences. In a first case study, we
demonstrate the feasibility and usefulness of the presented approach for
five different languages and 18 document sources.

1 Introduction

The ongoing harmonization of the various legal systems across the European
Union constantly increases the importance of the pan-European legal system.
For example, national case law decisions from the higher courts become imme-
diately effective in the legal system of each member state. Consequently, legal
professionals must consider more and more all member state’s national legisla-
tion and jurisdiction when researching the context of their causes.
However, such researches are difficult due to the intrinsic diversity of the
European Union. Legal information is spread over several national and Euro-
pean repositories and is, of course, published in different languages. Additionally,
the format and the structure of legal documents depend heavily on their origin
and often there are no standardized cross-references to the related documents.
Accordingly, there is a growing need for information systems that enhance the
access to the pan-European legislation and jurisdiction. A thorough overview on
the corresponding backgrounds is elaborated by Boella et al. [3], a more techni-
cal view on the characteristics of the legal domain is presented by the authors of
the Eunomos system [2]. Beside text classification and document clustering (see

C. Simon, K. Simon and K. Tomanek—The contributions were developed while this

author was working at Averbis GmbH.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 205–219, 2018.
https://doi.org/10.1007/978-3-030-00178-0_13
206 F. Schmedding et al.

Boella et al. [4] and Schweighofer et al. [31], for example), text summarization
is an important feature of these systems.
Being confronted with large bodies of documents that must be scrutinized
with great precision, legal professionals greatly benefit from document sum-
maries [24]. In turn, with respect to the size of the legal repositories, automated
document summarization greatly simplifies the operation of legal information
systems. Regarding the diverse origins of the legal documents, the system must
process different formats and languages uniformly.
In this work, we report on our experiences and ongoing work on both aspects.
We approach the first by generating two kinds of summaries, controlled keyword
and sentence summaries. The keywords are based on the EuroVoc terminol-
ogy and hence not only descriptive but do also relate documents across lan-
guages. For both summaries we pursue the language-independent graph-based
TextRank algorithm [23]. However, it benefits from a language-specific prepro-
cessing. Regarding the second, we are currently dealing with five important
European languages: Bulgarian, English, French, German, and Italian. Our nat-
ural language processing is based on UIMA [13], a framework for the orchestra-
tion of individual text analysis engines, facilitating the clear separation between
language-dependent and language-independent parts of the pipeline. Thus, the
extension to another language is fairly easy—only the language-specific compo-
nents have to be added.
We evaluated our results with questionnaires send to legal experts asking
them to rate the automatically selected keywords and sentences. Their answers
show the feasibility of our approach. Additionally, we invited them to select the
keywords and sentences that summarize the document in their opinion. Together
with the previous answers we hope to gain insights and clues for the improvement
of our system and its further adjustment to the legal domain.
The rest of the paper is structured as follows: Sect. 2 gives a short overview
of the related work on text summarization of legal documents and on approaches
based on the UIMA framework. The multilingual pipeline is introduced in Sect. 3
and the text summarization of European case law is described in Sect. 4. Section 5
presents the results of the case study and Sect. 6 concludes with a summary.

2 Related Work
Text summarization is an active and thriving field of research due to the fact that
its importance grows with the increasing amount of available textual information.
An overview of this research area, its historical development and categorization of
approaches is given in various survey articles (cf. Gupta and Lehal [17], Nenkova
and McKeown [25], Dalal and Malik [7], and Elfayoumy and Thoppil [9]). Espe-
cially Alemany et al. [1] provide a comprehensive overview of earlier systems.
Another recent summary focused on multilingual approaches for summarization
can be found in Sarkar [30]. Most publications classify approaches for text sum-
marization into diverse categories, for example, extractive and abstractive, single
document and multi document, language dependent and language independent,
 EuroVoc-Based Summarization of European Case Law 207

and other categories based on the purpose or target users (cf. Sarkar [30]). Fol-
lowing this, the approach pursued in this work can be classified as extractive,
single document and language independent.
The legal domain is predestined for text summarization and attracts a con-
siderable amount of interest. In the following, we present a selection of recent
and representative approaches for text summarization in the legal domain (in
chronological order). An extended overview of previous work can be found in
Moens [24].
The LetSum system [12] extracts summarizing sentences dependent on the
thematic structure of juridical decision. After preprocessing for tokenization,
sentence splitting and part-of-speech tagging is applied, the separate segments
concerning decision data, introduction, context, juridical analysis and conclusion
are identified. Their content is filtered by removing passages like citations and
large paragraphs. Then, separate heuristics optimized for a specific segment type
are applied in order to extract best candidates, which are proportionally joined
for the final summary.
The work of Saravanan et al. [29] and Hachey and Grover [18] follows up on
the idea of improving legal text summarization using segmentation in rhetori-
cal roles. Saravanan and Ravindran [29] apply Conditional Random Fields with
adapted feature sets to identify segments of seven rhetorical roles. In these seg-
ments, key sentences are extracted using term distribution models. Hachey and
Grover [18] define the identification of roles as well as the extraction of sum-
marizing sentences as a classification task and compare different machine learn-
ing models like Naive Bayes, Support Vector Machines and Maximum Entropy.
Yousfi-Monod et al. [34] also learn supervised Naive Bayes models for summa-
rization of legal documents. They evaluated different combinations of surface,
emphasis and content features for English and French documents concerning
immigration, tax and intellectual property. In Chieze et al. [6], an automatic
summarization system for English and French legal documents is described.
Kim et al. [19] developed an extractive summarization algorithm based on
directed and disconnected graphs with asymmetric edge weights. Furthermore,
they do not need to specify the compression rate and are able to provide summa-
rizing sentences with a high cohesion. Galgani et al. [14] argue that approaches
for summarization that are based on a single technique are not able to provide
sufficient results in divergent domains. They propose to build summarizers cus-
tomized for specific domains by a rule-based combination of different elements
like frequencies, centrality, citations and linguistic information. The increased
engineering effort for this approach is faced with an incremental knowledge
acquisition framework. While most publications automatically evaluate their
approaches using Rouge [22], they additionally presented results using human
evaluators. In another approach, Galgani et al. [15] emphasized the usefulness
of citation graphs in legal documents for text summarization. They combined
elements of cited and citing documents and validated that best performing com-
binations depend on the domain.
208 F. Schmedding et al.

In contrast to this previous work on text summarization in the legal domain,

we do not target the development of novel techniques. This work utilizes an
established technique for language independent summarization and investigates
architectural requirements, feasibility and results for multilingual case law. The
experimental results are reported using human evaluators only. From this archi-
tectural perspective, an additional branch of related work concerning UIMA can
be identified. DKPro Keyphrases [10] is a UIMA-based framework for keyphrase
extraction experiments and thus facilitates the evaluation as well as the develop-
ment of new approaches. The ATLAS platform [26] provides a framework based
on UIMA for multilingual natural language processing supporting several Euro-
pean languages. The authors report that the planned summarization component
will be based on LexRank [11] and Open Text Summarizer1 . A demonstrator
of the platform covers legal documents of the European Union that are cate-
gorized using the EuroVoc thesaurus. Multilingual natural language processing
is in general a well understood and supported use case in UIMA, which can be
ascertained by diverse publications [20,21,28,32].

3 Multi-lingual Document Processing

Clearly, a legal information system in a pan-European setting is confronted with

different languages (and distinct scripts—Latin and Cyrillic). Therefore, when
it comes to natural language processing, this should be done in a flexible and
robust way, i. e., the language coverage should be easily extensible. However,
there must be a common ground such that new components can be included free
of interference.
The UIMA framework [13] has proven to provide the necessary means to
these requirements paying special attention to the interoperability of compo-
nents. Individual components—the analysis engines—are orchestrated in the
processing pipeline and communicate with each other through adding or modi-
fying meta information stored in the Common Analysis Structure (CAS) which
contains the currently processed document. This information is represented by
typed feature structures based on a common type system and thus suited for
robust applications with predictable behavior and results.
Additionally, the UIMA framework comes with a powerful built-in support
for horizontal scaling achieved by distributing the requests via an ActiveMQ2
message broker to available processing endpoints. This is not only handy when
facing vast amounts of data but also helpful if some pipeline components are not
thread-safe. In this case, the workload can be easily parallelized by registering
multiple instances of the same pipeline at the broker which then takes care of
spreading the requests to a free instance, completely transparent to the user.
Our pipeline is depicted in Fig. 1. The reminder of this section introduces its
language-dependent components and the EuroVoc concept mapper.

1
http://libots.sourceforge.net/.
2
http://activemq.apache.org/.
 EuroVoc-Based Summarization of European Case Law 209

Fig. 1. Overview of the multilingual pipeline for text summarization.

3.1 NLP Components

Sentence detection segments the documents into sentences, which is an impor-
tant task for extractive text summarization. While this is a well-understood
problem, it can provide challenges if the sentence splitter also needs to handle
headlines and listings correctly. We adapted and evaluated rule-based approaches
and probabilistic models based on Maximum Entropy for case laws in different
languages with the result that the rule-based components prevailed. The tok-
enizer creates meaningful elements of text (tokens) for further processing and the
stopword tagger identifies tokens that carry no semantics for this given domain,
most often tokens of closed word classes like determiners. The stemmer, lemma-
tizer and morpho-semantic segmenter [8] are applied to reduce words to their
stem, basic form or morphologically meaningful units, respectively. Finally, the
part-of-speech tagger identifies the word class of a token.
The actual configuration of components for a specific language depends on the
availability of the components for that language. Partly, we rely on third-party
components, which influences the selection of previous components. A specific
part-of-speech tagger, for example, may require a certain type of tokenzier. The
language-specific part-of-speech tags are reduced to a universal part-of-speech
tagset [27] for further usage by the language-independent components.

3.2 EuroVoc
EuroVoc3 is a multilingual, multidisciplinary thesaurus created and managed
by the Publications Office of the European Union. It is utilized by European
Parliament and national governments amongst other for indexing. The thesaurus
consists of 21 area fields and 127 microthesauri with translations for over 20
languages. It provides almost 7000 concepts with varying amount of synonyms
per language.
3
http://eurovoc.europa.eu/.
210 F. Schmedding et al.

In the context of this work, we utilize the translation only of the five lan-
guages Bulgarian, English, French, German and Italian. Many concepts are too
general concerning the intended task. The concept “Law”, for example, is hardly
useful for describing the content of case law. Thus, we blacklisted 37 concepts
mostly of the legal domain. We apply an extension of the ConceptMapper [33]
in order to identify the mentions of the concepts and their synonyms in a docu-
ment. The matching process supports different lookup techniques including exact
matches, part-of-speech tags, stems, lemma and morpho-semantic segments. Pos-
sible ambiguities are resolved through removing concepts that have been matched
by an auxiliary synonym in favor of concepts that have been matched by their
preferred synonym. For any ambiguity remaining hereafter each affected concept
is also removed. While this approach may look simple compared to other dis-
ambiguation components of our pipeline (for example, statistics and dictionary
based) it has shown sufficiently good performance for plenty of use cases. So we
achieve a precise, robust and reliable extraction of EuroVoc concepts.

4 Summarization of Case Law

The pipeline introduced in the last section contains two components that are
responsible for generating the essential elements for the text summaries of Euro-
pean case law. While the first one identifies valuable keywords, the second one
extracts sentences which makeup the final summary. Both components are based
on the TextRank algorithm [23]. In this section, we first repeat the basic ideas
of this algorithm and then describe how it is applied to identify summarizing
keywords and sentences.

4.1 TextRank

The TextRank algorithm conceived by Mihalcea and Tarau [23] is an approach to

rank elements of text documents, for example, in order to extract keywords that
outline their topic. Unlike supervised systems, it relies exclusively on information
drawn from the text itself and is principally independent of the document content
and language. However, as Mihalcea and Tarau show, its results are generally
better if natural language processing is applied to the text before.
The basic idea of the algorithm is the transformation of the text into a
graph containing the elements in question as nodes. Edges between the nodes
are introduced according to the relationship between the respective elements and
thus the PageRank algorithm [5] is applicable for assigning a ranking score to
each element.
PageRank is adequate for undirected and directed edges. For cases where
weighted edges are more appropriate for modeling the relationship between the
text elements Mihalcea and Tarau propose an adjusted version of the original
scoring formula. For instance, they use it to compute the most important sen-
tences of a document where the edges between two nodes describe the similarity
among two sentences.
 EuroVoc-Based Summarization of European Case Law 211

While TextRank scores single elements only, a post-processing may be used

to refine the results. For instance, Mihalcea and Tarau join a selected token with
its adjacent ones in order to generate more meaningful keywords.

4.2 EuroVoc Keyword Extraction

In contrast to the pure TextRank algorithm that generates free keywords, we

want to choose the keywords from distinguished candidates contained in a con-
trolled vocabulary. This has two advantages: First, each candidate includes a
list of synonyms that enable the normalization of different tokens onto a precise
keyword. Second, the normalization reveals similarities among documents con-
taining the same concepts but different synonyms or being written in another
language.
Despite using controlled keywords we do not need to specifically adjust the
TextRank algorithm to it. The recognition of concepts from the vocabulary is
performed independently before the keyword extraction takes place. As proposed
by Mihalcea and Tarau, we restrict the domain of the nodes to nouns and adjec-
tives that are identified by the natural language preprocessing. Additionally, we
reject tokens that are shorter than three characters or that contain no letters or
numbers.
According to Mihalcea and Tarau, the TextRank results improve when the
nodes refer to normalized word forms. We use therefore segments (based on mor-
phological analysis) for English and German, lemmas for Italian and Bulgarian,
and stems for French. Each form is the preferred normalization for the corre-
sponding language in our natural language processing pipeline. For creating the
edges we relate each normalized form with the neighboring ones inside a window
size of three. The mapping between nodes and previously determined concepts
occurs in the post-processing step. It is achieved as follows:

1. The score of a concept is computed as the average score of the nodes rep-
resenting it, compensating the number of words within the corresponding
synonym.
2. As each concept may appear at several positions inside the text, equal con-
cepts are grouped into clusters.
3. Each cluster is represented by the concept occurrence having the highest score
among those contained in the cluster.
After sorting the obtained clusters according to the representative score in
descending order, the best concepts appear at the beginning of the resulting
list. Finally, the preferred synonyms pertaining to selected concepts are used to
generate a set of legible keywords.

4.3 Sentence Extraction

Like above for the keyword extraction, we apply the EuroVoc vocabulary for
detecting summary sentences. Distinct from the summarization method used
212 F. Schmedding et al.

Fig. 2. Excerpt of an exemplary questionnaire for rating the text summarization.

by Mihalcea and Tarau, where sentences are the immediate result of the graph
processing (cf. Subsect. 4.1), we identify them based on the keyword extraction.
Having computed the sorted list of clusters, we start with the first and look
at the text position of its representing concept. The enclosing sentence gets
selected for the summary. In order to avoid duplicates, concepts that are already
contained in a previously chosen sentence are skipped. The procedure repeats
with the next cluster until the desired number of sentences is collected or no
more keywords or sentences are available.

5 Case Study
While the pipeline for text summarization is still subject to further improvement,
we prepared an initial case study that provides some insights in the general feasi-
bility and applicability of the presented approach. Furthermore, we hope to gain
clues about possible starting points to improve the running system. Figure 2
shows an excerpt of an exemplary questionnaire for rating the results. As an ini-
tial test data set, we prepared a collection of 100 documents—20 documents from
different courts for each of the languages English, French, Italian, German, and
Bulgarian. The exact amount of documents and their sources are summarized
in Table 1.
To assess the quality of the automatically generated text summaries and key-
words, we conducted a survey of the selection and utility of the document descrip-
tions. Therefore, we asked the available human experts from the legal domain
(native speakers or business fluent) to evaluate summaries and keywords on the
test corpus of documents using predefined questionnaires. They were encouraged
to rate the utility of the automatic descriptions in the categories of “Very useful”,
“Useful”, “Not useful” to “Misleading”. Additionally, they are allowed to select
the five most important summary sentences and keywords from the documents
themselves. This supplementary information is utilized to calculate precision and
recall for the summaries and keywords. These auxiliary scores provide, however,
only limited expressiveness4 since multiple equivalent summaries may exist and
despite a suggested one is rated as useful a different may be chosen by the expert.

4
Therefore, precision and recall are always marked with asterisks.
 EuroVoc-Based Summarization of European Case Law 213

Additionally, the pipeline suggests up to twelve keywords and eight sentences,

and the experts did not always select five keywords or sentences, even if some of
the proposed elements were rated as useful. These circumstances greatly decrease
the evaluation scores but nevertheless they provide helpful insights for further
development.

Table 1. Amount, language and source of the documents utilized in the case study.

L a ngu a ge Source Do cuments Total

EUR-Lex 8
Judiciary.gov.uk 4
English 20
Supreme Court (UK) 4
House of Lords Judgments 2
EUR-Lex 10
French 20
Legifrance 10
Guistizia Amministrativa 10
Italian 20
Corte Costituzionale 10
EUR-Lex 8
Varhoven Kasatsionen Sad 4
Bulgarian 20
Varhoven Administrativen Sad 4
Portal ”Sadebni aktove” - Vish Sadeben Savet 4
EUR-Lex 4
Rechtsinformationssystem (at) 2
Bundesverfassungsgericht 2
Bundesgerichtshof 2
German Bundesverwaltungsgericht 2 20
Bundesfinanzhof 2
Bundesarbeitsgericht 2
Bundessozialgericht 2
Bundespatentgericht 2

5.1 Results

From a total of 100 questionnaires (20 judgments for each of the languages
English, French, German, and Italian, and Bulgarian) we got expert utility rat-
ings for 660 automatically generated summary sentences and 1070 keywords.
This is a response rate of 95.5% and 96.4%, respectively. The participants also
selected 468 sentences and 366 keywords for content description. Thus, whereas
for the sentences almost five were selected per document, less than four were
selected in the keyword case.
Table 2 shows the rated utility by language for the automatic keywords. The
majority of the English and German keywords are judged helpful (>60% useful or
very useful). On the other hand, the keywords in French, Italian, and Bulgarian
failed to describe the contents in roughly 75% of the cases. Additionally, Table 2
contains the aggregated utility rating of the automatic summaries broken down
214 F. Schmedding et al.

by document language. For the languages of English, French, and German more
than 75% have been found useful or even very useful. For Italian, half of the
summaries were rated not useful while for Bulgarian 60% were helpful. The
approximated precision and recall scores range in the anticipated low regions as
explained above. Concerning the keywords, a negative amplitude can be observed
for Bulgarian, and concerning summaries, for Bulgarian and English.

Table 2. Keywords and sentences by language

Keywords Sentences
de en fr it bg de en fr it bg
Very useful 0.07 0.43 0.02 0.03 0.03 0.10 0.38 0.00 0.14 0.05
Useful 0.53 0.20 0.22 0.18 0.27 0.75 0.40 0.78 0.26 0.55
Not useful 0.33 0.30 0.76 0.72 0.62 0.12 0.13 0.22 0.47 0.33
Misleading 0.06 0.00 0.00 0.04 0.01 0.03 0.04 0.00 0.02 0.02
Not rated 0.00 0.07 0.00 0.04 0.08 0.00 0.05 0.00 0.10 0.06
Precision* 0.24 0.23 0.19 0.24 0.16 0.24 0.11 0.25 0.26 0.17
Recall* 0.64 0.63 0.68 0.64 0.55 0.31 0.18 0.27 0.40 0.25

Tables 3 and 4 display the ratings for keywords and sentences broken down to
the different sources. The keywords in only half of the sources are rated helpful
or very helpful at the majority. However, the summaries received more positive
feedback. Only three sources are rated negative at the majority.

5.2 Discussion

The results of the questionnaires indicate that the EuroVoc terminology might
not be useful for extracting summarizing keywords in European case law. Only
the keywords in German and English documents have been rated positive. The
identified keywords can, however, be utilized to create useful summaries, which
can be observed for the positive utility ratings of German and English docu-
ments, but also for French and Bulgarian documents. The extracted sentences
for the two latter languages are rated useful or very useful for the majority, while
their keywords received overall negative feedback. Only Italian documents could
not keep pace.
No dependencies can be observed between the ratings and the precision and
recall scores. The evaluation scores for summarizing keywords are distributed in
the same region regardless of their utility. The summarizing sentences of English
received the most positive feedback by the legal experts, who stated that the
sentences seem to be manually selected. Interestingly, these English documents
obtained the lowest precision and recall scores. This leads to another point of
 EuroVoc-Based Summarization of European Case Law 215

Table 3. Keywords by source

waltungsge-
beitsgericht

stituzionale
fassungsge-
Bundesver-

Bundesver-
tentgericht
Bundespa-
Bundesge-

zialgericht
Bundesso-
Bundesar-

Corte Co-
Bundesfi-

EUR-Lex
richtshof
nanzhof

richt

richt
Very useful 0.21 0.25 0.00 0.05 0.08 0.04 0.05 0.06 0.23
Useful 0.67 0.46 0.62 0.25 0.63 0.58 0.79 0.15 0.19
Not useful 0.13 0.21 0.24 0.60 0.29 0.29 0.00 0.73 0.52
Misleading 0.00 0.08 0.14 0.10 0.00 0.08 0.16 0.01 0.01
Not rated 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.05
Precision* 0.13 0.29 0.10 0.15 0.33 0.38 0.42 0.25 0.20
Recall* 0.30 0.70 0.40 0.60 0.80 0.90 0.80 0.61 0.61

Rechtsinfor-

Administra-
mationssys-
Legifrance

tiven Sad
Varhoven

Varhoven
Kasatsio-
Judiciary
Guistizia
Ammini-

House of

Supreme

nen Sad
Sadebni
strativa

aktove

Court
Lords

tem
Very useful 0.00 0.23 0.40 0.03 0.00 0.00 0.31 0.00 0.00
Useful 0.21 0.27 0.21 0.31 0.29 0.25 0.35 0.25 0.35
Not useful 0.70 0.23 0.31 0.66 0.63 0.75 0.33 0.75 0.65
Misleading 0.08 0.00 0.02 0.00 0.04 0.00 0.00 0.00 0.00
Not rated 0.02 0.27 0.06 0.00 0.04 0.00 0.00 0.00 0.00
Precision* 0.22 0.17 0.15 0.27 0.17 0.15 0.25 0.19 0.19
Recall* 0.69 0.53 0.47 0.77 0.80 0.50 0.60 0.64 0.56

reference that the utility is a more important rating than the evaluation scores
for assessing the summarization. The reason behind may be the aforementioned
existence of several useful summaries. The average number of candidate key-
words and sentences shown in Table 5 supports this presumption. English doc-
uments are generally longer than those of other languages. Consequently, more
appropriate keywords can be extracted and more summarizing sentences exist.
A closer look at the results dependent on the source reveals some interesting
insights. The EUR-Lex documents were processed in all languages but Italian.
Here, 84% of the summarizing sentences haven been rated useful or very useful.
This is a clear indication that the presented approach is capable to extract useful
summaries in the different languages. Most of the summarizing document of
German sources are rated positive, up to 94% and 100% useful or very useful for
“Bundessozialgericht” and “Bundesverfassungsgericht” respectively. Documents
from the Austrian source “Rechtsinformationssystem” received more negative
than positive ratings. More work needs to be invested in order to investigate
the reasons, e.g., the increased amount of not useful keywords. The discrepancy
between the evaluation scores and the utility rating can also be observed for the
sources. The two sources “Bundessozialgericht” and “Bundesverfassungsgericht”
obtained the best feedback, but also completely differing evaluation scores. While
the first one counts to be worst sources with a precision and recall of 0.08 and
0.09, the second one achieved the best precision and recall scores with 0.50 and
0.60 respectively. This is also an indication of a possibly low interrater agreement.
216 F. Schmedding et al.

Table 4. Sentences by source

waltungsge-
beitsgericht

stituzionale
fassungsge-
Bundesver-

Bundesver-
tentgericht
Bundespa-
Bundesge-

zialgericht
Bundesso-
Bundesar-

Corte Co-
Bundesfi-

EUR-Lex
richtshof
nanzhof

richt

richt
Very useful 0.06 0.19 0.15 0.00 0.00 0.17 0.17 0.28 0.30
Useful 0.81 0.75 0.69 0.71 0.94 0.83 0.67 0.24 0.52
Not useful 0.06 0.00 0.15 0.29 0.06 0.00 0.17 0.41 0.12
Misleading 0.06 0.06 0.00 0.00 0.00 0.00 0.00 0.04 0.03
Not rated 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.03
Precision* 0.13 0.19 0.08 0.14 0.06 0.50 0.42 0.28 0.18
Recall* 0.20 0.30 0.10 0.10 0.09 0.60 0.45 0.43 0.25

Rechtsinfor-

Administra-
mationssys-
Legifrance

tiven Sad
Varhoven

Varhoven
Kasatsio-
Judiciary
Guistizia
Ammini-

House of

Supreme

nen Sad
Sadebni
strativa

aktove

Court
Lords

tem
Very useful 0.00 0.06 0.31 0.00 0.00 0.00 0.22 0.00 0.00
Useful 0.29 0.59 0.47 0.78 0.45 0.41 0.41 0.75 0.69
Not useful 0.53 0.06 0.22 0.22 0.55 0.56 0.25 0.25 0.28
Misleading 0.00 0.03 0.00 0.00 0.00 0.00 0.13 0.00 0.00
Not rated 0.17 0.25 0.00 0.00 0.00 0.03 0.00 0.00 0.03
Precision* 0.24 0.03 0.13 0.35 0.27 0.16 0.13 0.22 0.13
Recall* 0.36 0.06 0.20 0.35 0.30 0.25 0.19 0.33 0.19

Table 5. Average number of candidate keywords and sentences per language

de en fr it bg de en fr it bg
Keywords 44.8 119.8 30.0 30.2 23.3 Sentences 97.2 431.8 60.2 51.6 87.6

Overall, the initial results of the questionnaires are very promising and sup-
plied us with useful insights for further development. Text summarization in
Italian documents needs to be improved in general. Furthermore, the reduced
recall for some sources can indicate a low coverage of the terminology for a
subdomain.

6 Conclusions
In this work, we presented an approach to automatically extract controlled key-
words and sentence summaries for legal documents from diverse origins and
in different languages. We also evaluated the utility of the generated keywords
through user ratings and alternative keyword selections by legal experts.
We found, that there are major differences in the utility of keywords in the
different languages. While the majority of English and German summary sen-
tences and keywords were found useful, the results are mixed for French, Italian,
and Bulgarian. By comparison of the utility ratings between keywords and sum-
mary sentences in these languages, we have shown for French and Bulgarian
 EuroVoc-Based Summarization of European Case Law 217

that our approach is capable to generate useful summary sentences although the
extracted keyword terms from the EuroVoc terminology do not provide useful
keywords by themselves. Moreover, from the expert selections of keywords, we
have seen that the EuroVoc terminology is suitable for some of them only. We
provided evidence that for Italian other vocabularies are likely to cover the legal
domain more appropriately.
The example of case law from the EUR-Lex source displayed indications
that our approach is able to automatically extract summary sentences in the
different European languages. Here, 84% of the keywords were rated as helpful
by the human experts.
Many interesting possibilities for future work remain. The summarization
pipeline is still ongoing work, but its results are already useful and prove the
feasibility of the architecture and approach. This enables us to further improve
any adjusting screw ranging from specialization of the sentence splitter to more
sophisticated algorithms for text summarization. An important step will be a
more standardized and automatized evaluation of the pipeline with established
evaluation metrics like ROUGE [22] and datasets like the HOLJ corpus [16]
or the Legal Case Reports Data Set5 . A promising option is the integration of
citation information as in [15].
Regarding the processing of documents in different languages, we have estab-
lished a reliable and extensible pipeline. Our approach has proven capable of run-
ning several distinct components within a uniform framework and we have shown
how to cleanly separate language-specific from language-independent processing
steps. By the help of the broker middleware we were able to increase the pipeline
throughput although some of the components did not have native support for
parallelization.

Acknowledgments. Parts of this work have been supported by the European Com-
mission under the 7th Framework Programme through the project EUCases–EUropean
and National CASE Law and Legislation Linked in Open Data Stack (grant agreement
no. 611760). We do also gratefully acknowledge the effort spent by all legal experts for
finishing the questionnaires.

References
1. Alemany, L.A., Castellón, I., Climent, S., Fort, M.F., Padró, L., Rodrı́guez, H.:
Approaches to text summarization: questions and answers. Inteligencia Artif. Rev.
Iberoamericana de Inteligencia Artif. 8(22), 79–102 (2004)
2. Boella, G., Caro, L.D., Humphreys, L., Robaldo, L., Rossi, P., Torre, L.: Eunomos,
a legal document and knowledge management system for the web to provide rel-
evant, reliable and up-to-date information on the law. Artif. Intell. Law 24(3),
245–283 (2016)
3. Boella, G., et al.: Linking legal open data: breaking the accessibility and language
barrier in European legislation and case law. In: Proceedings of the 15th Inter-
national Conference on Artificial Intelligence and Law, ICAIL 2015, pp. 171–175.
ACM, New York (2015)
5
http://archive.ics.uci.edu/ml/datasets/Legal+Case+Reports.
218 F. Schmedding et al.

4. Boella, G., Di Caro, L., Rispoli, D., Robaldo, L.: A system for classifying multi-label
text into EuroVoc. In: Proceedings of the Fourteenth International Conference on
Artificial Intelligence and Law, ICAIL 2013, pp. 239–240. ACM, New York (2013)
5. Brin, S., Page, L.: The anatomy of a large-scale hypertextual web search engine.
Comput. Netw. ISDN Syst. 30(1–7), 107–117 (1998)
6. Chieze, E., Farzindar, A., Lapalme, G.: An automatic system for summarization
and information extraction of legal information. In: Francesconi, E., Montemagni,
S., Peters, W., Tiscornia, D. (eds.) Semantic Processing of Legal Texts: Where
the Language of Law Meets the Law of Language. LNCS (LNAI), vol. 6036, pp.
216–234. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-12837-
0 12
7. Dalal, V., Malik, L.: A survey of extractive and abstractive text summarization
techniques. In: 6th International Conference on Emerging Trends in Engineering
and Technology (ICETET), pp. 109–110. IEEE (2013)
8. Daumke, P., Schulz, S., Markó, K.: Subword approach for acquiring and cross-
linking multilingual specialized lexicons. In: Workshop on Acquiring and Repre-
senting Multilingual, Specialized Lexicons at LREC 2006 (2006)
9. Elfayoumy, S., Thoppil, J.: A survey of unstructured text summarization tech-
niques. Int. J. Adv. Comput. Sci. Appl. 5(4), 149–154 (2014)
10. Erbs, N., Santos, P.B., Gurevych, I., Zesch, T.: DKPro keyphrases: flexible and
reusable keyphrase extraction experiments. In: Proceedings of the 52nd Annual
Meeting of the Association for Computational Linguistics: System Demonstrations,
pp. 31–36. ACL (2014)
11. Erkan, G., Radev, D.R.: LexRank: graph-based lexical centrality as salience in text
summarization. CoRR abs/1109.2128 (2011)
12. Farzindar, A., Lapalme, G.: LetSum, an automatic legal text summarizing system.
In: Legal Knowledge and Information Systems, pp. 11–18 (2004)
13. Ferrucci, D., Lally, A.D.A.M.: UIMA: an architectural approach to unstructured
information processing in the corporate research environment. Nat. Lang. Eng.
10(3–4), 327–348 (2004)
14. Galgani, F., Compton, P., Hoffmann, A.: HAUSS: incrementally building a summa-
rizer combining multiple techniques. Int. J. Hum.-Comput. Stud. 72(7), 584–605
(2014)
15. Galgani, F., Compton, P., Hoffmann, A.G.: Summarization based on bi-directional
citation analysis. Inf. Process. Manag. 51(1), 1–24 (2015)
16. Grover, C., Hachey, B., Hughson, I., et al.: The HOLJ corpus: supporting sum-
marisation of legal texts. In: Proceedings of the 5th International Workshop on
Linguistically Interpreted Corpora (LINC) at Coling 2004 (2004)
17. Gupta, V., Lehal, G.S.: A survey of text summarization extractive techniques. J.
Emerg. Technol. Web Intell. 2(3), 258–268 (2010)
18. Hachey, B., Grover, C.: Extractive summarisation of legal texts. Artif. Intell. Law
14(4), 305–345 (2006)
19. Kim, M.-Y., Xu, Y., Goebel, R.: Summarization of legal texts with high cohesion
and automatic compression rate. In: Motomura, Y., Butler, A., Bekki, D. (eds.)
JSAI-isAI 2012. LNCS (LNAI), vol. 7856, pp. 190–204. Springer, Heidelberg (2013).
https://doi.org/10.1007/978-3-642-39931-2 14
20. Kontonasios, G., Korkontzelos, I., Ananiadou, S.: Developing multilingual text
mining workflows in UIMA and U-compare. In: Bouma, G., Ittoo, A., Métais, E.,
Wortmann, H. (eds.) NLDB 2012. LNCS, vol. 7337, pp. 82–93. Springer, Heidelberg
(2012). https://doi.org/10.1007/978-3-642-31178-9 8
 EuroVoc-Based Summarization of European Case Law 219

21. Kontonatsios, G., Thompson, P., Batista-Navarro, R.T., Mihaila, C., Korkontzelos,
I., Ananiadou, S.: Extending an interoperable platform to facilitate the creation of
multilingual and multimodal NLP applications. In: Proceedings of the 51st Annual
Meeting of the Association for Computational Linguistics: System Demonstrations,
pp. 43–48 (2013)
22. Lin, C.Y.: ROUGE: a package for automatic evaluation of summaries. In: Marie-
Francine Moens, S.S. (ed.) Text Summarization Branches Out: Proceedings of
the ACL 2004 Workshop, pp. 74–81. Association for Computational Linguistics,
Barcelona (2004)
23. Mihalcea, R., Tarau, P.: TextRank: bringing order into texts. In: Proceedings of
Empirical Methods for Natural Language Processing, pp. 404–411 (2004)
24. Moens, M.F.: Summarizing court decisions. Inf. Process. Manag. 43(6), 1748–1764
(2007)
25. Nenkova, A., McKeown, K.: A survey of text summarization techniques. In: Aggar-
wal, C.C., Zhai, C. (eds.) Mining Text Data, pp. 43–76. Springer, Boston (2012).
https://doi.org/10.1007/978-1-4614-3223-4 3
26. Ogrodniczuk, M., Karagiozov, D.: ATLAS multilingual language processing plat-
form. Procesamiento del Leng. Nat. 47, 241–248 (2011)
27. Petrov, S., Das, D., McDonald, R.T.: A universal part-of-speech tagset. In: Calzo-
lari, N., et al., (eds.) Proceedings of the 8th International Conference on Language
Resources and Evaluation (LREC 2012). European Language Resources Associa-
tion (ELRA), Istanbul, May 2012
28. Rocheteau, J., Daille, B.: TTC TermSuite: a UIMA application for multilingual
terminology extraction from comparable corpora. In: Proceedings of the 5th Inter-
national Joint Conference on Natural Language Processing (IJCNLP), Chiang Mai,
Thailand, pp. 9–12 (2011)
29. Saravanan, M., Ravindran, B., Raman, S.: Improving legal document summariza-
tion using graphical models. Front. Artif. Intell. Appl. 152, 51–60 (2006)
30. Sarkar, K.: Multilingual summarization approaches. In: Computational Linguistics:
Concepts, Methodologies, Tools, and Applications, pp. 158–177 (2014)
31. Schweighofer, E., Rauber, A., Dittenbach, M.: Automatic text representation, clas-
sification and labeling in European law. In: Proceedings of the 8th International
Conference on Artificial Intelligence and Law, ICAIL 2001, pp. 78–87. ACM, New
York (2001)
32. Strötgen, J., Gertz, M.: Multilingual and cross-domain temporal tagging. Lang.
Resour. Eval. 47(2), 269–298 (2013)
33. Tanenblatt, M., Coden, A., Sominsky, I.: The ConceptMapper approach to named
entity recognition. In: Calzolari, N., et al., (eds.) Proceedings of the Seventh confer-
ence on International Language Resources and Evaluation (LREC 2010). European
Language Resources Association (ELRA), Valletta, May 2010
34. Yousfi-Monod, M., Farzindar, A., Lapalme, G.: Supervised machine learning for
summarizing legal documents. In: Farzindar, A., Kešelj, V. (eds.) AI 2010. LNCS
(LNAI), vol. 6085, pp. 51–62. Springer, Heidelberg (2010). https://doi.org/10.
1007/978-3-642-13059-5 8
 Towards Aligning legivoc Legal Vocabularies
by Crowdsourcing

Hughes-Jehan Vibert1, Benoit Pin2, and Pierre Jouvelot2(&)

1
Ministère de la Justice, Paris, France
hughes-jehan.vibert@justice.gouv.fr
2
MINES ParisTech, PSL University, Paris, France
{benoit.pin,pierre.jouvelot}@mines-paristech.fr

Abstract. legivoc is the first Internet-based platform dedicated to the diffusion,

edition and alignment of legal vocabularies across countries. Funded in part by
the European Commission and administered by the French Ministry of Justice,
legivoc offers a seamless path for governments to disseminate their legal
foundations and specify semantic bridges between them. In this system
description paper, we outline the general principles behind the legivoc frame-
work and provide some ideas about its implementation, with a particular focus
on the state-of-the-art tools it includes to help crowdsource the alignment of
legal corpora together.

Keywords: Legivoc
Classification
Legal foundations

1 Introduction

Easy and seamless access to law is of key import to states, institutions, businesses and
citizens. Such a challenge is compounded when dealing with states that admit multiple
languages, such as Belgium or Switzerland, or supranational institutions that embed
states with different, even if sometimes similar, legal systems such as the European
Union (EU) or the United Nations. Providing tools that help handling such diversity is
of paramount importance in a connected world where open e-government and access to
knowledge anytime anywhere are taken for granted.
Yet, when dealing with legal systems from different countries and/or languages, “in
order to build a bilingual tool [and hence a multilingual one] granting access to legal
terminology databases, translation will not be the only issue. Equivalence relationships
between languages are not sufficient to convey the modalities of judicial systems
connecting to one another. They shall be kept as such under the right circumstances,
but one must search for the depth of the degree of relationship intrinsically binding two
legal systems: that degree is called functional equivalence” [15]. Aligning functionally
equivalent concepts in a seamless and consistent manner is one of legivoc main goals.
legivoc (all in lower case) is an Internet-based database platform dedicated to the
management of multiple legal information terminologies, with a particular focus on
vocabularies and their alignments [4]. The legivoc system is designed to be used both
interactively and also as an automated Web service, interoperable with other document
management tools or international legislation or translation systems, via a dedicated
© Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 220–232, 2018.
https://doi.org/10.1007/978-3-030-00178-0_14
 Towards Aligning legivoc Legal Vocabularies by Crowdsourcing 221

Application Programing Interface (API). The legivoc web site1 is open and running,
although in an “alpha” version for the time being.
The main goals of legivoc are: (1) to provide access, within a unique framework
and using a general formalism, to (ultimately) all the legal vocabularies of the Member
States of the EU; (2) to foster the use of best practices regarding the encoding of these
vocabularies using Internet standards such as the Simple Knowledge Organization
System (SKOS) and Uniform Resource Identifier (URI); (3) to encourage the creation
of alignment information between these vocabularies, helping provide bridges between
judicial systems based on different laws and languages.
The French Ministry of Justice spearheads the project, partly funded by the
European Commission and the Ministries of Justice of the Czech Republic, Spain,
Finland, France, Italy and Luxembourg. ARMINES and MINES ParisTech are the lead
scientific advisors and implementation specialists for the legivoc project.
legivoc is intended to be used directly by law and judicial experts, for instance
when dealing with cross-border legal issues or planning new legislative regulations. For
such a purpose, knowing how a legal notion in a given vocabulary relates to similar
ones in other countries, a so-called alignment is a key asset. We show how legivoc can
be extended so that entering such information is a very intuitive operation, with the idea
of relying on crowdsourcing efforts (where dedicated individuals perform useful tasks
for the community, as in Wikipedia) to enrich its vocabularies.
The rest of this system description paper is structured as follows. In Sect. 2, we
survey the current and planned systems that provide services somewhat related to the
ones we focus on with legivoc. Section 3 introduces briefly the structure of legivoc and
typical use cases. Section 4 focuses on the alignment process, which intends to build
upon international crowdsourcing efforts to enrich the legivoc database. Section 5
outlines the communication capabilities embedded within legivoc, for use by the
systems that want to take advantage of it. We introduce short- and long-term future
work in Sect. 6, while concluding in Sect. 7.

2 Related Work

For years, computer-based tools have been a staple of legal departments that provide
juridical help to both legal experts and citizens. We review here some of such systems,
focusing on the ones that support multilingual functionalities, and put them in per-
spective with legivoc goals.
EuroVoc [16] is a multilingual (27 languages), multidisciplinary thesaurus that
classifies all EU activities into about 6,800 labeled categories; organized into a 8-level
hierarchy, it is used to loosely group related descriptors. EuroVoc users include the
European Parliament, the Publications Office of the EU (which manages the thesaurus),
national and regional parliaments in Europe as well as national governments. It was
developed to (manually) categorize all relevant documents in order to perform multi-
lingual and cross-lingual search and retrieval in potentially very large document

1
www.legivoc.org.
222 H.-J. Vibert et al.

collections. legivoc differs from Eurovoc by its focus on legal terminologies and their
alignment, while Eurovoc only seeks to define labels that can be used to summarize all
kind of subject matters addressed in EU documents.
CELLAR [5] is multilingual, semantic-based service provided by the Publications
Office of the EU. It is based on a bibliographical ontology specifically designed to
encompass all the material issued by the Office, together with its associated metadata.
legivoc uses the same format as CELLAR, namely SKOS, to encode its taxonomies,
but plans to address law systems both inside and outside of the EU, while using
automated tools to enrich its database from a variety of sources. Moreover, legivoc
intends to ultimately bootstrap its content-building effort via crowdsourcing and
gamification.
EUR-Lex [6] is another effort from the EU which, when completed, will provide
free access to the European law material (treatises, directives, jurisprudential data…) in
24 languages. It uses European Legislation Identifiers (ELI [7, 8]), which uniquely
identify documents. The intent behind legivoc is to provide the same wealth of mul-
tilingual coverage but on a more limited data set, namely legal vocabularies, together
with alignment information between concepts, which will be of use to both legal
experts and translators, and even possibly the public-at-large. legivoc concepts are
identified by specific SKOS concepts, although linking them with ELI-identified
documents could be a valuable addition to the alignment goal targeted by legivoc.
The Inter-Active Terminology for Europe (IATE [9]) is a 10-year old, EU-funded
web-based system that focuses on EU-specific terminology. Its almost 8.5 million
entries target issues well beyond the world of law, up to agriculture or information
technology, targeted by legivoc. Moreover, alignment information between similar
concepts is not present in IATE.
The World Law Dictionary Project [10] (supported by TransLegal and nine law
faculties) intends to provide alignment information from an online database in English
to their equivalent in other languages. The use of a pivot language, here English, is a
key distinguishing feature from legivoc, which in particular for political reasons, must
remain language-agnostic.

3 legivoc 1.0

legivoc materializes as a Web site providing a state-of-the-art multilingual system for

the creation, edition and alignment of international legal vocabularies (see Fig. 1.
legivoc home page).

3.1 User Interface and Capabilities

In its current (6/2015) state, legivoc enables registered users to access the legal
vocabularies of 13 Member States plus Switzerland, in addition to a global one
managed by Eurovoc2. These vocabularies exist in multiple languages; these translated

2
www.eurovoc.europa.eu.
Towards Aligning legivoc Legal Vocabularies by Crowdsourcing 223

Fig. 1. legivoc home page

224 H.-J. Vibert et al.

versions have been either provided by the Member States themselves or automatically
translated via the European Commission MT@EC multilingual service [3].
Words in vocabularies are considered as SKOS concepts. They can be (1) visualized
in various forms (text, dendogram, SKOS source), (2) edited, (3) related to more or less
abstract concepts or (4) aligned to similar concepts in other vocabularies (see Sect. 4).
We illustrate part of a typical display of a concept, here the one of “civil law” (see Fig. 2.
Concept for “droit civil” (excerpt)). It was obtained, after a search for “droit civil” in
legivoc, via the French version of the Eurovoc vocabulary, identified by its International
Standard Organization (ISO) code eu, followed by its legivoc number, 523.

Fig. 2. Concept for “droit civil” (excerpt)

3.2 Technological Foundations

legivoc is built on top of plinn3, an open-source collaborative infrastructure developed
at ARMINES and MINES ParisTech. This environment, based on the Zope multi-
purpose Web framework system, provides sophisticated ways to manage documents in
various formats, with a particular emphasis on users’ access rights, for security pur-
poses, and workflows, to enforce best practices. The existing legal vocabularies pro-
vided by the State Members use various formats (ASCII, Excel, Word, XML,
SKOS…); thus heuristics have been developed and implemented to encode them into a
common legivoc-specific SKOS format. Introducing new legislative corpora, e.g., from

3
www.plinn.org.
 Towards Aligning legivoc Legal Vocabularies by Crowdsourcing 225

countries wishing to join legivoc, requires the design and implementation of new
approximate algorithms, unless the data is already encoded into the legivoc SKOS
format, which is strongly advised.
As an example, the SKOS encoding used by legivoc for the “droit civil” concept
(see Fig. 2. Concept for “droit civil” (excerpt)) is provided below (see Table 1). After a
constant list of definitions specifying the chosen encoding formats, each concept is
given a unique number (here 523 in the Eurovoc vocabulary). Alignment information
(see Sect. 4) is then followed by a list of alternate and preferred labels. The list of exact
strings corresponding to each label completes the whole specification.

Table 1. SKOS encoding of the “droit civil” concept (excerpt)

@prefix dct: <http://purl.org/dc/terms/> .

@prefix legivoc: <http://legivoc.org/namespaces/2014/legivoc#> .
@prefix rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .
@prefix skos: <http://www.w3.org/2004/02/skos/core#> .
@prefix skosxl: <http://www.w3.org/2008/05/skos-xl#> .
@prefix xml: <http://www.w3.org/XML/1998/namespace> .
@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .

<http://eurovoc.europa.eu/523> a skos:Concept ;
skos:closeMatch <http://legivoc.org/be/932>,
<http://legivoc.org/fi/1619>,
<http://legivoc.org/fi/2270>,
<http://legivoc.org/nl/3792>, ...
skos:narrower <http://eurovoc.europa.eu/164>,
<http://eurovoc.europa.eu/186>,
<http://eurovoc.europa.eu/3497> ...
skos:related <http://eurovoc.europa.eu/5496>
<http://eurovoc.europa.eu/576> ;
skosxl:altLabel <http://eurovoc.europa.eu/523/label/11>,
<http://eurovoc.europa.eu/523/label/13>,
<http://eurovoc.europa.eu/523/label/15>, ...
skosxl:prefLabel <http://eurovoc.europa.eu/523/label/0>,
<http://eurovoc.europa.eu/523/label/1>,
<http://eurovoc.europa.eu/523/label/10>, ...
<http://eurovoc.europa.eu/523/label/0> a skosxl:Label ;
skosxl:literalForm "гражданско право"@bg .
<http://eurovoc.europa.eu/523/label/1> a skosxl:Label ;
skosxl:literalForm "Derecho civil"@es .
...

Even though the presence of such diverse formats and possibly incompatible
semantics is a clear challenge and the existence of a general, widely adopted vocab-
ulary format would be a desirable feature, we take, with legivoc, a “can-do” approach,
and try to leverage existing corpora instead of waiting for an eventual standard. In fact,
our own legivoc SKOS format can be seen as a first proposal in this direction.
226 H.-J. Vibert et al.

legivoc strongly adheres to W3C standards, such as (1) the Resource Description
Framework (RDF), which uses 3-tuples (subject, predicate, object), also called triples,
to represent data, (2) the URI naming convention, to denote such resources, (3) SKOS,
an RDF-based representation format for vocabularies, and (4) the SPARQL Protocol
and RDF Query Language (SPARQL), to search and access triples in an effective and
expressive manner.
IT specialists can write and execute (possibly from a remote site, see Sect. 5)
legivoc SPARQL commands to answer specific requests emanating from law or judicial
experts (see Fig. 3. Querying the Greek vocabulary).

Fig. 3. Querying the Greek vocabulary

In this particular example, one asks for the list of all concepts, given by their URI,
and their preferred labels, as strings, present in the Greek vocabulary. An excerpt of the
output is given below (see Fig. 4. Greek concepts and labels (excerpt)).

Fig. 4. Greek concepts and labels (excerpt)

 Towards Aligning legivoc Legal Vocabularies by Crowdsourcing 227

4 Alignment Management

Alignment is a feature supported by SKOS, via the predefined closeMatch, ex-

actMatch, broadMatch, narrowMatch and relatedMatch mapping prop-
erties that can relate two concepts in a somewhat hierarchical manner. Table 1 provides
an example of alignment information; there, for instance, the concept of “droit civil” is
mentioned as a close match to Concept 932 in the Belgian corpus (also called
“Burgerlijk Recht”, in Flemish – recall that Belgium admits two official languages). We
believe the new approach to alignment management, described below, is one of the key
features of legivoc.

4.1 Collecting Alignments

Introducing alignment information between vocabularies greatly enriches the semantic
knowledge already embedded in legivoc. Yet, the size of the legivoc database, with its
(up-to-now) 13 vocabularies sporting each an average number of concepts in excess of
7,000, would make this an expensive undertaking if one were to only use the text-based
concept (and alignment) editor provided by legivoc.
One possible approach to handle data in such high volumes is to rely on automatic
semantic analysis tools and techniques, e.g., machine learning, to suggest possible
alignments to the user, an approach akin to the one advocated, for instance, for the
EUCases project [14]. Given the more homogeneous nature of our data (only vocab-
ularies), we decided to try another route.
We designed a streamlined and intuitive alignment environment within legivoc so
that any benevolent user can help improve, in a friendly and even possibly “fun” way,
the budding ontology induced by alignments. We indeed expect legivoc alignment
content to be built from the ground up via crowdsourcing, building upon ideas stem-
ming from human computation [1]. Providing a natural affordance for the introduction
of alignments, using a simple but dynamic graphical user interface (see below), is a first
step towards motivating the law community.
Another tool to increase participation we have had in mind from the very start of
the project is gamification [2, 11], which has been shown to significantly improve
intrinsic motivation in a wide variety of situations [12] and, interestingly enough, even
when dealing with alignment in natural languages [18]. We view the entirely graphical
and playful interface we designed, in opposition to mostly menu- and text-based
interfaces seen in more traditional ontology-building tools such as Tematres [13] for
instance, as a first step in this direction.
With its unique alignment extension environment, legivoc is now open (in read
mode) to any user interested in participating in this ontology-building effort by pro-
viding new alignment information. To register as a legivoc user, the only user data
required is a valid email address and an optional Twitter account. Indeed, all align-
ments are reported on Twitter, enabling following users to discover in real time when
new alignments are being added and by whom, thus creating a sense of community and
challenge. New alignments are currently checked a posteriori by law experts, to prevent
as much as possible the input of possibly incorrect or conflicting information in the
legivoc database; if need be, malevolent users could be blocked in short order and their
228 H.-J. Vibert et al.

actions reverted by a SPARQL command. In the future, this group of experts could be
supported by selected digital volunteers (see for instance [17] for examples of such
programs).

4.2 Using Alignments in legivoc

There are currently in legivoc three new modalities for introducing, in an intuitive
manner, new concept alignments involving two different vocabularies. They vary
according to how the relevant concepts are found:
• list, where concepts are displayed in alphabetic order;
• search, where the user inputs keywords to legivoc, which returns a list of concepts
matching those words;
• pick, where the system chooses at random a given concept, and challenges the user
to find a related concept in the other vocabulary.
The user then only has to mouse-select one concept in a given vocabulary and drag
it on top of the other concept, in the second vocabulary. This process is made even
easier and more informed via the use of dendograms, which display in an intuitive
manner the hierarchical structure of law concepts (see Fig. 5. Aligning the UK and
English-translated Spanish vocabularies, for a particularly rich example).
A typical illustration of the alignment environment is provided above (same figure).
The top half of the screen is dedicated to the UK vocabulary, in list mode (by clicking
twice on a concept name, here “communication”, its partly elided dendogram has
appeared), while the second half is in search mode on the Spanish vocabulary. Note
that we use the English version of the Spanish law, obtained via MT@EC. The user has
already dragged the blue bullet along the word “Media and the press” on top of “Public
communication”, thus creating an alignment, reported in the left window.
As described above, we are currently experimenting on the motivational aspects of
our approach. As already mentioned and as graphically indicated by the ticked box
dedicated to alignment tweeting, all alignments are tweeted on the @legivoc account,
in the hope of motivating users. We also intend to use in the future this medium to keep
users informed of legivoc developments and upgrades. Ultimately, users could also
contribute their own tweets on this account, on a limited basis.

5 Using legivoc as a Web Service

We already alluded to the possibility of accessing legivoc remotely. This can currently
be done in two ways: structured or not.
• The first one is via SPARQL commands (see Sect. 3) sent to a dedicated legivoc
URL: legivoc.org/sparql_form. The first argument, query, is a string
corresponding to the command to run, while the second one, db, is the ISO country
code of the vocabulary on which the command is to be run. The output uses the
XML format required by the SPARQL specifications. This service allows arbitrary,
Fig. 5. Aligning the UK and English-translated Spanish vocabularies
Towards Aligning legivoc Legal Vocabularies by Crowdsourcing
229
230 H.-J. Vibert et al.

non-modifying requests to be run on the legivoc database, thus providing a powerful

and generic API for remote users and servers.
• The second access method is via the legivoc.org/search_concepts URL.
It can be used to perform non-structured, textual searches on the legivoc database.
The output, which includes the matching concepts URIs, with some additional data,
is formatted according to the JavaScript Object Notation (JSON) standard.
Alignments, being encoded within legivoc as full-fledged triples, can be retrieved
remotely as well as all other data elements.
Obvious security concerns can naturally be raised by such powerful remote access
mechanisms. legivoc relies on dedicated cookies to ensure proper, fine-grained user
rights management; these small pieces of data are transferred and locally stored after
performing POST requests at the legivoc.org/logged_in URL, with proper
name and password parameters. These transferred cookies will have to be passed along
subsequent data access requests, ensuring that only registered users are granted access
to the system.
In order to validate legivoc’s remote access approach, it is currently being tested
within the legicoop4 network for legislative cooperation between the Ministries of
Justice of the EU.

6 Future Work

Future work will first be dealing with the actual full-fledged opening of the legivoc web
site, and assessing how the corresponding traffic is handled by our platform. Unitary
testing of all of legivoc features has been performed, and a few key users have been
granted access to the “alpha” version of the site.
At a more theoretical level, we want to address reasoning over alignment infor-
mation, e.g., regarding (1) transitive alignment properties (“if a is aligned with b and
b is aligned with c, then the alignment of a with c should be present in the legivoc
database”) or (2) alignment semantics, using for instance natural language processing
tools to infer implicit alignments.
On the motivational front, we want to measure the effectiveness of the tools used in
legivoc to fuel the alignment process. In addition to relying on community effects, one
could, if need be, envision looking at a higher degree of gamification.
Another venue for development is the extension of handled vocabularies, for either
more countries (Africa is one target we are considering) or different application
domains (commercial law, ecology, human rights). Another issue related to vocabulary
extension is the matter of the proper management of their updates; fortunately, even
though this is an important problem, it is not a very frequent one, given the generally
relatively slow evolution of law terminology.
Finally, linking legivoc with other law-oriented multilingual systems presents
interesting opportunities. We already mentioned the EU-Cases project. Another fas-
cinating prospect could be Eunomos [19], currently dedicated to customer law and

4
www.legicoop.eu.
 Towards Aligning legivoc Legal Vocabularies by Crowdsourcing 231

where legivoc data could be used to extend its current analysis procedures and appli-
cation domain [20].

7 Conclusion

We presented a new approach to the alignment input process for the international legal
vocabularies stored within the legivoc infrastructure. Heavily based on existing W3C
standards, it offers both an intuitive and even possibly “fun” interactive interface and a
remote API. We rely on motivational techniques based on social networks tools such as
Twitter to (hopefully) increase the amount of alignment information required to make
legivoc a success.
The goal pursued by the legivoc project, and the joined forces of computer systems,
AI and advanced HCI techniques, is to provide a key tool for economic, legal and
political intelligence. Ultimately, we will be looking at ways of offering our enhanced
data in Open access mode, while enforcing some country-specific copyright
restrictions.

Acknowledgments. We thank Claire Medrala for her help with the implementation of the
legivoc Twitter interface. We also thank the anonymous reviewers of the 2015 LST4LD and
AICOL workshops for helping us improve our paper.

References
1. Michelucci, P. (ed.): Handbook of Human Computation. Springer, New York (2013). https://
doi.org/10.1007/978-1-4614-8806-4
2. Deterding, S., Sicart, M., Nacke, L., O’Hara, K., Dixon, D.: Gamification using game-design
elements in non-gaming contexts. In: CHI 2011 Extended Abstracts on Human Factors in
Computing Systems (CHI EA 2011). ACM (2011)
3. Pilos, S.: European Commission Machine Translation and Public Administrations. Legivoc
conference, Brussels, Belgium (2014). www.youtube.com/watch?v=B_rDUisXaB8. Acces-
sed 7 Dec 2015
4. Vibert, H.-J., Jouvelot, P., Pin, B.: Legivoc - connecting law in a changing world. J. Open
Access Law 1(1), 1–19 (2013)
5. www.joinup.ec.europa.eu/sites/default/files/da/c3/de/CESAR-community_CELLAR.pdf.
Accessed 7 Dec 2015
6. www.eur-lex.europa.eu/homepage.html. Accessed 7 Dec 2015
7. www.eur-lex.europa.eu/eli/reg/2013/216. Accessed 7 Dec 2015
8. www.eli.fr/en. Accessed 7 Dec 2015
9. www.iate.europa.eu/about_IATE.html. Accessed 7 Dec 2015
10. www.translegal.com/legal-english-dictionary. Accessed 7 Dec 2015
11. Walz, S., Deterding, S. (eds.): The Gameful World. The MIT Press, Cambridge (2014)
12. Rigby, C.S.: Gamification and motivation. In: [11] (2014)
13. www.vocabularyserver.com. Accessed 8 Dec 2015
14. www.eucases.eu. Accessed 7 Dec 2015
15. Mazet, G.: Jurilinguistique et informatique juridique. IRETIJ, Université de Montpellier
(2001)
232 H.-J. Vibert et al.

16. www.eurovoc.europa.eu. Accessed 17 Mar 2016

17. www.onlinevolunteering.org. Accessed 12 May 2017
18. Poesio, M., Chamberlain, J., Kruschwitz, U., Robaldo, L., Ducceschi, L.: Phrase detectives:
utilizing collective intelligence for internet-scale language resource creation. In: Proceedings
of the 24th International Joint Conference on Artificial Intelligence (IJCAI 2015). AAAI
Press (2015)
19. Boella, G., Di Caro, L., Humphreys, L., Robaldo, L., Rossi, P., van der Torre, L.: Eunomos,
a legal document and knowledge management system for the web to provide relevant,
reliable and up-to-date information on the law. Artif. Intell. Law 24, 245–283 (2016)
20. Boella, G., Di Caro, L., Robaldo, L.: Semantic relation extraction from legislative text using
generalized syntactic dependencies and support vector machines. In: Morgenstern, L.,
Stefaneas, P., Lévy, F., Wyner, A., Paschke, A. (eds.) RuleML 2013. LNCS, vol. 8035,
pp. 218–225. Springer, Heidelberg (2013). https://doi.org/10.1007/978-3-642-39617-5_20
 Data Protection in Elderly Health
Care Platforms

Angelo Costa1 , Aliaksandra Yelshyna2 , Teresa C. Moreira2 ,

Francisco C. P. Andrade2 , Vicente Julian3 , and Paulo Novais1(B)
1
Centro ALGORITMI, University of Minho, Braga, Portugal
{acosta,pjon}@di.uminho.pt
2
School of Law, University of Minho, Braga, Portugal
yelshyna@gmail.com, {tmoreira,fandrade}@direito.uminho.pt
3
Departamento de Sistemas Informáticos y Computación,
Universitat Politècnica de València, Valencia, Spain
vinglada@dsic.upv.es

Abstract. Ambient Assisted Living provides solutions to the increas-

ing cognitive problems that affect the elderly population. To provide
all features possible, Ambient Assisted Living projects require access
to personal and private information of their users. Currently, the legal
issues arisen in Ambient Assisted Living are a hot topic in the Euro-
pean Union, especially aspects regarding unsupervised data processing
and cross-sharing of personal information. In this paper it is presented
the iGenda project, which is a Cognitive Assistant inserted in the Ambi-
ent Assisted Living area which aims to build safe environments that
adapt themselves to one’s individual needs. However, one of the issues
is the protection of the data flowing within the system and the protec-
tion of user’s fundamental rights. It is also presented the principles and
legal guarantees of data protection and transmission, and legal aspects
are explained, embracing appropriate solutions to technological features
that may be a threat.

Keywords: Healthcare platform · Ambient Assisted Living

Data protection · Privacy · IGenda

1 Introduction
The dramatic increase in average life expectancy during the 20th century was
achieved due to social and economic stability and medical improvements [18].
According to the UN report, in the year 2050 the elderly population is expected
to be over 2 billion [18]. This growth of human longevity and birth rate decrease
is threatening the sustainability of health systems and forces to rethink health
care planning and provision [19].
Furthermore an important issue is the great medical care that the elderly
population needs. Currently there are efforts to provide technological solutions,
based on the Ambient Intelligence (AmI) and Ambient Assisted Living (AAL)
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 233–244, 2018.
https://doi.org/10.1007/978-3-030-00178-0_15
234 A. Costa et al.

concepts, which allow their users to stay at home, providing medical assistance
through the use of devices and services that connect them with their physicians
and medical staff [7,11,12,16,20].
The AAL is focused in people with some form of disabilities; the frameworks
developed usually target the elderly population. The two main reasons why the
elderly are chosen are: the considerable size of population they represent and the
challenge that they pose as most possess more than one disability.
In recent years, some projects developed became a reference to the AAL area,
because their goals, architecture or innovation. These projects are: AAL4ALL
[15], Care4Balance [1], RelaxedCare [13]. The AAL4ALL has presented new ways
of communicating with heterogeneous devices and services using the IEEE 11073
and the HL7 as base standards, which are commonly used in the medical area.
The Care4Balance presented a new perspective in terms of gathering information
of the caregiver and care-receiver. The RelaxedCare aims to create a novel social
network that connects its platform users with their relatives. There are more
projects but these are what seem more relevant in the area. The AAL area differs
from the medical area because the information it has and the environment where
the platforms are deployed.
Medical devices are tested and certified to achieve a high level of protec-
tion for human health and safety and a good functioning without any harm or
malpractice to its users (Directives 90/385/EEC, 93/42/EEC and 98/79/EC).
Therefore, they are very restricted in terms of features and the type of informa-
tion the possess or generate. AAL projects usually require complete information
about the users because most of the features rely on Artificial Intelligence pro-
cesses, which consume big amounts of data.
The use of the AmI and AAL systems may present some difficult issues from
a legal point of view due to the monitoring procedures and the cross-sharing
of sensible information intend a serious risk of privacy loss. The AAL4AAL
made several efforts to create a standard that encompasses the exigencies that
AAL projects require, this would allow them to be equivalent to medical devices
in terms of legal frameworks. Until now there are no advances in this field.
Therefore, AAL projects have been barred the access to the medical environment,
although that did not stop the development of medical features hoping that the
regulations change.
Field tests performed on these projects, and others, reveal that there is a
generalized acceptance by the elderly population and by the medical staff. The
successful results show that these type of projects are needed and there is a
market for them. The issue relies on the privacy and data protection. Something
that the RelaxedCare project is working on, because there is nothing more prone
to invade privacy than a social network. Enforcing encryption and social tools,
as social spheres can be a way of keeping private data secure. One approach
of attempted encryption is introduced by Doukas et al. [6], which proposes the
introduction of Public Key Infrastructure (PKI) encryption [17] on sensor gate-
ways, disabling middle-man attacks and packet sniffing. This security level is
appropriate to secure remote data transmission, where data has to pass several
 Data Protection in Elderly Health Care Platforms 235

internet nodes but it has a high computational cost. This is unfeasible to be

implemented in internal data exchanges. The number of connections established
in just one second rely on fast response, thus the overhead required is imprac-
tical. Furthermore, the complexity of AAL systems communications exceed the
peer-to-peer type, which would require multiple keys to each user, increasing
substantially the complexity of the encryption/decryption process.
This paper presents a discussion of the dichotomy between the current legal
framework and ALL technology using the iGenda as an example, focusing on the
privacy and data protection concerns. The aim is to find solutions to harmonize
the technological advances and data protection requirements for current and
future laws.
In the following section it is going to be presented an AAL platform, iGenda,
showing the data transferred and user access to it. In Sect. 2 it is presented
the current data protection framework and the legal warranties related to the
AAL action area. Section 3 explains legal aspects of data storage and access law
procedures. And finally, Sect. 4 presents the conclusions of the paper.

2 The iGenda Example

The AAL and AmI aim to build safe environments that adapt themselves to one’s
individual needs. Typically used in home environments (that can be adapted to
nursing homes and others alike) AAL platforms are built with cost in mind, thus
resorting to commercially available devices and software to implement their fea-
tures. The goal is to deliver medical assistance to one’s home, therefore decreas-
ing hospital stays and visits sustaining the familiar feeling that a home provides.
The use of AAL systems require a large amount of personal information
about the users of those systems, such as personal health record, data about
social contacts, domestic activities, and physical location.
To better demonstrate the AAL concept we present the iGenda project [4,
5,10] that is an AAL platform that uses mobile devices and sensor systems to
collect and process vital data, displaying them via mobile devices or the iGenda
administration web-page. These procedures aim to improve the well-being of the
users (the care-receivers) by creating a compendium of health data that can help
to identify health problems or critical events.
In terms of features, the iGenda primary feature is to be a communication
platform with an calendar manager that intelligently schedules regular events,
plans social events and, directed to the medical staff, schedules medical appoint-
ments with the care-receivers, facilitating the creation of shared events.
There are three major actors in iGenda: the care-receivers (elderly or men-
tally impaired people), the caregivers (physicians or family/relatives), and the
relatives (family and friends). They have access to specific information tailored
to them, according to their needs. For instance, the care-receivers have no need
to receive extensive medical information as it would only confuse them.
Apart from these three actors there is also the technician who is a trained
professional responsible for the iGenda system and who is bound by a contract.
236 A. Costa et al.

iGenda relies on data, in fact, without a large amount of data about its users
it will not operate correctly. The platform uses a profiling method based on likes
and dislikes of the users so it can suggest activities that please them. Thereon,
the platform can schedule shared events of leisure activities that please all the
participants and that also comply with the active-aging objective.
To find activities that are pleasant to all of the users, the system searches
their activities database for similar events. The events have their own ontology,
which relies on well-defined tags to each activity. Therefore, all activities are
described the same way and their introduction is done by a iGenda technician.
The similar activities are ranked by a weighted algorithm that analyses each
activity classification (according to each user) and produces a new classification.
The higher classified activity is then scheduled in a timeframe common to all
participants (that anyone has no activities). For instance, if 4 users (that know
eachother) like playing cards and have the Monday afternoon free the iGenda is
able to schedule a card game on that time period.
Furthermore, the caregivers have the responsibility to care for the care-
receivers that are assigned to them (they can be formal or informal, such as
relatives or friends) and receive extended health or personal information about
each care-receiver, effectively entering the private sphere of each user.
One of the great privacy protection issue is that in iGenda (and in most
of the AAL projects) the information is shared and viewed by several users,
some are bound by confidentiality obligations and others are not. Furthermore,
the information will be present in iGenda as long as possible, e.g., at least as
a specific user is registered in the system but in may be present for a longer
period. These choices were taken so the platform is able to relate all information
and social connections, thus being able to provide accurate event suggestions
and health reports that are grounded to the common medical history.

3 The Data Protection Framework for AAL Systems

European Union legislation on personal data protection is presented by the

Directive 95/46/EC (Personal Data Protection Directive) that was adapted in
Portugal by the Law 67/98, 26 October on Data Protection (Portuguese Data
Protection Act) and Directive 2002/58/EC (Directive on Privacy and Electronic
Communications). The currently applicable Directive 95/46/EC is being revised
and in the near future will be replaced by the General Data Protection Regu-
lation (GDPR) which establishes rules adapted to the digital era and aims to
harmonize data protection rules in the EU, introducing some new principles of
data protection: data protection by default and data protection by design, which
will then guarantee that data protection safeguards are being incorporated in all
planning phases of development of the AAL solutions.
iGenda (and AAL platforms in general) collect and process personal and
health data which is particularly sensitive and therefore requires special protec-
tion in accordance with the Directive 95/46/CE and Convention for the Pro-
tection of Individuals with regard to Automatic Processing of Personal Data
 Data Protection in Elderly Health Care Platforms 237

(Convention N.◦ 108). In Portugal, article 35 of the Portuguese Constitution

stipulates a general prohibition of processing personal data, forbidding the use
of informatics for the treatment of data concerning the private life of the citi-
zens [9]. On the other hand, both the Portuguese Law 67/98 and the Directive
95/46/CE have specified that, within the prohibition of processing sensitive data,
in addition to all data concerning the private life of the citizens, health, sexual
and genetic data must be also included. However, there is an obvious exception
to this general prohibition: the case when the data subject expressly consents,
through free informed will, without any kind of coercion, being totally aware of
all the effects arising out of his/her manifestation of will [2]. This requirement of
a free and express consent is obviously related with the legal principles of per-
sonal data protection. First of all, the principle of transparency, meaning that
the person responsible for the data collection and processing must be clearly
identified and the data subject must be informed on its purpose and also on
the delays for keeping the data, as well as the possibility and conditions of its
communication to third parties.
One of the this principles that constitutes the truly fundamental and main
principle of data protection is the limitation principle or purpose principle (par-
tially embodied in article 6◦ , n. 1, paragraph (b) of the Directive and in article
5◦ , n. 1, paragraph (b) of the Portuguese Law 67/98). The referred principle
prohibits further processing that is incompatible with the original purpose(s)
of the collection. Personal and health data collected via AAL should only be
processed for the purpose of providing AAL services and should not be used for
any way incompatible with those purposes.
Article 6◦ of the Directive and article 5◦ , n. 1, paragraph (c) of the Portuguese
Law 67/98 incorporate this principle of proportionality by stating that personal
data must be adequate, relevant, and not excessive in relation to the intended
and legitimate purposes for which personal data are collected and/or processed.
Also, the processing of personal data must be strictly limited to the minimum
required to achieve the AAL objectives, according to the minimization principle.
In addition, each consultation of personal data that is available through the AAL
should be justified by a real necessity of providing care, treatment or medicine
prescription.
The proportionality principle is associated to the quality of the personal data
which imposes personal data to be pertinent, kept up-to-date and not excessive in
relation to the purpose for which they are collected. The treatment of personal
data can only take place when it is indispensable for the initial purpose and
irrelevant data should not be collected (article 6◦ , n. 1, paragraph (c) of the
Directive).
According to the retention principle the collected data should not be retained
in these systems for longer than necessary (see article 6◦ , n. 1, paragraph (e)
of the Directive 95/46/CE and article 5◦ , n. 1, paragraph (e) of the Portuguese
Law 67/98). This is an ambiguous principle as the maximum time is not defined
and can be abused. This principle is established to prevent abuses and enforce
legal protection if abuses are done. In the case of AAL projects, and in iGenda,
238 A. Costa et al.

some features require the collected data to be permanent, and that constitutes
an abuse of privacy. Therefore, this issue is more relatable to ethical concerns,
as while it is not illegal to keep the information a large period of time it may be
considered unethical because users may not be aware of such time period. It is
hard to fathom the concept of “forever” and what it means, thus most people
cannot make an informed decision about the data their are surrendering.
In AAL platforms the care-receiver has the right to access and verify, without
any need of substantiation, if the data concerning himself/herself are (or not)
correct and updated. The provision of this information is necessary to satisfy the
requirement of fair and lawful processing under the Data Protection Directive
and also ensures informational self-determination [2,14].

4 Legal Aspects for Data Storage

The main issue of AAL platform consists on using sensors and profiling tech-
niques that create a large amount of personal information (including health
data) flowing the system. However, the health data is considered by European
and Portuguese law as “sensitive data”, thus requiring reinforced protection.
Nevertheless, monitoring and profiling must be done in order to accomplish the
minimal requirements for the platform operation, which does not mean that legal
aspects are breached. It’s important to guarantee the protection of the personal
data in the iGenda project, in a way allowing the care-receiver to benefit from
the available services and, at the same time, having all warranties of fundamental
rights being respected.

4.1 Health Data

All AAL platforms must collect health data about their users and store it for
historical operations, personal health records and future medical actions based
on previous conditions. Therefore, iGenda is confronted with the difficult decision
of which categories of personal data, particularly health data, should be collected
and stored.
Health data is sensitive data according to Portuguese and European law
and its processing may not be authorized in all situations, unless there is an
explicit consent of the data subject and additional data security measures are
available (article 8◦ of the Directive and article 15 nr. 3 of the Portuguese Law
67/98). An exception to the requirement of free and informed consent occurs
when the care-receiver is temporarily unable to express consent (for instance,
because he/she is in coma or totally unconscious) and, yet, the data collection
or processing is absolutely essential in order to protect a vital interest of the care-
receivers (usually life or death situations) and in this case, the fundamental right
to life will always prevail [3]. Another important exception is the treatment of
medical data for purposes of preventive medical actions, medical diagnosis, care
or management of healthcare services that are carried out by health professional
obligated to professional secrecy [2].
 Data Protection in Elderly Health Care Platforms 239

The delicate issue of AAL systems and iGenda platform is centered in the
establishment of limits to this huge flow of collection, storing and transmission
of health data, and these are related with the application of the data protection
principles. Since it is crucial to observe fundamental rights of the individual
(especially regarding the right to be left alone or the right to be forgotten)
so, the data must only be stored while it is absolutely indispensable, assuring
a balance between the collected data and the purposes of its collection and
processing [2]. The care-receiver must always be informed about the presence of
sensors and cameras and what type of personal data is being processed and for
what purposes the data is planned to be used, according to article 11◦ of the
Data Protection Directive. Additionally, to guarantee that only personal data
that is necessary for each specific purpose is processed, we recommend that AAL
platforms should be created with a mechanism of privacy by design and also a
privacy impact assessment before it is used.
The Data Protection Directive directly and indirectly affects the process of
keeping information about the medical history of the people that are super-
vised by AAL platforms. In fact, data protection principles represent an impor-
tant limit to the processing and conservation of personal data under any form,
mainly imposing restrictions in the elaboration of automatic profiles based on
the personal data treated. To provide a secure and reliable medical diagnosis, it
is imperative to have knowledge about previous medical problems. Therefore, by
shortening the lifespan of the information, the Directive restricts the provision
of any type of diagnosis and just responds to immediate problems.

4.2 Profiling

The essential feature in the iGenda that requires a large amount of personal data
is the profiling technique. It automatically creates a database that mirrors the
user’s personality to better emulate the user’s choices in non-critical decisions.
In this database, each user is clearly identified and each one has its own profile
type, such as care-receiver, the caregiver, and other users.
In accordance with the general rules of Data Protection Directive, when the
construction of user’s profiles take place, the iGenda (seen in this context as
an entity) always informs the care-receiver with the following information: the
precise purpose of the collection and processing of his/her personal data (e.g.,
for diagnosis, prevention), identity and contact details of data consumer, the
precise categories of personal data the platform will collect and process, the
recipients of the data entitled with the right of access and rectification, ensuring
the transparency of all of the process of collection and treatment of personal
data and the revelation of information to third parties [8].
These profiles contains various categories of personal data that require differ-
ent degrees of confidentiality, therefore, each user has different access conditions
to the database that includes explicit consent and special technical barriers for
data protection.
240 A. Costa et al.

4.3 Data Access Feature

iGenda features social interaction in highly heterogeneous group of people, con-

necting several of its users and sharing non-vital information among them. The
issue that arises is that there is the possibility of building true knowledge from
the crossing of non-vital information. Thus, by propagating previous rules, each
individual may require being exempt from this feature and the removal of all
information related to him/her. This right is associated to the right to be for-
gotten and the right to be let alone [2]. The care-receiver should be able to ask
about the access of each party and be allowed the possibility of rectification,
deletion or blocking of any parties or entities involved in the exchange of infor-
mation within iGenda. Furthermore, the possibility of sharing personal data is
also kept under control of the care-receiver: him/her can deny the collection and
processing of his/her personal data and refuse access to optional information
using privacy-friendly default options [2], with the downside of losing some or
all iGenda features and services.
In the data protection domain, the integrity of the system and the control
by the individual on data of his own can be achieved by the use of privacy-
enhancing technologies and transparency-enhancing technologies as instruments
capable of helping in the fulfilment of the requirements of informational self-
determination [8]. The iGenda platform includes regular internal checks and
controls of database access, which serve as a protection against intrusions. There-
fore, the module of the Agenda Manager keeps a record of every connection
made through a logging registry, which registers every communication tunnel
established.
Nevertheless, iGenda intensely uses personal data. That is why data security
must be implemented directly in the architecture of AAL (privacy by design),
from the early design stages. The privacy by design approach has been addressed
by the European Commission in their proposal for a General Data Protection
Regulation and it can be a solution to some of the legal problems raised by the
cross exchange of health data in iGenda, while preserving a high level of data
protection. Currently, the iGenda provides some features that followed the pri-
vacy by design concept, e.g., encryption of data, login requirements for sensitive
data, communication obfuscation, etc. These are not enough and more is needed
to keep the information secure. Moreover, privacy by design has to be carefully
considered, as it is relatable to technology and social factors. In terms of iGenda,
the privacy by design is considered only applied to the technological features,
implementing encryption here possible, assuring digital signatures and enforcing
secure database access and communication tunnelling.

5 Technological Implementations

As explained before, iGenda has already some security features implemented.

One is the secure access defined by user/password pair. Dealing with mobile
 Data Protection in Elderly Health Care Platforms 241

devices, there is the possibility of others operating the device unlawfully. The
visual interfaces are designed to be simple, and some information is directly
displayed but, sensible and private information is protected with identification
and passwords. The digital signature assures that the sender of events is the
real person. The issue with the iGenda is that currently it does not enforce
encryption on the message per se, but the content is stamped with the digital
signature. The multi-agent system that sustains the iGenda provides ontologies
and encryption to the underlying message system, which for internal messages
of the platform is secure enough but not to exchange them over the internet.
Technologically speaking, security measures have their positives and nega-
tives. While the positive are easy to see, the negatives are usually increased
complexity and time and resources consumption. While common users may con-
sider that spending a few seconds more in sending a message is acceptable, in a
large scale system that time is not trivial nor unnoticeable. For instance, PKI
encryption is a proven secure method, but it takes an considerable amount of
time to be encoded or decoded (up to a few seconds) that shows an impact when
implemented in low computing power systems like sensor platforms [6]. We have
considered different approaches, like each user possessing a computer system at
home that could decentralize the information but, that would only increase the
number of security measures that would have to be implemented. So, the only
solution to this issue is to wait for more advanced sensor systems that will have
embedded encryption protocols. Until then, we will in the near future implement
encryption in the messaging service of the smartphones, reinforcing the strength
of the digital signature, but accept non-encrypted messages from other systems.
In terms of database security, the implementation relies on the database
provider tools that encrypts in real-time its contents. The issue is the access
to the information, and that is considered a social issue. The automatized ser-
vices do not rely on people to operate and serves the information to the ones
who require them. This situation shifts the burden of responsibility to the users.
iGenda technicians will be scrutinized and under a non-disclosure confidential
contract, which enforces these agents to be private about the information that
they edit. Thus, the main issue related to the databases is the time that infor-
mation is kept.
The profiling methods (and the medical information) require that the infor-
mation about the users are kept at least during the time he/she is registered in
the platform, and in some cases even more. For instance, if a specific user influ-
enced or shared activities with others its information has to stay on the platform
even if the user quits the service. It is our belief that other users must not have
reduced services due to others actions, meaning that to keep the actions’ his-
tory of each user intact, all the participants must be correctly identified. This
is a difficult issue to resolve as it goes against two directives, the right to be
forgotten and the retention principle. While the retention principle is somewhat
easy to be met (as explained before), the right to be forgotten is only partially
possible. There are two approaches to achieve it, delete all data related to the
242 A. Costa et al.

user who wishes to opt-out and void the assumption that users should not be
impaired by other users actions; or to delete only part of the data and go against
the legal ruling. Currently the iGenda is able to partially deleting the informa-
tion of the users that which to opt-out, eliminating all private information but
keeping social information (e.g. name and friends information) and to delete all
information available. Although the latter option is not recommended by us.
As it happens in real life, one cannot be simply deleted from others lives and
the interactions performed do not suddenly become vacant of that person. As
interaction occurs there will always be some type of information trail. There-
fore, until the Data Protection Directive encompasses new rules about social
networks and information sharing it will be impossible to have iGenda features
fully compliant.

6 Conclusions
The present study of the iGenda project has focused mainly on analysis in what
measure its technical features stay in accordance with the legal requirements, in
terms of privacy and data protection and what must be enforced to keep data
transmission and processing within the legal boundaries. Along the previous
sections, we have shown that the procedures of the AAL platforms, besides the
potential risks of privacy loss and unauthorized access to personal data, can be
used in accordance with the current data protection framework.
However, in order to benefit from the iGenda services and, at the same time,
to guarantee all fundamental rights, the care-receiver needs to be allowed to
make use of his legal right on informational self-determination by taking control
on his own data flowing within the system. The acceptability of AAL projects
also depend on an adequately high level of data protection and privacy that
is why security-relevant issues must be identified in the preliminary stages of
development. The environments here AAL projects operate are sensitive and
most require that the frameworks are secure and without information breaches.
We may consider that by the current standards the iGenda is able to be imple-
mented in home environments, but to nursing homes, where the complexity
increases exponentially, there legal considerations (and ethical) become a very
important subject. Considering that the main players are the nursing homes, the
acceptability by them to embrace the iGenda can be questioned and refuted.
The main concern is definitely the protection of the care-receiver and of the
data flowing within the system, in a way allowing him/her to benefit from the
available services and, at the same time, having all his/her fundamental rights
being guaranteed. It was possible to observe that AAL projects are able to follow
current laws but not without losing important features in terms of functionality
that would be essential to improve ones’ health condition.
 Data Protection in Elderly Health Care Platforms 243

Acknowledgements. This work has been supported by COMPETE: POCI-01-0145-

FEDER-007043 and FCT-Fundação para a Ciência e Tecnologia within the Project
Scope UID/ CEC/00319/2013. A. Costa thanks the Fundação para a Ciência e a Tec-
nologia (FCT) the Post-Doc scholarship with the Ref. SFRH/BPD/102696/2014.
This work is also partially supported by the MINECO/FEDER TIN2015-65515-
C4-1-R.

References
1. Care4Balance (2015). http://www.aal-care4balance.eu/
2. e Castro, C.S.: Direito da Informática - Privacidade e Dados Pessoais. Almedina
(2005)
3. Correia, L.B.: Direito da Comunicação Social. Direito da comunicação social, vol.
1. Almedina (2005)
4. Costa, Â., Castillo, J.C., Novais, P., Fernández-Caballero, A., Simoes, R.: Sensor-
driven agenda for intelligent home care of the elderly. Expert. Syst. Appl. 39(15),
12192–12204 (2012). https://doi.org/10.1016/j.eswa.2012.04.058
5. Costa, Â., Novais, P., Corchado, J.M., Neves, J.: Increased performance and better
patient attendance in an hospital with the use of smart agendas. Log. J. IGPL
20(4), 689–698 (2011). https://doi.org/10.1093/jigpal/jzr021
6. Doukas, C., Maglogiannis, I., Koufi, V., Malamateniou, F., Vassilacopoulos, G.:
Enabling data protection through PKI encryption in IoT m-Health devices. In: 2012
IEEE 12th International Conference on Bioinformatics & Bioengineering (BIBE),
pp. 25–29. IEEE, November 2012
7. Grauel, J., Spellerberg, A.: Attitudes and requirements of elderly people towards
assisted living solutions. In: Mühlhäuser, M., Ferscha, A., Aitenbichler, E. (eds.)
AmI 2007. CCIS, vol. 11, pp. 197–206. Springer, Heidelberg (2008). https://doi.
org/10.1007/978-3-540-85379-4 25
8. Hert, P., Gutwirth, S., Moscibroda, A., Wright, D., González Fuster, G.: Legal
safeguards for privacy and data protection in ambient intelligence. Pers. Ubiquitous
Comput. 13(6), 435–444 (2008)
9. Marques, G., Martins, L.: Direito da Informática. Almedina (2006)
10. Novais, P., Costa, R., Carneiro, D., Neves, J.: Inter-organization cooperation for
ambient assisted living. J. Ambient. Intell. Smart Environ. 2(2), 179–195 (2010).
https://doi.org/10.3233/AIS-2010-0059
11. O’Grady, M.J., Muldoon, C., Dragone, M., Tynan, R., O’Hare, G.M.P.: Towards
evolutionary ambient assisted living systems. J. Ambient. Intell. Humanized Com-
put. 1(1), 15–29 (2009)
12. Rashidi, P., Mihailidis, A.: A survey on ambient-assisted living tools for older
adults. IEEE J. Biomed. Health Inform. 17(3), 579–590 (2013)
13. RelaxedCare (2015). http://www.relaxedcare.eu/en/
14. Rouvroy, A., Poullet, Y.: The right to informational self-determination and the
value of self-development: reassessing the importance of privacy for democracy.
In: Gutwirth, S., Poullet, Y., Hert, P., Terwangne, C., Nouwt, S. (eds.) Reinvent-
ing Data Protection? pp. 45–76. Springer, Netherlands (2009). https://doi.org/10.
1007/978-1-4020-9498-9 2
15. Sousa, F., et al.: An ecosystem of products and systems for ambient intelligence -
the AAL4ALL users perspective. Stud. Health Technol. Inform. 177, 263–71 (2012)
244 A. Costa et al.

16. Sun, H., Florio, V.D., Gui, N., Blondia, C.: Promises and challenges of ambient
assisted living systems. In: 2009 Sixth International Conference on Information
Technology: New Generations, pp. 1201–1207. IEEE (2009)
17. Tepandi, J., Tšahhirov, I., Vassiljev, S.: Wireless PKI security and mobile voting.
Computer 43(6), 54–60 (2010)
18. United Nations: World Population Ageing, vol. 7 (2009)
19. United Nations: Population estimates and projections section. Technical report
(2012)
20. Villacorta, J.J., del Val, L., Jimenez, M.I., Izquierdo, A.: Security system technolo-
gies applied to ambient assisted living. In: Lytras, M.D., Ordonez De Pablos, P.,
Ziderman, A., Roulstone, A., Maurer, H., Imber, J.B. (eds.) WSKS 2010. CCIS,
vol. 111, pp. 389–394. Springer, Heidelberg (2010). https://doi.org/10.1007/978-
3-642-16318-0 46
 Assigning Creative Commons Licenses
to Research Metadata: Issues and Cases

Marta Poblet1(&), Amir Aryani2, Paolo Manghi3, Kathryn Unsworth2,

Jingbo Wang4, Brigitte Hausstein5, Sunje Dallmeier-Tiessen6,
Claus-Peter Klas5, Pompeu Casanovas7,8,
and Victor Rodriguez-Doncel9
1
RMIT University, Melbourne, Australia
marta.pobletbalcell@rmit.edu.au
2
Australian National Data Service (ANDS), Caulfield East, Australia
3
ISTI, Italian Research Council, Pisa, Italy
4
Australian National University, Canberra, Australia
5
GESIS – Leibniz Institute for the Social Sciences, Mannheim, Germany
6
CERN, Geneva, Switzerland
7
IDT, Autonomous University of Barcelona, Barcelona, Spain
8
La Trobe University, Bundoora, Australia
9
Universidad Politécnica de Madrid, Madrid, Spain

Abstract. This paper discusses the problem of lack of clear licensing and
transparency of usage terms and conditions for research metadata. Making
research data connected, discoverable and reusable are the key enablers of the
new data revolution in research. We discuss how the lack of transparency
hinders discovery of research data and make it disconnected from the publica-
tion and other trusted research outcomes. In addition, we discuss the application
of Creative Commons licenses for research metadata, and provide some
examples of the applicability of this approach to internationally known data
infrastructures.

Keywords: Semantic web
Research metadata
Licensing
Discoverability

Data infrastructure
Creative commons
Open data

1 Introduction

The emerging paradigm of open science relies on increased discovery, access, and
sharing of trusted and open research data. New data infrastructures, policies, principles,
and standards already provide the bases for data-driven research. For example, the
FAIR Guiding Principles for scientific data management and stewardship [21] describe
the four principles—findability, accessibility, interoperability, and reusability—that
should inform how research data are produced, curated, shared, and stored. The same
principles are applicable to metadata records, since they describe datasets and related
research information (e.g. publications, grants, and contributors) that are essential for
data discovery and management. Research metadata are an essential component of the
open science ecosystem and, as stated in [17], “for a molecule of research metadata to

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 245–256, 2018.
https://doi.org/10.1007/978-3-030-00178-0_16
246 M. Poblet et al.

move effectively between systems, the contextual information around it - the things that
are linked to, must also be openly and persistently available”.
Yet, finding relevant, trusted, and reusable datasets remains a challenge for many
researchers and their organisations. New discovery services address this issue by
drawing on open public information, but the lack of transparency about legal licenses
and terms of use for metadata records compromises their reuse. If licenses and terms of
use are absent or ambiguous, discovery services lack basic information on how
metadata records can be used, to what extent they can be transformed or augmented, or
whether they can be utilised as part of commercial applications. Ultimately, legal
uncertainty hinders investment and innovation in this domain.
The rest of this paper is organised as follows: Sect. 1 presents the most widely
adopted research metadata protocols and practices; Sect. 2 provides some global fig-
ures about the types of licenses used for research metadata; Sect. 3 identifies the main
stakeholders; Sect. 4 reviews the most common choices for metadata licenses and
discusses both advantages and disadvantages of such choices; Sect. 5 offers six com-
pact case studies from different research data services. Finally, the conclusion raises
some questions to guide future work.

2 Research Metadata Protocols and Practices

A number of instruments covering the management of research metadata are currently

available. For example, the Open Archives Initiative (OAI) developed the Protocol for
Metadata Harvesting OAI-PMH to facilitate interoperability between repositories and
metadata service providers [14]. OAI-PMH enables harvesting the metadata of open
access repositories such as PubMed, Arxiv, HAL, the Wikipedia [4], or the World
Bank’s Open Knowledge Repository (OKR).
The Dublin Core Metadata Initiative (DCMI) promotes interoperability and
reusability in metadata design and best practices by developing semantic standards and
recommendations, model-based specifications, and syntax guidelines, such as the
Singapore Framework for Dublin Core Application Profiles or the DCMI Abstract
Model.1
The RIOXX Metadata Guidelines,2 implemented by more than 50 institutional
repositories in the UK [20], have adopted NISO’s “Recommended Practice on Meta-
data Indicators for Accessibility and Licensing of E-Content”3 to add a tag (<li-
cense_ref>) with a reference to a URI carrying the license terms [13]. The main goal is
to provide a mechanism of compliance with the RCUK policy on open access.
While the adoption of these instruments paves the way for technical standardisa-
tion, the discussion subsists with regard to the licensing options available and the
implications of such choices. The issues arise out of the complexity of contractual
obligations that the different types of licenses create, the extent of copyright laws in

1
http://dublincore.org/specifications/.
2
http://rioxx.net/.
3
http://www.niso.org/apps/group_public/project/details.php?project_id=118.
 Assigning Creative Commons Licenses to Research Metadata 247

different jurisdictions, or the difficulties of attribution when metadata are combined or

remixed. Since a distinctive feature of high-quality metadata is that “it is created once
and then reused as needed” [2], transparency and predictability are essential. The
available options come with different requirements, conditions, and scope.

3 Current Use of Creative Commons (CC) Licenses

in Research Metadata

Although many scientific data repositories live behind firewalls in proprietary envi-
ronments, the Web houses thousands of scientific data repositories whose study is now
possible. Marcial et al. [12] manually chose 100 diverse scientific data repositories and
analysed 50 of their characteristics. While copyright issues were out of their scope, two
of the observations referred to the input and output metadata’s rights–distinguishing
between the terms a contributor has to accept before uploading a new record and the
license under which the entire metadata collection is offered. The text excerpts included
in the 100 data repositories referring to these matters showed a huge variety of custom-
made licenses and only two mentions to CC licenses were reported. The earliest data
for this study were collected in 2007 and there is some evidence that the use of
standardized licenses has dramatically increased since then. Yet, an updated study is
still needed.
The Registry of Research Data Repositories by re3data.org (a service of DataCite)
makes its data available for research under an API.4 The Registry, now “the largest and
most comprehensive registry of data repositories available on the web” [15] publishes
an overview of existing international repositories for research data from all academic
disciplines and as of September 2016, listed 1692 data repositories. An analysis of
these repositories reveals that 269 (16%) of repositories made an explicit mention to
CC licenses with a valid URI, while only 17 to Open Data Commons or 9 to GNU
licenses. While these data require some caution (for example, the World Bank’s Open
Knowledge repository applies CC-BY 3.0 in most cases but it does not provide the
corresponding URI) they offer a good snapshot of the current adoption of CC licenses
in the research metadata ecosystem.

4 Main Stakeholders

The following stakeholders can benefit from assigning Creative Commons (CC) li-
censes to the research public metadata:
• Research Management Software Vendors: Assigning CC licenses to research public
metadata will encourage software vendors to incorporate this data into their sys-
tems, leading to better automation in data entry and discovery capabilities of
research management platforms.

4
https://www.re3data.org/.
248 M. Poblet et al.

• Research Institutions: Better research management systems can reduce the cost of
data entry for universities, and enable discovery of research collaboration oppor-
tunities. In addition, universities will be able to demonstrate their collaboration
networks on the public domain using derivative analytics from CC licensed research
metadata.
• Research Infrastructures (including data repositories): research metadata are the key
enablers in creating interoperability between research infrastructures; particularly
for research data repositories, public metadata enables connecting datasets across
multiple systems and enables better discovery and reuse of the research output.
• Researchers: At present, finding related and relevant research, research data and
other scholarly works is not a trivial task for most researchers. Better discovery
tools augmented with public metadata would enable researchers to find related
research and research collaborators, hence finding new research opportunities.
• Publishers: A clear indication on the applicable CC licenses would help to eliminate
the uncertainties about possible consequences of reusing/republishing metadata.
• Funders: The collective effort by universities, publishers, infrastructure providers
and software vendors can enable funders to have a better understanding of the
impact of their funding; moreover, better research collaboration discovery can
improve the return on investment.

5 Applying Creative Commons Licenses to Research

Metadata

We address the issue of assigning clear licenses and terms of use for public research
information by reviewing two frequently used Creative Commons (CC) licenses for
public metadata records: CC0 and CC-BY. Creative Commons discourages the use of
its NonCommercial (NC) or NoDerivatives (ND) licenses on databases intended for
scholarly or scientific use, and they are not open licenses according to the definition of
‘open” by the Open Knowledge Foundation.5It is important to note that CC0 and CC-
BY are not the only open licenses available, as Open Data Commons, to refer to
another popular option, offers three legal tools – the Public Domain Dedication and
License (PDDL), the Attribution License (ODL-By) and the Open Database License
(ODBL)—which covers the European sui generis database right (although now this is
also the case of CC 4.0 licenses). The choices will depend on the objects to be licensed
(creative contents, data, databases, etc.), the clauses and terminology that come with
each choice, the derived contractual obligations, and the mechanisms of enforcement
available to the licensor.
The most accessible form of CC instrument is CC0—“No Rights Reserved” (also
known as Public Domain Dedication).6 This is the choice of research data services such
as Dryad or Figshare for their generated metadata. Increasingly, a number of cultural

5
http://opendefinition.org.
6
https://creativecommons.org/about/cc0.
 Assigning Creative Commons Licenses to Research Metadata 249

institutions such as the Tate Gallery, the Museum of Modern Art (MoMA), the Walters
Art Museum, or the Thyssen Foundation are also releasing their metadata with the CC0
document.
Nevertheless, there are some doubts about the force of the CC0 waiver in some
jurisdictions (e.g. under Australian law), especially with regard to moral rights. As
AusGOAL alerts, “the disclaimer that accompanies CC0, at present, may be ineffective
in protecting the user from liability for claims of negligence.”7 The main issue with
assigning a CC0 document to research metadata is the responsibility to collect the
original records with the CC0 waiver. According to the Creative Commons definition
(CC0 2016) “You should only apply CC0 to your own work, unless you have the
necessary rights to apply CC0 to another person’s work.”8 Hence, unless adequate
provisions are taken, metadata aggregators or repositories would not be able to assign
the CC0 license to records created by other sources. This is why, for instance, Euro-
peana releases all its metadata with the CC0 document and requires its data providers to
waive all IP rights to the metadata provided. Likewise the Digital Public Library of
America (DPLA) requires all data and metadata donors to attach a CC0 document to
any donation [6].
Another popular CC tool for open access works is the CC-BY license that enables
third parties to distribute the work with attribution to the original author. CC 4.0 now
makes this requirement more flexible as it can be done ‘in any reasonable manner based
on the medium, means, and context in which you share the licensed material”.9 A
potential issue when assigning CC-BY licenses to aggregated metadata is that the
sources of metadata records are not always clear. Who owns metadata records? The
researcher who described the work? The research institution who owns the IP?
Moreover, the CC-BY license requires to “indicate if changes were made” which adds
to the complexity of enriching metadata by aggregators (CC-BY 2016). Given these
options, assuming Copyright in metadata seems to be the safest approach. As Aus-
GOAL advises, “recent developments in Australia have led to the situation where it is
unclear which data is subject to copyright. In this situation, Australian researchers have
to take a pragmatic approach and it would seem desirable to assume copyright as
subsisting in all data created in the course of research, and ensure that it is licensed
accordingly. No harm can come from this approach.” [3]. ANDS adds to this, “It will
still serve as a useful way to make known how you would like to be attributed, in
addition to applying a limitation of liability and warranty clause to the data” [1]. In
cases where it is clear that copyright does not subsist in the aggregated metadata,
applying a CC Public Domain mark would suffice, provided the rights to do so have
been established, including consideration that copyright for the material may subsist in
other jurisdictions.

7
https://www.ands.org.au/working-with-data/publishing-and-reusing-data/licensing-for-reuse/faq-for-
research-data-licensing-and-copyright.
8
https://creativecommons.org/share-your-work/public-domain/cc0/.
9
https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode.
250 M. Poblet et al.

6 Case Studies
6.1 CERN
The European Organization for Nuclear Research10 is a hub for the High-Energy
Physics community. The laboratory is used to handling complex and large-scale
datasets. Among many other things, CERN operates a range of platforms and services
related to scholarly information that serve specific community needs. For the purpose
of this case study, two of them should be highlighted here. Both are openly accessible
to the public and aim at fostering (re)use of the disseminated materials. One is
INSPIREHEP, the main scholarly information platform in high-energy physics,
aggregates information from all relevant community resources. Historically this mainly
concerns preprints, but recently has also been extended to include research data. On top
of the content and its metadata, the service provides ‘author pages’ (with ORCID
integration) compiling information about researchers from the scholarly records
available on INSPIRE [8]. The metadata on this platform are shared with a CC0 waiver,
with the expectation that third parties or researchers themselves will use the available
information to compile new services, such as citation statistics. The second example
concerns CERN Open Data which is a dedicated portal to publish data, software and
accompanying research materials such as documentation, trigger files, and tutorials to
enable reuse by any interested audience. Objects are shared with Open Science
licences, data and metadata with the CC0 waiver, and software with the GNU General
Public License (GPL). CC0 was determined to be the best option for promoting the
widest possible exploitation of the curated datasets. In addition, users are asked to cite
the datasets whenever they publish a result based on the shared datasets. Each dataset is
provided with a DOI to facilitate that process (according to Force 11 Data Citation
Recommendations). So far the feedback on the choices made has been positive and
highlighted the appreciation of clear, international and liberal license conditions for
such materials.

6.2 da|ra
da|ra is a registration agency for social science and economics data in Germany.11 It is
run by the GESIS Leibniz Institute for the Social Sciences and ZBW Leibniz Infor-
mation Center for Economics, in cooperation with DataCite (the international con-
sortium promoting research data as independent citable scientific objects). This
infrastructure lays the foundation for long-term, persistent identification, storage,
localization and reliable citation of research data via allocation of DOI names.
Each DOI name is linked to a set of metadata and presents the properties of resources,
their structure and contextual relations. The da|ra Metadata Schema [9] provides a
number of mandatory elements – six core properties –that have to be submitted by the
data centres at the time of data registration. Although da|ra complies with the official

10
https://home.cern/.
11
http://www.da-ra.de/.
 Assigning Creative Commons Licenses to Research Metadata 251

DataCite Metadata Schema, it has broadened the DataCite metadata by adding some
specific properties related to the social sciences and economics. Therefore data centres
may also choose other optional properties to identify their data. da|ra reserves the right
to share provided metadata of the registered research data with information indexes and
other entities. Under German law most of the formal metadata are not subject to
copyright because the threshold of originality is not sufficient enough. [10]. As da|ra
supports the open metadata principles the metadata of the registered research data are
available under CC0 1.0 to encourage all metadata providers (data centres, data
repositories, libraries, etc.) to make their metadata available under the same terms.
Since 2016 da|ra has been offering access to the metadata using the Open Archives
Initiative Protocol for Metadata Harvesting [14]. The da|ra OAI-PMH Data Provider is
able to disseminate records in various formats such as DDI-Lifecycle 3.1 and OAI DC.

6.3 NCI
The National Computational Infrastructure (NCI)12 at the Australian National
University (ANU) has evolved to become Australia’s peak computing centre for
national computational and Data-intensive Earth system science. More recently NCI
collocated 10 Petabytes of 60+ major national and international environmental, climate,
earth system, geophysics and astronomy data collections to create the National Envi-
ronmental Research Interoperability Data Platform (NERDIP). Data Collection man-
agement has become an essential activity at NCI. NCI’s partners (CSIRO, Bureau of
Meteorology, Australian National University, and Geoscience Australia), supported by
the Australian Government and Research Data Storage Infrastructure (RDSI) and
Research Data Services (RDS), have established a national data resource that is co-
located with high-performance computing. NCI uses license conditions, national/
international regulations as guidance for data governance. The license file is required as
the very first step of our data publishing process, this requirement has greatly pushed
the progress of the license investigation within the data provider’s agency. The license
files are published jointly with data through NCI’s OpenDAP server13. More than half
of NCI’s data collection projects have license files published with the metadata records.
The access and use constraints of the rest collections are still being investigated. Most
of the data are quality assured for being ‘published’ and made accessible as services
under Creative Commons Attribution (CC-BY) 4.0 as they are sourced from com-
monwealth government agencies [9]. However, some geophysical surveys were con-
ducted through Australian state government, where the state license is applied while
discussion about moving state license to CC-BY4.0 is happening. Other type of
licenses exist for our international data collections, such as Earth System Grid Fed-
eration (ESGF14) require every single user to register an OpenID as the way to agree
with the ESGF license, Copernicus Sentinel Data and Service Information regulated by

12
http://nci.org.au.
13
http://dapds00.nci.org.au/thredds/catalog/licenses/catalog.html.
14
https://esgf.llnl.gov/.
252 M. Poblet et al.

EU law15, European Centre for Medium-Range Weather Forecasts16 data has more
strict license17 for data access and usage so that it is only available to NCI users who
agree with the license and can only access data through Raijin supercomputer. Different
license often means different project code on the file system as we need to grant
different type of access to each individual project. Therefore, the license condition has
become one of the important indicators for our data management practice. The meta-
data associated with data collection are available under CC-BY4.0. They are publicly
available for users to query the metadata catalogue entries. Our collection level
metadata has also been harvested by national and international aggregators such as
Research Data Australia (RDA) and International Directory Network of Committee on
Earth Observation Satellites (CEOS).

6.4 OpenAIRE
The OpenAIRE18 infrastructure is the point of reference for Open Access and Open
Science in Europe (and beyond) [11]. Its mission is twofold: enabling the Open Science
cultural shift of the current scientific communication infrastructure by linking,
engaging, and aligning people, ideas, resources, and services at the global level;
monitoring of Open Access trends and measuring research impact in terms of publi-
cations and datasets to serve research communities and funders. To this aim, Open-
AIRE offers services [19] that collect, harmonize, de-duplicate, and enrich by inference
(text mining) or end-user feedback, metadata relative to publications, datasets, orga-
nizations, persons, projects and several funders from all over the world.
As for any repository aggregation system, access and usage rights of the 25 million
records collected by the OpenAIRE system plays a rather important role in the lifecycle
of the overall infrastructure. However, from an analysis run in January 2017, out of
2500 publication repository services (supporting the OAI-PMH protocol standard)
registered in the OpenDOAR directory19, only 9 expose metadata license information:
3 with CC-0, 2 with CC-BY, and 4 which require a permission for commercial use,
3 with CC-0 and 1 with CC-BY. Repository managers, as well as their institutions,
seem to convene that metadata records are generally free of access and reuse, and
underestimate the importance of licensing. In order to cope with this “legal gap”,
starting from the end of 2017, OpenAIRE services will request data source managers to
accept Terms of Agreement where they are informed of the intention of OpenAIRE to
collect the records, transform/enrich them, and make the accessible to third-parties with
provenance information. The aggregated OpenAIRE graph is exported via standard
protocols (e.g. HTTP-REST search, LOD, OAI-PMH) and formats, and the metadata
records are available under CC-0, with no restriction of embargo or re-use. This ToA
based approach seemed the only reasonable solution in order to offer properly usable

15
http://dapds00.nci.org.au/thredds/fileServer/licenses/fj7_licence.pdf.
16
https://www.ecmwf.int/.
17
http://dapds00.nci.org.au/thredds/fileServer/licenses/ub4_licence.pdf.
18
https://www.openaire.eu/.
19
http://www.opendoar.org/.
 Assigning Creative Commons Licenses to Research Metadata 253

metadata obtained from a large set of unlicensed data sources: the aggregating system
can safely carry on its function, as metadata records are collected on request of the
repository managers.

6.5 ResearchGraph
Research Graph20 is an example of value added to data infrastructures by third-party
services. Research Graph is a collaborative project by a number of international
partners that links research information (datasets, grants, publications and researchers)
across multiple platforms (Fig. 1). This initiative uses the research metadata to con-
struct a graph of scholarly works, and this graph connects data and publications with
multiple degrees of separation. The outcome enables a researcher to search the graph
for a particular publication or research project and discovers a collaboration network of
researchers who are connected to this work (or topic). The main consideration for such
a service is the ability to read, connect and transform metadata, and without clear
licensing or terms of use, this platform would not be able to include a data infras-
tructure in the graph. In addition, given the mixture of licences provided by different

Version 2.1 (Aug 2016)

Legend •rst name

last name
Relation between nodes Researcher full name*
orcid
* Mandatory eld scopus author id

from_key* title*
to_uri* R R
author list
label Research doi
Data publication year
license
size

R R

title* title
author list* purl
doi Publication R Grant participant list
publication year start year
scopus eid end year

Fig. 1. Research graph metamodel

20
http://researchgraph.org/.
254 M. Poblet et al.

contributors the constructed graph is not publically available as a single dataset.

Instead, it had to be split into separate clusters where the license for each cluster can be
defined from the individual sources. For example, Research Graph holds a cluster of
nodes about Australian Research Grants and Australian Research Data without any
explicit license provided by the metadata source. As such the given cluster, cannot be
integrated into systems that require CC0 license. On the contrary, Research Graph
composed of clusters from da-ra and CERN metadata records that all contain well-
described license leading to a much more usable network of scholarly communication
that can be used by commercial and open access systems with clear terms of use. For
systems with heterogeneous data sources, it is often difficult to find a minimum
technical requirement for all connected components, as such a hybrid technical
approach can be adopted to enable interoperability between systems. However, in the
case of Research Graph, this is very difficult to achieve due to lack of legal interop-
erability between information systems.
Figure 1 shows the Research Graph metamodel. The metamodel contains three four
main Research Graph node types: Researcher, Research Data, Publication and Grant.
The metadata of all these nodes require well defined terms of use to enable consuming
and processing by research management systems and other information systems.

7 Conclusions and Future Work

In this paper we have raised the need for a transparent regime of metadata licensing and
have briefly reviewed the state-of-the art application of open licenses to research
metadata. We have offered some global figures on the use of CC licenses in scientific
research metadata, explored the differences between CC0 and CC-BY, and the
approach taken by five data registries and/or repositories.
The use of standardized licenses fosters reusability and science in general, and
choosing CC or alternative well-known licenses favours the automatic discovery of
usable datasets (e.g. search by license), which can be accomplished by means of the
re3data.org API. An additional advantage of CC licenses is their availability in a
machine-readable form, namely, the license document contains a RDFa which enable
further intelligent processing of the license content (e.g. search by specific conditions).
The examples reviewed did not include research funding bodies, an integral part of
the meta-research ecosystem [16] and the integration of researcher identifiers (such as
ORCID); hence, we believe that these are areas for future investigation. Likewise, it
could be further investigated if research metadata, as compared to other types of
metadata, have enough specificity to require a dedicated set of licenses.
Data licensing has attracted the attention of many researchers in the fields of
Semantic Web, computational linguistics and deontic logic. Datasets of RDF licenses
do exist already [18]. The challenge of developing automated frameworks able to
generate licensing data terms from heterogeneous distributed sources has also been
addressed [7]. NLP techniques to extract rights and conditions granted by licenses and
return them into RDF have been applied [5]. We believe that these technical solutions
offer a number of advantages and deserve to be monitored, reused, and tested in real
scenarios.
 Assigning Creative Commons Licenses to Research Metadata 255

However, the context-dependent problems and ambiguities highlighted in this

paper still survive, such as the discussion of how to share the metadata generated in
scientific research, whether it should accrue the public domain or, rather, whether
scientists and research organisations should retain a legal expression of attribution,
including the implications of such choices. For example, if assigning CC licenses to
remixed metadata requires modifying the records at the source (or get consent from all
creators involved), the process can be difficult to escalate.
This is a domain that requires some previous positioning before making decisions
about the appropriate heuristics for end-users interfaces. The underlying philosophy
and assumptions about building semi-automated ecosystems for data science face
practical legal, commercial, and political issues that require attention.

References
1. ANDS: Australian National Data Service guide on Metadata stores solutions (2016). http://
www.ands.org.au/guides/metadata-stores-solutions. Accessed 5 May 2016
2. ANDS: Guides 2016, Copyright, data and licensing guide (2016). http://ands.org.au/guides/
copyright-data-and-licensing
3. AusGoal: Research data FAQs (2016). http://www.ausgoal.gov.au/research-data-faqs.
Accessed 9 May 2016
4. Bianchini, L.: Metadata in scientific publication (2012). https://www.mysciencework.com/
omniscience/metadata-in-scientific-publication. Accessed 9 May 2016
5. Cabrio, E., Palmero Aprosio, A., Villata, S.: These are your rights. In: Presutti, V., d’Amato,
C., Gandon, F., d’Aquin, M., Staab, S., Tordai, A. (eds.) ESWC 2014. LNCS, vol. 8465,
pp. 255–269. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-07443-6_18
6. Cohen, D.: CC0 (+BY) (2013). http://www.dancohen.org/2013/11/26/cc0-by/. Accessed 5
May 2016
7. Governatori, G., Rotolo, A., Villata, S., Gandon, F.: One license to compose them all. In:
Alani, H., et al. (eds.) ISWC 2013. LNCS, vol. 8218, pp. 151–166. Springer, Heidelberg
(2013). https://doi.org/10.1007/978-3-642-41335-3_10
8. Hecker, B.L.: Four decades of open science. Nat. Phys. 13(6), 523–525 (2017)
9. Helbig, K., Hausstein, B., Koch, U., Meichsner, J., Kempf, A.O.: da|ra Metadata Schema.
Version: 3.1. GESIS - Technical reports 2014/17 (2014). http://doi.org/10.4232/10.mdsdoc.
3.1. Accessed 24 July 2017
10. Klimpel, P.: Eigentum an Metadaten? Urheberrechtliche Aspekte von Bestandsinformatio-
nen und ihre Freigabe. In: Handbuch Kulturportale, Online-Angebote aus Kultur und
Wissenschaft. Eds. Ellen Euler, Monika Hagedorn-Saupe, Gerald Maier, Werner Schweienz
and Jörn Siegelschmidt, Berlin/Boston, pp. 57–64 (2015). https://irights.info/artikel/
eigentum-an-metadaten-urheberrechtliche-aspekte-von-bestandsinformationen-und-ihre-
freigabe-2/26829. Accessed 24 July 2017
11. Manghi, P., Bolikowski, L., Manola, N., Shirrwagen, J., Smith, T.: Openaireplus: the
European scholarly communication data infrastructure. D-Lib Mag. 18, 9–10 (2012)
12. Marcial, L.H., Hemminger, B.M.: Scientific data repositories on the Web: an initial survey.
J. Am. Soc. Inf. Sci. 61, 2029–2048 (2010). https://doi.org/10.1002/asi.21339
13. NISO RP-22-2015: Access License and Indicators, January 2015. http://www.niso.org/apps/
group_public/download.php/14226/rp-22-2015_ALI.pdf. Accessed 6 Sept 2016
256 M. Poblet et al.

14. OAI: The open archives initiative protocol for metadata harvesting (2015). https://www.
openarchives.org/OAI/openarchivesprotocol.html. Accessed 5 May 2016
15. Pampel, H., Vierkant, P., Scholze, F., Bertelmann, R., Kindling, M., Klump, J., et al.:
Making research data repositories visible: The re3data.org Registry. PLoS ONE 8(11),
e78080 (2013)
16. Poblet, M., Aryani, A., Caldecott, K., Sellis, T., Casanovas, P.: Open-Access grant data:
towards meta-research innovation. In 27th International Conference on Legal Knowledge
and Information Systems: Jurix 2014, pp. 125–130 (2014)
17. Porter, S.: A new research data mechanics. Digital Science White Paper (2016). https://s3-
eu-west-1.amazonaws.com/pfigshare-u…/ANewResearchDataMechanics.pdf
18. Rodríguez Doncel, V., Gómez-Pérez, A., Villata, S.: A dataset of RDF licenses. In:
Hoekstra, R. (ed.) Legal Knowledge and Information Systems. The Twenty-Seventh Annual
Conference, JURIX-2014, pp. 187–189. IOS Press, Amsterdam (2014)
19. Schirrwagen, J., Manghi, P., Manola, N., Bolikowski, L., Rettberg, N., Schmidt, B.: Data
curation in the OpenAire scholarly communication infrastructure. Inf. Stand. Q. 25(3), 13–
19 (2013)
20. Walk, P.: RIOXX adoption reaches a half-century (2016). http://www.rioxx.net/2016/04/12/
rioxx-adoption-reaches-a-half-century/. Accessed 7 Sept 2016
21. Wilkinson, M., Dumontier, D., Mons, M.: The FAIR guiding principles for scientific data
management and stewardship. Sci. Data 3, 160018 (2016). https://doi.org/10.1038/sdata.
2016.18%5b21%5d
 Dataset Alignment and Lexicalization
to Support Multilingual Analysis of Legal
Documents

Armando Stellato1(&), Manuel Fiorelli1, Andrea Turbati1,

Tiziano Lorenzetti1, Peter Schmitz2, Enrico Francesconi2,3(&),
Najeh Hajlaoui2, and Brahim Batouche2
1
ART Group, Department of Enterprise Engineering, University of Rome Tor
Vergata, Via del Politecnico 1, 00133 Rome, Italy
stellato@uniroma2.it,
{fiorelli,turbati}@info.uniroma2.it,
tiziano.lorenzetti@gmail.com
2
Publications Office of the European Union, Luxembourg City, Luxembourg
{Peter.SCHMITZ,
Enrico.Francesconi}@publications.europa.eu,
{Najeh.HAJLAOUI,
Brahim.BATOUCHE}@ext.publications.europa.eu
3
Institute of Legal Information Theory and Techniques (ITTIG),
Consiglio Nazionale delle Ricerche (CNR),
Via dei Barucci 20, 50127 Florence, Italy

Abstract. The result of the EU is a complex, multilingual, multicultural and yet

united environment, requiring solid integration policies and actions targeted at
simplifying cross-language and cross-cultural knowledge access. The legal
domain is a typical case in which both the linguistic and the conceptual aspects
mutually interweave into a knowledge barrier that is hard to break. In the context
of the ISA2 funded project “Public Multilingual Knowledge Infrastructure”
(PMKI) we are addressing Semantic Interoperability at both the conceptual and
lexical level, by developing a set of coordinated instruments for advanced
lexicalization of RDF resources (be them ontologies, thesauri and datasets in
general) and for alignment of their content. In this paper, we describe the
objectives of the project and the concrete actions, specifically in the legal
domain, that will create a platform for multilingual cross-jurisdiction accessi-
bility to legal content in the EU.

1 Introduction

The construction of the European Union is one of the most relevant political success
stories of the last decades, able to guarantee a space of freedom, justice and democracy
for millions of European citizens, based on the free exchange of people, information,
goods and services.
However, complex, multilingual and multicultural as Europe is, it cannot rely on
political success and good intentions alone: the objectives of its unification must be

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 257–271, 2018.
https://doi.org/10.1007/978-3-030-00178-0_17
258 A. Stellato et al.

underpinned by solid integration policies and targeted actions, considering and dealing
with the heterogeneities that lay at the basis of the foundation of EU itself.
The legal domain is an emblematic example of this heterogeneity: while united
under common goals and ethics, each of the Member States retains its own laws and
regulations. These need to be aligned to the common directions and indications pro-
vided by the EU Parliament, while keeping their independence and bindings to the
constitutions characterizing each nation. The differences are not technically limited to
the regulations per se, being the whole fabric of knowledge bond to the cultural and
societal heritage of a nation. For instance, the French concept “tribunaux adminis-
tratifs” cannot be translated in English as “administrative tribunals”. The English word
for “tribunaux” in fact is “courts” while the “administrative tribunals” are adminis-
trative commissions which are comparable, mutatis mutandis, to the French “autorités
administratives indépendantes” [1]. There is however, as this example shows, a lin-
guistic problem as well, as it is important that the reached semantic consensus on
recognized similarities and affinities be available and accessible in different languages.
In such a scenario, the European digital eco-system should be made ready to
support seamless and cross-lingual access to Member States’ legislations, accounting
for their differentia as well as their relatedness under the common umbrella of the EU.
With this objective to pursue, and in a broader context including, but not limited to,
the domain of jurisprudence and law, in 2010 the EU defined the so-called European
Interoperability Framework, namely a set of recommendations and guidelines to sup-
port the pan-European delivery of electronic government services. This framework
aims at facilitating public administrations, enterprises and citizens to interact across
borders, in a pan-European context. Such guidelines cover different aspects of social,
commercial and administrative relations among different European actors, like multi-
lingualism, accessibility, security, data protection, administrative simplification,
transparency, reusability of the solutions.
One of the main objectives of such guidelines is to establish semantic interoper-
ability between digital services, having the potential to overcome the barriers ham-
pering their effective cross-border exploitation, which means making information
exchange not only understandable by humans but also understandable and processable
by machines, as well as establishing correspondences between concepts in different
domains and languages, or represented in different digital tools (like controlled
vocabularies, classification schemas, thesauri).
In the context of the Public Multilingual Knowledge Infrastructure (PMKI), a
project funded by the ISA2 programme1 with the aim to overcome language barriers
within the EU by means of multilingual tools and services, we are addressing Semantic
Interoperability at both the conceptual and lexical level, by developing a set of coor-
dinated instruments for advanced lexicalization of RDF resources (be them ontologies,
thesauri and datasets in general) and for alignment of their content.
In this paper, we will show how the realization of such an objective will enable
seamless, multilingual, cross-legislative retrieval and analysis of legal content, and will
show how the PMKI project will contribute to such a vision by detailing it objectives

1
https://ec.europa.eu/isa2/
 Dataset Alignment and Lexicalization 259

and milestones. The rest of the paper is organized as follows: Sect. 2 provides more
motivations for our effort and describes use case scenarios. In Sect. 3 a brief overview
on the evolution of models for representing lexical resources is given. Section 4
introduces the PMKI project, while Sect. 5 details the actions of the project and their
outcomes in the legal domain. Section 6 concludes the paper.

2 Use-Case Scenarios

There are several scenarios in the management and access to legal content that would
benefit from a thorough approach to conceptual and lexical integration.

2.1 Semantic Integration

As shown in [2], a typical use case for the adoption of legal ontologies is their
application to specialist domains (e.g. industry standards), which in turn opens up
different interaction possibilities between the legal knowledge (e.g. norms and regu-
lations) and the one pertaining to the specialized domain (e.g. a domain ontology
related to the aforementioned industry standards). Alignments between legal ontologies
and specialized domain ontologies should be considered as precious resources per se,
as well as the systems and frameworks that support the creation of such artifacts.
Semantic integration between analogous legal knowledge resources developed in
different countries is also important, in order to facilitate the understanding of alien
concepts through those closer to one’s own culture or, at least, through general shared
conceptualizations. For instance, the Italian thesaurus TESEO2 (TEsauro SEnato per
l’Organizzazione dei documenti parlamentari: Senate Thesaurus for the organization of
parliamentary documents) is a classification system, originally developed by the Italian
Senate and now used in the most relevant databases of the Senate, Chamber of
Deputies and of some regions of Italy. Even in the multilingual environment charac-
terizing the EU, the monolingual TESEO (its concepts are expressed in Italian only)
keeps its relevance due to its tight connection to the Italian law regulation system and
culture (TESEO includes a mix of specific legal concepts and more general topics). It is
thus an irreplaceable resource for semantically indexing information from the above-
mentioned Italian data and document bases. At the same time, aligning TESEO to other
resources, such as the EU’s multilingual thesaurus EuroVoc3, the multilingual the-
saurus of the EU, allows for cross-cultural and cross-lingual (EuroVoc is available in
26 languages, chosen from those spoken by EU Member States and candidate coun-
tries) mediated access to any content indexed through it. While EuroVoc obviously
lacks the specificities that TESEO can offer to the Italian interested user, it still provides
a best-mediated access modality, universally accepted and officially adopted within
the EU.

2
https://www.senato.it/3235?testo_generico=745
3
http://eurovoc.europa.eu/
260 A. Stellato et al.

2.2 Natural Language Understanding

Another very important scenario (which, in turn, includes a plethora of use cases)
concerns the identification and extraction of relevant information. The identified
information can be exploited in a variety of tasks, such as:
Annotation of Document Corpora. In the legal domain law references, articulated in
proper structures (e.g. law, article, paragraph), are important indexing elements for
documents. Being able to search corpora by including explicit constraints on mentioned
regulations is a powerful feature for legal search engines. For instance, a lawyer could
access to the full list of trials registered for a given court, and extract all judgements
that include a mention of a given law (or a set of laws), eventually specifying portions
of them. However, discovering references is not trivial, as it implies both being able to
parse the structure of a law mention (a mix of natural language processing capabilities
and of background knowledge about the existing laws is required) and being able to
recognize the so called “popular expressions” for referring to these laws. For instance,
in Italy the expression “Bossi-Fini” is a specific term referring to the law n° 189 of 30th
July 2002, establishing policies about immigration and employment of migrants. This
popular expression is originated by the names of the first signatories, Gianfranco Fini
and Umberto Bossi, at that time vice-president of the cabinet and minister for insti-
tutional reforms and devolution respectively, and is often adopted even in specialized
literature. A proper lexicalization even – and actually, most importantly – in the same
language of the country adopting that law is thus important in order to recognize the
variety of expressions that refer to the same precise legal entity.
Knowledge Building. While the previous scenario deals with the identification of
mentions of entities already defined and structured in specific areas of knowledge, it is
also important to be able to build new knowledge by analyzing language content. The
development of specialized domain ontologies can be partially automated by applying
terminology extraction techniques to document corpora [3] in order to identify the
entities that will be later elaborated into ontology classes and properties. The analysis
of relations [4] in the text can help both the development of new knowledge as well as
– when legal content is available – the production of alignments between legal and
domain ontologies.
Cross-lingual Recognition. The availability of terms in multiple languages allows for
efficient retrieval of the same conceptual information in various languages. However,
the analysis of language content (for any of the two tasks above) requires more fine-
grained lexical background knowledge than just mere terminology. Being able to
describe, in different languages, the single components forming compound terms, the
several forms in which these can be declined/conjugated, their lexical variations etc. is
a necessary step which has to be carried on even for those languages different from the
one spoken in the country where that knowledge originated.
It appears evident how all the tasks above would benefit from proper lexicalization
of the knowledge resources involved, performed by adopting well established standards
for the representation of lexical information and of the lexical-semantic interface with
ontologies. Reuse of existing resources modeled according to these standards should be
 Dataset Alignment and Lexicalization 261

also encouraged, to minimize the effort for lexicalizing knowledge resources. In the
next section, we will provide an excursus over models for lexical resources that have
been proposed in the last 20 years of research on (computational) linguistics.

3 State of the Art on Linguistic Resources and Language

Representation

“The term linguistic resources refers to (usually large) sets of language data and
descriptions in machine readable form, to be used in building, improving, or evaluating
natural language (NL) and speech algorithms or systems” [5].
Multiple efforts have been spent in the past towards the achievement of consensus
among different theoretical perspectives and systems design approaches. The Text
Encoding Initiative (www.tei-c.org) and the LRE-EAGLES (Expert Advisory Group
on Linguistic Engineering Standards) project [6] are just a few, bearing the objective of
making possible the reuse of existing (partial) linguistic resources, promoting the
development of new linguistic resources for those languages and domains where they
are still not available, and creating a cooperative infrastructure to collect, maintain, and
disseminate linguistic resources on behalf of the research and development community.
A popular resource which got a broad diffusion characterized by exploitation in
both applications and scientific studies is WordNet [7, 8]. Being a structured lexical
database, presents a neat distinction between words, senses and glosses, and is char-
acterized by diverse semantic relations like hypernymy/hyponymy, antonymy etc.…
Though not being originally realized for computational uses, and being built upon a
model for the mental lexicon, WordNet has become a valuable resource in the human
language technology and artificial intelligence. Due to its vast coverage of English
words, WordNet provides general lexico-semantic information on which open-domain
text processing is based. Furthermore, the development of WordNets in several other
languages [9–11] extends this capability to trans-lingual applications, enabling text
mining across languages.
A more recent effort towards achieving a thorough model for the representation of
lexical resources is given by the Lexical Markup Framework [12]. LMF, which has
obtained ISO standardization (LMF; ISO 24613:2008), can represent monolingual,
bilingual or multilingual lexical resources. The same specifications are to be used for
both small and large lexicons, for both simple and complex lexicons, for both written
and spoken lexical representations. The descriptions range from morphology, syntax,
computational semantics to computer-assisted translation. The covered languages are
not restricted to European languages but cover all natural languages. The range of
targeted NLP applications is not restricted. LMF is able to represent most lexicons,
including the above mentioned WordNet.
With the advent of the Semantic Web and Linked Open Data, a number of models
have been proposed to enrich ontologies with information about how vocabulary
elements have to be expressed in different natural languages. These include the Lin-
guistic Watermark framework [13, 14], LexOnto [15], LingInfo [16], LIR [17],
262 A. Stellato et al.

Fig. 1. The OntoLex-lemon model

LexInfo [18] and, more recently, lemon [19]. The lemon model envisions an open
ecosystem in which ontologies4 and lexica for them co-exist, both of which are pub-
lished as data on the Web. It is in line with a many-to-many relation between: (i) on-
tologies and ontological vocabularies, (ii) lexicalization datasets and (iii) lexical
resources. Lexicalizations in our sense are reifications of the relationship between an
ontology reference and the lexical entries by which these can be expressed within
natural language. lemon foresees an ecosystem in which many independently published
lexicalizations and lexica for a given ontology co-exist.
In 2012, an important community effort has been made to provide a common model
for Ontology-Lexicon interfaces: the OntoLex W3C Community Group5 was started
with the goal of providing an agreed-upon standard by building on the aforementioned
models, the designers of which are all involved in the community group.
The OntoLex-lemon [20] model (see Fig. 1) developed by the OntoLex Commu-
nity Group is based on the original lemon model, which by now has been adopted by a
number of lexica [21–24], and as such was taken by the group as the basis for
developing an agreed-upon and widely accepted model. The lemon model is based onto
the idea of a separation between the lexical and the ontological layer following
Buitelaar [25] and Cimiano et al. [26], where the ontology describes the semantics of
the domain and the lexicon describes the morphology, syntax and pragmatics of the
words used to express the domain in a language. The model thus organizes the lexicon

4
It would be more appropriate to adopt the term “reference dataset” (including thus also SKOS
thesauri and datasets in general), to express data containing the logical symbols for describing a
certain domain. In line with the traditional name OntoLex (and thus the ontology-lexicon dualism),
we will however often refer to them with the term ontology.
5
http://www.w3.org/community/ontolex/
 Dataset Alignment and Lexicalization 263

primarily by means of lexical entries, which are a word, affix or multiword expression
with a single syntactic class (part-of-speech) to which a number of forms are attached,
such as for example the plural, and each form has a number of representations (string
forms), e.g. written or phonetic representation. Entries in a lexicon can be said to
denote an entity in an ontology, however normally the link between the lexical entry
and the ontology entity is realized by a lexical sense object where pragmatic infor-
mation such as domain or register of the connection may be recorded.
In addition to describing the meaning of a word by reference to the ontology, a
lexical entry may be associated with a lexical concept. Lexical concepts represent the
semantic pole of linguistic units, mentally instantiated abstractions which language
users derive from conceptions [27]. Lexical concepts are intended primarily to repre-
sent such abstractions when present in existing lexical resources, e.g. synsets for
wordnets.
Finally, linguists have acknowledged [28] the benefits that the adoption of the
Semantic Web technologies could bring to the publication and integration of language
resources, thus denoting a convergence of interests and results rarely occurring before.
A concrete outcome of this convergence is given by the Open Linguistics Working
Group6 of the Open Knowledge Foundation, which is contributing to the development
of a LOD (Linked Open Data) (sub)cloud of linguistic resources, known as LLOD7
(Linguistic Linked Open Data).

4 The PMKI Project

Public Multilingual Knowledge Infrastructure (PMKI) is launched as an ISA2 action to

answer claims from the European Language Technology Community such as the
multilingual extension of the Digital Single Market, the increase of the EU cross-border
online service. It aims to provide support for the EU economy in particular to SMEs to
overcome language barriers and to help to unlock the e-Commerce potential within the
EU implementing the necessary multilingual tools and features and helping to build the
Connecting Europe Facility Automated Translation (CEF.AT) Platform - a common
building block implemented through the CEF programme.
The project aims to create a set of tools and facilities, based on Semantic Web
technologies, aimed to support the language technology industry as well as public
administrations, with multilingual tools in order to improve cross border accessibility
of public administration services and e-commerce solutions. In practical terms, over-
coming language barriers on the Web means creating multilingual vocabularies and
language resources, establishing links between them as well as using them to support
accessibility to services and goods offered through the Internet.
The objective of PMKI is to implement a proof-of-concept infrastructure to expose
and to harmonize internal (European Union institutional) and external multilingual
resources aligning them in order to facilitate interoperability. It could support the

6
http://linguistics.okfn.org/
7
http://linguistic-lod.org/llod-cloud
264 A. Stellato et al.

knowledge layer of the multilingual infrastructure for Europe. Additionally, the project
aims to create a governance structure to extend systematically the infrastructure by the
integration of supplementary public multilingual taxonomies/terminologies.
PMKI is a pilot project to check the feasibility of the proposed solutions and to
prepare the roadmap to convert such proof-of-concept into a public service.

5 Specific Actions with Reusable Outcomes in the Legal

Domain

The proposed PMKI action meets the recommendations included in the European
Interoperability Strategy (EIS)8. The adherence to specific standards for describing
language resources, and the creation of an interoperability platform to manage them,
comply with the main approaches and “clusters” of the EIS (reusability of the solutions,
interoperability service architecture in the EU multilingual context, implication of ICT
on new EU legislation, as well as promotion of the awareness on the maturity level and
of the shareability of the public administration services).
Similarly, the proposal meets the recommendations and principles of the European
Interoperability Framework (EIF)9, regarding multilingualism, accessibility, adminis-
trative simplification, transparency, and reusability of the solutions. The creation of a
public multilingual knowledge infrastructure will allow EU public administrations to
create services that can be accessible and shareable independently from the language
actually used, as well as the SMEs to sell goods and service cross-border in a Digital
Single Market.
As we have shown in Sect. 2, the outcomes of such initiatives are prodrome for
supporting document analysis, indexing and retrieval as well as cross-legislation access
to legal content. In the next sections we will present the main actions foreseen in our
contribution to the PMKI project and their potential in supporting the above objectives.

5.1 Comparative Study and Selection of Semantic Web Standards

for Describing Multilingual Resources
A study has been conducted, embracing available web standards for multilingual
resources at large, thus including multilingualism in ontologies, terminologies and
specifically in lexical resources.
Due to the nature of the resources in PMKI, within the project different recom-
mendations or even popular vocabularies will be adopted:
SKOS [29]: the W3C recommendation for formalizing thesauri, terminologies, con-
trolled vocabularies and other knowledge resources characterized by shallow seman-
tics. It is worth noticing that the terminological level solely supports the identification
of concepts by giving them names (and alternative lexical references) but cannot be
considered to be a lexicon nor any sort of advanced lexical resource, as any sort of

8
http://ec.europa.eu/isa/documents/isa_annex_i_eis_en.pdf
9
http://ec.europa.eu/isa/documents/isa_annex_ii_eif_en.pdf
 Dataset Alignment and Lexicalization 265

lexical description taking into account phenomena such as morphology, lexical rela-
tions etc. are considered, by definition, to be out of the scope of a thesaurus.
SKOS-XL [30]: As for the general definition of thesauri, SKOS does not address
complex lexical descriptions of its elements. However, SKOS is extended by the
SKOS-XL vocabulary which provides reified labels by means of the class skosxl:Label.
SKOS terminological properties have their equivalents (identified by homonymous
local names) in the new namespace, that is: skosxl:prefLabel, skosxl:hiddenLabel,
skosxl:altLabel in order to relate concepts with these reified labels.
OntoLex-Lemon. We already described this model in Sect. 3. As the model is rela-
tively recent, there is still not much support for developing resources according to its
vocabulary. As explained in the next section, we will develop a system, integrated into
an already mature ontology/thesauri development environment, for the development of
lexicons and for interfacing lexical knowledge with ontological one.

Lexical Lexical
Resource Resource

Conceptual
Resource

Conceptual Alignment ValidaƟon Tools

LexicalizaƟon Tools Resource
LexicalizaƟon Process

Lexical Lexical Conceptual

Resource Resource Resource
IntegraƟon
Lexical
Lexical Services
Resource Lexical
Resource
Resource Lexical
Resource

Fig. 2. PMKI integration framework, general architecture

Other Models and Schemes. The above vocabularies represent the core of the
selected models for development and alignment of resources in PMKI. The support for
OntoLex will however not be limited to the enrichment of SKOS thesauri, and OWL
ontologies or generic RDF datasets can be lexically enriched with OntoLex lexical
descriptions with no loss of generality. Similarly, the above choices do not obviously
prevent the adoption of specific metadata vocabularies, domain/application ontologies
etc.…

5.2 Systems for Semantic and Lexical Integration of Multilingual

Resources
Support for integration will be implemented two-fold: by realizing a framework for
alignment of semantic resources (thesauri, ontologies etc.) and by the development of a
system for the development of lexicons according to the OntoLex vocabulary and for
the lexical enrichment of semantic resources with lexical information.
266 A. Stellato et al.

Even though a pilot project in nature, PMKI is not a research project, it in fact aims
at building up on well-established research results and existing technologies and at
converging towards a concrete proposal for an integration framework.
The general concept behind the framework is depicted in Fig. 2, focused on the
interaction of systems aimed at supporting the two tasks previously defined.
Semantic Integration Framework. The architecture foresees the presence of
semantic integration services accessed by RDF management systems. The separation
between the two is dictated by the different requirements in terms of interaction
modalities, performance and results. RDF Management Services, whether single-user
desktop applications or centralized collaborative platform, require high interaction with
the user, averagely-low response times and, in the case of collaborative systems, the
capacity to serve in real time several users accessing diverse projects. These platforms
may offer manual or semi-automatic alignment functionalities, which though have to be
performed with a low impact on system resource, and possibly replicated across several
parallel requests. Conversely, Semantic Integration systems may instead act as token-
based service providers, receiving requests to load and align datasets of considerable
size, performing their function in non-trivial execution time due to the intensive
analysis of the involved resources and dedicating considerable amount of resources to
these tasks. After each alignment process has been completed, the alignment services
may release the token to the requesting peer and start the next alignment task at the
head of the request queue. A pool of processors may be considered in order to allow
parallelization of alignment tasks.
The Semantic Integration System developed within the pilot project will be based
on GENOMA [31], a, highly configurable alignment architecture, and on MAPLE [32],
a metadata-driven component for the configuration of mediators, which will allow for
seamless application of the same alignment techniques on datasets modeled according
to different modeling vocabularies, by providing vocabulary-specific implementations
of the general analysis engine tasks. The manual/semi-automatic alignment capabilities
will be provided by VocBench [33], a collaborative RDF management system for the
development and maintenance of RDF ontologies thesauri and datasets, based on a
service-oriented RDF management platform [34], recently updated to its third version
[35] through another funded action of the ISA2 program. As part of a coordinated
action with the PMKI project, VocBench will also feature interaction modalities with
the Semantic Integration system developed within PMKI.
Lexicon Development and Lexical Enrichment of Knowledge Resources. The
OntoLex model is relatively young and, as such, it is still not widely supported by most
mature technologies for data management. In a recent paper [36] describing the
expressive power of VocBench 3 custom forms, the authors show how the custom form
mechanism could be used to define complex lemon-patterns. As VocBench 3 provides
a general-purpose editing environment with specific facilities for the editing of SKOS
and SKOSXL thesauri and OWL ontologies, extending the system with dedicated
support for OntoLex-Lemon seems thus a natural way to cover this need.
In PMKI, VocBench will thus be improved to support the OntoLex-Lemon model in
two different scenarios: developing Lexicons based on the OntoLex vocabularies and
enriching semantic resources with lexical content. The two scenarios may be
 Dataset Alignment and Lexicalization 267

interwoven, as it will be possible to develop lexical entries specifically for semantic

resources as well as reuse lexical content from existing lexicons in order to enrich the
semantic resource with it. In most real applications, the two possibilities will not be
alternative to each other: while a lexicon can usually provide domain-independent
lexical entries, the description of specific concepts in a domain/application ontology
often requires the definition of new complex terms, thus requiring in turn to state how
the proper combination (at the lexical level) of single lexical entries would generate the
accurate description of the conceptual elements. This implies the creation of further
lexical entries describing the multiword, syntagmatic structure of the lexical represen-
tation of the complex concept. The inspiring work for such evolution of VocBench
comes from past works concerning semi-automatic enrichment of ontologies by reuse of
lexical resources [37] and exploitation of language metadata [13, 38], which can now
benefit from the standardization of this metadata for the Linked Open Data [39, 40].

Fig. 3. Aligning concepts between EU EuroVoc and Italian Senate’s TESEO thesauri

5.3 Realization of Concrete Semantic Alignments and Lexical

Enrichments and Assessing
In the pilot project, a certain amount of alignments and lexicalizations will be pro-
duced. The objective is not only to produce the resources per-se, but to provide golden-
standards that can be used in the later stage to evaluate the alignment systems.
Developing a gold standard mapping dataset is however not an easy task, due to the
difficulty, even for humans, to lose potential matches in datasets of even modest size.
Not incidentally, the Ontology Alignment Evaluation Initiative (OAEI), an initiative,
implemented as a contest, which aims at evaluating the state of the art of ontology
alignment tools, [41] mostly offers test beds of relatively small size and rarely updates
the list of these mappings. Furthermore, evaluation of mappings sent by the partici-
pants, in the cases involving large datasets does not rely entirely on the standard and
involves instead manual scrutiny, in order to take into account potentially correct
mappings missing from the oracle.
268 A. Stellato et al.

We thus decided to divide the kind of contributions for dataset alignment in two
steps: a vertical exploration of a humanly-computable subdomain of the two thesauri,
and a larger attempt at mapping complete resources. The first result guarantees the
creation of a reliably sound and complete set of mappings, while a larger alignment on
the whole resources will be produced later on in the project, by means of semi-
automatic processes, reusing the same systems that we will validate through the first
result. An alignment that will be considered for production is the one – already
mentioned in the example in Sect. 2.1 – between EuroVoc and TESEO (see Fig. 3).
Concerning lexicalizations, EuroVoc, as a central hub in the EU scenario, has, also
in this case, been selected as the target conceptual resource. Candidate lexical resources
for the results to be produced within the pilot are WordNet, being probably the most
popular lexical resource and, for analogous reasons and more specifically in the EU
scenario, IATE10, the InterActive Terminology for Europe. However, both these
resources do not provide the rich lexical and morphological descriptions representing
the added value brought by the OntoLex model. For this reason, other resources such as
BabelNet, or other lexicons still not modeled in OntoLex, will be taken into consid-
eration. In the latter case the process will be two-fold: porting resources to OntoLex,
which is a result per se, and then using them to lexicalize (part of) EuroVoc.

6 Conclusions

In this paper, we have presented the objectives and roadmap of the PMKI project and
how its outcomes will directly and positively affect access to legal content and foster its
exploitation in various scenarios. The multicultural, multi-jurisdictional and multilin-
guistic nature of the European Union has always been considered an asset rather than
an obstacle, as it is through their differences that the Member States can learn from each
other, benefiting from distinct experiences and approaches. Making these experiences
truly and effectively comparable by lowering the language barriers and by
harmonizing/connecting different though overlapping concepts and regulations is the
objective of initiatives such as PMKI. Even though a pilot project, PMKI will aim to
pave the way for analogous efforts while still contributing the community with tangible
results in terms of systems and frameworks for alignment and lexicalization of
heterogeneous resources.

References
1. Francesconi, E., Peruginelli, G.: Opening the legal literature portal to multi-lingual access.
In: Proceedings of the Dublin Core Conference, pp. 37–44 (2004)
2. Antonini, A., Boella, G., Hulstijn, J., Humphreys, L.: Requirements of legal knowledge
management systems to aid normative reasoning in specialist domains. In: Nakano, Y.,
Satoh, K., Bekki, D. (eds.) JSAI-isAI 2013. LNCS (LNAI), vol. 8417, pp. 167–182.
Springer, Cham (2014). https://doi.org/10.1007/978-3-319-10061-6_12

10
http://termcoord.eu/iate/
 Dataset Alignment and Lexicalization 269

3. Velardi, P., Navigli, R., Cucchiarelli, A., Neri, F.: Evaluation of ontolearn, a methodology
for automatic population of domain ontologies. In: Ontology Learning from Text: Methods,
Applications and Evaluation. IOS Press, Amsterdam (2005)
4. Pennacchiotti, M., Pantel, P.: Automatically harvesting and ontologizing semantic relations.
In: Buitelaar, P., Cimiano, P. (eds.) Ontology learning and population: bridging the gap
between text and knowledge. Frontiers in Artificial Intelligence. IOS Press, Amsterdam
(2008)
5. Cole, R.A., Mariani, J., Uszkoreit, H., Zaenen, A., Zue, V. (eds.): Survey of the State of the
Art in Human Language Technology. Cambridge University Press, Cambridge (1997)
6. Calzolari, N., McNaught, J., Zampolli, A.: EAGLES Final Report: EAGLES Editors
Introduction. Pisa, Italy (1996)
7. Miller, G.A., Beckwith, R., Fellbaum, C., Gross, D., Miller, K.: Introduction to WordNet:
An On-line Lexical Database (1993)
8. Fellbaum, C.: WordNet: An Electronic Lexical Database. WordNet Pointers. MIT Press,
Cambridge, MA (1998)
9. Vossen, P.: EuroWordNet: A Multilingual Database with Lexical Semantic Networks.
Kluwer Academic Publishers, Dordrecht (1998)
10. Roventini, A., et al.: ItalWordNet: a large semantic database for the automatic treatment of
the Italian language. In: First International WordNet Conference, Mysore, India, January
2002
11. Stamou, S., et al.: BALKANET: a multilingual semantic network for the Balkan languages.
In: First International Wordnet Conference, Mysore, India, pp. 12–14 (2002)
12. Francopoulo, G., et al.: Lexical markup framework (LMF). In: LREC2006, Genoa, Italy
(2006)
13. Pazienza, M.T., Stellato, A., Turbati, A.: Linguistic Watermark 3.0: an RDF framework and a
software library for bridging language and ontologies in the semantic web. In: 5th Workshop
on Semantic Web Applications and Perspectives (SWAP2008), Rome, Italy, 15–17
December 2008, CEUR Workshop Proceedings, FAO-UN, Rome, Italy, vol. 426, p. 11
(2008)
14. Oltramari, A., Stellato, A.: Enriching ontologies with linguistic content: an evaluation
framework. In: The Role of Ontolex Resources in Building the Infrastructure of Web 3.0:
Vision and Practice (OntoLex 2008), 31 May, Marrakech, Morocco, pp. 1–8 (2008)
15. Cimiano, P., Haase, P., Herold, M., Mantel, M., Buitelaar, P.: LexOnto: a model for
ontology lexicons for ontology-based NLP. In: Proceedings of the OntoLex07 Workshop
(held in conjunction with ISWC 2007) (2007)
16. Buitelaar, P., et al.: LingInfo: design and applications of a model for the integration of
linguistic information in ontologies. In: OntoLex 2006, Genoa, Italy, pp. 28–34 (2006)
17. Montiel-Ponsoda, E., Aguado de Cea, G., Gómez-Pérez, A., Peters, W.: Enriching
ontologies with multilingual information. Nat. Lang. Eng. 17, 283–309 (2011)
18. Cimiano, P., Buitelaar, P., McCrae, J., Sintek, M.: LexInfo: a declarative model for the
lexicon-ontology interface. Web Semant. Sci. Serv. Agents World Wide Web 9(1), 29–51
(2011)
19. McCrae, J., et al.: Interchanging lexical resources on the Semantic Web. Lang. Resour. Eval.
46(4), 701–719 (2012)
20. Cimiano, P., McCrae, J.P., Buitelaar, P.: Lexicon Model for Ontologies: Community Report,
10 May 2016. Community Report, W3C (2016). https://www.w3.org/2016/05/ontolex/
21. Borin, L., Dannélls, D., Forsberg, M., McCrae, J.P.: Representing Swedish lexical resources
in RDF with lemon. In: Proceedings of the ISWC 2014 Posters & Demonstrations Track a
Track Within the 13th International Semantic Web Conference (ISWC 2014), Riva del
Garda, Italy, pp. 329–332 (2014)
270 A. Stellato et al.

22. Ehrmann, M., Cecconi, F., Vannella, D., McCrae, J.P., Cimiano, P., Navigli, R.:
Representing multilingual data as linked data: the case of BabelNet 2.0. In: Proceedings
of the Ninth International Conference on Language Resources and Evaluation (LREC-2014),
Reykjavik, Iceland, 26–31 May 2014, pp. 401–408 (2014)
23. Eckle-Kohler, J., McCrae, J.P., Chiarcos, C.: lemonUby—a large, interlinked syntactically-
rich lexical resources for ontologies. Semant. Web J. (2015 accepted)
24. Sérasset, G.: Dbnary: wiktionary as a LMF based multilingual RDF network. In:
Proceedings of the Eighth International Conference on Language Resources and Evaluation
(LREC-2012), Istanbul, Turkey, 23–25 May 2012, pp. 2466–2472 (2012)
25. Buitelaar, P.: Ontology-based Semantic Lexicons: Mapping between Terms and Object
Descriptions. In: Huang, C.-R., Calzolari, N., Gangemi, A., Lenci, A., Oltramari, A., Prevot,
L. (eds.) Ontology and the Lexicon: A Natural Language Processing Perspective. Cambridge
University Press, Cambridge (2010)
26. Cimiano, P., McCrae, J., Buitelaar, P., Montiel-Ponsoda, E.: On the role of senses in the
ontology-Lexicon. In: Oltramari, A., Vossen, P., Qin, L., Hovy, E. (eds.) New Trends of
Research in Ontologies and Lexical Resources, pp. 43–62. Springer, Berlin (2013). https://
doi.org/10.1007/978-3-642-31782-8_4
27. Evans, V.: Lexical concepts, cognitive models and meaning-construction. Cognit. Linguist.
17(4), 491–534 (2006)
28. Chiarcos, C., McCrae, J., Cimiano, P., Fellbaum, C.: Towards open data for linguistics:
linguistic linked data. In: Oltramari, A., Vossen, P., Qin, L., Hovy, E. (eds.) New Trends of
Research in Ontologies and Lexical Resources, pp. 7–25. Springer, Berlin (2013). https://
doi.org/10.1007/978-3-642-31782-8_2
29. World Wide Web Consortium (W3C): SKOS Simple knowledge organization system
reference. In: World Wide Web Consortium (W3C) (2009). http://www.w3.org/TR/skos-
reference/. Accessed 18 Aug 2009
30. World Wide Web Consortium (W3C): SKOS simple knowledge organization system
eXtension for labels (SKOS-XL). In: World Wide Web Consortium (W3C). http://www.w3.
org/TR/skos-reference/skos-xl.html. Accessed 18 Aug 2009
31. Enea, R., Pazienza, M.T., Turbati, A.: GENOMA: GENeric Ontology Matching Architec-
ture. In: Gavanelli, M., Lamma, E., Riguzzi, F. (eds.) AI*IA 2015. LNCS (LNAI), vol. 9336,
pp. 303–315. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-24309-2_23
32. Fiorelli, M., Pazienza, M.T., Stellato, A.: A meta-data driven platform for semi-automatic
configuration of ontology mediators. In Chair, N.C., et al. (eds.) Proceedings of the Ninth
International Conference on Language Resources and Evaluation (LREC’14), May 2014.
European Language Resources Association (ELRA), Reykjavik, Iceland, pp. 4178–4183
(2014)
33. Stellato, A., Rajbhandari, S., Turbati, A., Fiorelli, M., Caracciolo, C., Lorenzetti, T., Keizer,
J., Pazienza, M.T.: VocBench: a web application for collaborative development of
multilingual thesauri. In: Gandon, F., Sabou, M., Sack, H., d’Amato, C., Cudré-Mauroux,
P., Zimmermann, A. (eds.) ESWC 2015. LNCS, vol. 9088, pp. 38–53. Springer, Cham
(2015). https://doi.org/10.1007/978-3-319-18818-8_3
34. Pazienza, M.T., Scarpato, N., Stellato, A., Turbati, A.: Semantic Turkey: a browser-
integrated environment for knowledge acquisition and management. Semant. Web J. 3(3),
279–292 (2012)
35. Stellato, A., et al.: Towards VocBench 3: pushing collaborative development of thesauri and
ontologies further beyond. In: 17th European Networked Knowledge Organization Systems
(NKOS) Workshop, 21st September 2017, Thessaloniki, Greece (2017)
 Dataset Alignment and Lexicalization 271

36. Fiorelli, M., Lorenzetti, T., Pazienza, M.T., Stellato, A.: Assessing VocBench custom forms
in supporting editing of lemon datasets. In: Gracia, J., Bond, F., McCrae, John P., Buitelaar,
P., Chiarcos, C., Hellmann, S. (eds.) LDK 2017. LNCS (LNAI), vol. 10318, pp. 237–252.
Springer, Cham (2017). https://doi.org/10.1007/978-3-319-59888-8_21
37. Pazienza, M.T., Stellato, A.: An environment for semi-automatic annotation of ontological
knowledge with linguistic content. In: Sure, Y., Domingue, J. (eds.) ESWC 2006. LNCS,
vol. 4011, pp. 442–456. Springer, Heidelberg (2006). https://doi.org/10.1007/11762256_33
38. Pazienza, M.T., Sguera, S., Stellato, A.: Let’s talk about our “being”: a linguistic-based
ontology framework for coordinating agents. Appl. Ontol. Spec. Issue Form. Ontol.
Commun. Agents 2(3–4), 305–332 (2007)
39. Fiorelli, M., Stellato, A., McCrae, J.P., Cimiano, P., Pazienza, M.T.: LIME: the metadata
module for OntoLex. In: Gandon, F., Sabou, M., Sack, H., d’Amato, C., Cudré-Mauroux, P.,
Zimmermann, A. (eds.) ESWC 2015. LNCS, vol. 9088, pp. 321–336. Springer, Cham
(2015). https://doi.org/10.1007/978-3-319-18818-8_20
40. Fiorelli, M., Pazienza, M.T., Stellato, A.: An API for OntoLex LIME datasets. In: OntoLex-
2017 1st Workshop on the OntoLex Model (co-located with LDK-2017), Galway (2017)
41. Shvaiko, P., Euzenat, J.: Ontology matching: state of the art and future challenges. IEEE
Trans. Knowl. Data Eng. 25(1), 158–176 (2013)
 A Multilingual Access Module to Legal Texts

Kiril Simov1(&), Petya Osenova1, Iliana Simova1,

Hristo Konstantinov2(&), and Tenyo Tyankov2
1
Linguistic Modelling Department, IICT-BAS, Sofia, Bulgaria
{kivs,petya,iliana}@bultreebank.org
2
APIS, Sofia, Bulgaria
{ico,tencho}@apis.bg

Abstract. The paper introduces a Multilingual Access Module. This module

translates the user’s legislation query from its source language into the target
language, and retrieves the detected texts that match the query. The service is
demonstrated in its potential for two languages – English and Bulgarian, in both
directions (English-to-Bulgarian and Bulgarian-to-English). The module con-
sists of two submodules: Ontology-based and Statistical Machine Translation.
Since both proposed submodules have some drawbacks, they are used in an
integrated architecture, thus profiting from each other.

Keywords: Multilingual access
Query translation
Query expansion

1 Introduction

Processing of legal data has been in the focus of NLP community for the recent years.
A lot of resources have been compiled and made available to the community. This
holds especially for the legal documentation of EU (EuroParl1, JRC Acquis2, EAC-
ECDC3, etc.). Rich thesauri and ontologies have been produced (such as, Eurovoc4,
lkif-core5). At the same time, processing modules have been developed for the aims of
information retrieval. There are also a number of projects dedicated to the smart access
to legal data. These are OpenLaws6, EUCases7, e-CODEX8, etc.; focused conferences,
such as JURIX9, among others. Our paper presents a multilingual module for accessing
legal texts. It uses two approaches for the translation of users’ queries – ontology-based
and statistical. The experiments showed that the results are better when both compo-
nents are used in combination. Although the multilingual nature of the module is
demonstrated for two languages – English and Bulgarian, it is scalable also to other
languages.
1
http://www.statmt.org/europarl/.
2
http://optima.jrc.it/Acquis/index_2.2.html.
3
https://ec.europa.eu/jrc/en/language-technologies/ecdc-translation-memory.
4
http://open-data.europa.eu/data/dataset/eurovoc.
5
https://github.com/RinkeHoekstra/lkif-core.
6
http://www.openlaws.eu/?page_id=1004.
7
http://www.eucases.eu/start/.
8
http://www.e-codex.eu/about-the-project.html.
9
http://jurix2015.di.uminho.pt/.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 272–286, 2018.
https://doi.org/10.1007/978-3-030-00178-0_18
 A Multilingual Access Module to Legal Texts 273

The structure of the paper is as follows: in Sect. 2 the main background issues are
outlined. Section 3 focuses on the Ontology-based Translation module. Section 4
discusses Statistical Machine Translation Module. Section 5 presents the parameters of
the integration between the two modules and describes the evaluation. Section 6
concludes the paper.

2 Background

In this paper we report on the design and development of Multilingual Access Module
(MLAM) for full text search within legal documents in English and Bulgarian.
The MLAM translates user queries in both directions: from Bulgarian to English and
from English to Bulgarian. A typical user query is a list of key words and phrases. The
user query is evaluated over a set of documents loaded in a full text search engine
which performs searches for relevant documents. Thus, our goal is to deliver an ade-
quate translation service for the user queries.
In the module two complementary technologies are exploited. The first technology
is based on Ontology-to-Text relation. In this case, the system relies on a common
ontology with augmented lexicons. The lexicons are mapped in such a way that the
conceptual information within the ontology corresponds to the meaning of the lexical
items. Having lexicons for different languages aligned to the same ontology is a pre-
requisite for the accurate translation of the corresponding lexical items. In addition to
the lexicons, special chunk grammars are needed to recognize the lexical items in the
text. Such grammars are important especially for languages with rich morphology
and/or free word order.
The exploitation of ontologies in translation provides some additional functionality
of performing query expansion on the basis of inference within the ontology. In our
module we implemented two query expansion procedures: (1) expansion via subclasses
and (2) expansion via related classes. Both of them are presented in our ontologies:
Syllabus ontology and Eurovoc multilingual taxonomy. After performing the query
expansion, the new set of classes is translated to the target language using the
appropriate lexicons.
We expect that the user queries will be mainly related to the above mentioned
domain ontologies which are also employed for document indexing. They reflect the
specific content of the documents in the EUCases database. Nevertheless, the users
should not be restricted to specify their queries only through the lexical items from the
available lexicons. Thus, MLAM needs to provide translation also for words and
phrases that are not in the lexicons. In order to solve this problem, we exploited a
statistical machine translation module trained on domain specific parallel corpora. This
module in combination with the ontology-based module provides alternative transla-
tions to the lexicon items and thus covers the missing translations for out-of-vocabulary
words and phrases.
274 K. Simov et al.

3 The Ontology-Based Translation Module

The design and implementation of the Ontology-based translation module of MLAM

exploits the ontology-to-text relation approach [7, 8]. We started with classification of
ontologies with respect to their precision provided by Nicola Guarino10:
Lexicon: machine readable dictionaries; vocabulary with natural language
definitions.
Simple Taxonomy: classifications.
Thesaurus: WordNet; taxonomy plus related terms.
Relational Model: Light-weight ontologies; unconstrained use of arbitrary
relations.
Fully Axiomatized Theory: Heavy-weight ontologies.
The two ontologies to be used in the project – Syllabus ontology and Eurovoc
multilingual taxonomy – are at the middle of this classification. The first one is a
relational model by its creation, and the second one is a thesaurus. Thus, they provide a
limited inference which we exploit for query expansion. Both of them are considered
domain ontologies, but are not aligned to an upper ontology.

3.1 Ontology-to-Text Relation

Here we represent briefly the two main components that define the ontology-to-text
relation necessary to support the tasks within our project. These components are:
(terminological) lexicon and a concept annotation grammar.
The lexicon plays a twofold role in our architecture. First, it interrelates the con-
cepts in the ontology to the lexical knowledge used by the grammar for recognizing the
role of the concepts in the text. Second, lexicon represents the main interface between
the user and the ontology. This interface allows for the ontology to be navigated or
represented in a natural for the user way. For example, the concepts and relations might
be named with terms used by the users in their everyday activities and in their own
natural language. This might be considered as a first step to a contextualized usage of
the ontology in a sense that the ontology might be viewed through different terms
depending on the context.
Thus, the lexical items contain the following information: a term, information
determining the context of the term usage, and grammatical features providing the
syntactic realization within the text. In the current implementation of the lexicons the
contextual information is simplified to a list of a few types of users (lawyers, judges,
etc.). With respect to the relations between the terms in the lexicon and the concepts in
the ontology, there are two main problems: (1) there is no lexicalized term for some of
the concepts in the ontology, and (2) there are lexical terms in the domain language
which lack corresponding concepts in the ontology.

10
Invited Mini-course on Ontological Analysis and Ontology Design. First Workshop on Ontologies
and lexical Knowledge Bases - OntoLex 2000. Sozopol, Bulgaria.
 A Multilingual Access Module to Legal Texts 275

The first problem is solved by adding also non-lexicalized (fully compositional)

phrases to the lexicon. These varieties of phrases or terms for a given concept are used
as a basis for construction of the annotation grammar. Having them, we would capture
different wordings of the same meaning in the text. The picture below shows the
mapping varieties. It depicts the realization of the concepts (similarly for relations and
instances) in the language. The concepts are language independent and they might be
represented within a natural language as form(s) of a lexicalized term, or as a free
phrase. In general, a concept might have a few terms connected to it and a (potentially)
unlimited number of free phrases expressing this concept in the language. Some of the
free phrases receive their meaning compositionally regardless their usage in the text,
other free phrases denote the corresponding concept only in a particular context. In our
lexicons we decided to register as many free phrases as possible in order to have better
recall on the semantic annotation task. In case of a concept that is not-lexicalized in a
given language we require at least one free phrase to be provided for this concept. See
Fig. 1.

Lexicalized Free Phrases

Ontology
Terms

Fig. 1. Ontology-to-text relation

We could summarize the connection between the ontology and the lexicons in the
following way: the ontology represents the semantic knowledge in form of concepts
and relations with appropriate axioms; and the lexicons represent the ways in which
these concepts can be realized in texts in the corresponding languages. Of course, the
ways in which a concept could be represented in the text are potentially infinite in
number. For that reason we aimed at representing in our lexicons only the most
frequent and important terms and phrases.
276 K. Simov et al.

The second component of the ontology-to-text relation, the concept annotation

grammar, is ideally considered as an extension of a general language deep grammar
which is adopted to the concept annotation task. Minimally, the concept annotation
grammar consists of a chunk grammar for concept annotation and (sense) disam-
biguation rules. The chunk grammar for each term in the lexicon contains at least one
grammar rule for recognition of the term. As a preprocessing step we consider anno-
tation with grammatical features and lemmatization of the text. The disambiguation
rules exploit the local context in terms of grammatical features, semantic annotation
and syntactic structure, and also the global context, such as the topic of the text,
discourse segmentation, etc.
For the implementation of the annotation grammar we rely on the grammar facil-
ities of the CLaRK System (Simov et al. 2001).
The creation of the actual annotation grammars started with the terms in the lexi-
cons for the corresponding languages. Each term was lemmatized and the lemmatized
form of the term was converted into a regular expression of grammar rules. Each
concept related to the term is stored in the return markup of the corresponding rule.
Thus, if a term is ambiguous, then the corresponding rule in the grammar contains a
reference to all the concepts related to the term.
Figure 2 depicts the relations between lexical items, grammar rules and the text.
We exploit these relations for multilingual access.
The relations between the different elements of the models are as follows. A lexical
item could have more than one grammar rule associated to it depending on the word
order and the grammatical realization of the lexical item. Two lexical items might share
a grammar rule if they have the same wording, but relate to different concepts in the
ontology. Each grammar rule recognizes zero or several text chunks.
The ontology-to-text relation, implemented in this way, provides facilities for
solving different tasks, such as ontology-based search (including crosslingual search).

3.2 Implementation of the Ontology-Based Translation Module

Our first implementation of the ontology-based translation module used the EuroVoc11
multilingual taxonomy as its main ontology. We consider EuroVoc as a light-weight
ontology. EuroVoc covers the concepts related to the European Union activities. In this
respect we consider it as a domain ontology for the main topics of interest to the users
of EUCases project services. The lexicons were aligned within EuroVoc for the official
languages of European Union and also for some other languages. In this respect, it
provides the necessary inventory for the two languages that are important to us: Bul-
garian and English. It can be extended further for the other represented languages. The
actual concepts are mapped via numerical identifiers. The concepts are arranged in
domains and microthesauri. Each domain is divided into a number of microthesauri.
A microthesaurus is considered as a concept scheme with a subset of the concepts that
are part of the complete EuroVoc thesaurus. The main relations encoded in EuroVoc

11
http://eurovoc.europa.eu/.
 A Multilingual Access Module to Legal Texts 277

that we exploit are: “skos:broader”12, “skos:related” and “xl:pre-

fLabel”, defined in http://www.w3.org/2009/08/skos-reference/skos.html and http://
www.w3.org/TR/skos-reference/skos-xl.html documents. Table 1 contains a few
examples of the interrelations of concepts in EuroVoc. With the help these relations we
define different query expansion approaches.

Lexical Items Grammar Domain Texts

Rules

Fig. 2. Relations between lexical items, grammar rules and text

As it was mentioned above, on the basis of all terms related to a given concept
identifier, a regular expression is created. Each rule annotated the recognized text with
the corresponding identifier. In the cases of ambiguous terms, the text was annotated
with several identifiers. The annotation grammars were applied over the user query
string and each recognized term was presented by the corresponding concept identi-
fiers. After the complete analysis of the input query, the text was converted into a list of
identifiers. In some cases, a specific substring of the user query might not be recognized
as a term in the lexicon. In this case, the substring remained unanalyzed. For example,
for the concept 146013 the grammar rule in the CLaRK system would look like:
<“EU”>, <“financial”>, <“instrument”> ! <concept v = ”1460”/>, where the first part
of the rule would recognize the term in the text and this text would be annotated with
the XML fragment from the right part of the rule.

12
Also its reverse relation “skos:narrower”.
13
See the first row in Table 1.
278 K. Simov et al.

Table 1. In the table the relations between concepts in EuroVoc thesaurus are presented. The
relations show the complexity of the graph in the thesaurus.
Concept Bulgarian term English term skos:narrower skos:
ID related
1460 Финaнcoв EU financial 1052, 2054, 2609 862, 1851,
инcтpyмeнт нa instrument 2511, 4370,
Oбщнocттa 5472
1052 Фoндoвe нa EC EC funds 5138, 5643, 978 973, 4055,
8549, 862
5138 Cтpyктypни Structural 1056, 4056, 5668 5472, 5499,
фoндoвe funds 5580, 5847
862 Пoмoщ нa EU aid 852 –
Oбщнocттa
5499 Икoнoмичecкo и Economic 5864 5643
coциaлнo and social
взaимoдeйcтвиe cohesion
5643 Фoнд зa cближaвaнe Cohesion – –
Fund

After the user query was executed, the text was annotated with concept identifiers.
Then we performed query expansion on the basis of the ontology. In this case, we
exploited the two relations that define the structure of the ontology: “skos:narrower”
and “skos:related”. As the table shows, both relations “skos:narrower” and
“skos:related” are transitive. These relations can be used for adding new concept
identifiers to those from the annotation of the user query. The relations can be used in two
ways: (1) getting only the directly related concepts, or (2) getting the concepts that are
related via transitive closure of the relations.
In the first implementation, we performed a transitive closure for the relation
skos:broader and only direct related concepts for the relation skos:related.
Here we present the processing steps of the user query: EU financial instrument:
Step 1: Text annotation
Step 2: Query expansion applying transitive closure of skos:narrower
Step 3: Query expansion applying skos:related
Step 4: Deleting the repeated concepts
Step 5: Translation to the other language
In this step each concept identifier is substituted with the corresponding terms in the
other language. The result for our example includes phrases in Bulgarian:
финaнcoв инcтpyмeнт нa EC пoмoщ нa EC пoддъpжaщ мexaнизъм зeмeдeлcкa вaлyтнa
пoлитикa eвpoпeйcкa пapичнa cиcтeмa paмкa зa пoдкpeпa нa oбщнocттa фoндoвe
(EC) eвpoпeйcки фoнд зa вaлyтнo cътpyдничecтвo eвpoпeйcки фoнд зa paзвитиe eвpo-
пeйcки фoнд зa пpиcпocoбявaнe към глoбaлизaциятa cтpyктypни фoндoвe икoнoмичecкo и
coциaлнo взaимoдeйcтвиe cтpyктypeн paзxoд пoдxoдящ paйoн зa paзвитиe eвpoпeйcки
фoнд зa peгиoнaлнo paзвитиe peгиoнaлнa пoмoщ peгиoнaлнo плaниpaнe peгиoнaлнa
пoлитикa нa EC cтpyктypнo пpиcпocoбявaнe eвpoпeйcки coциaлeн фoнд фиop eвpoпeйcки
 A Multilingual Access Module to Legal Texts 279

фoнд зa opиeнтиpaнe и гapaнтиpaнe нa зeмeдeлиeтo ceкция „opиeнтиpaнe“ фoнд зa

cближaвaнe ФEOГA eвpoпeйcки фoнд зa opиeнтиpaнe и гapaнтиpaнe нa зeмeдeлиeтo –
ceкция „гapaнтиpaнe“ нoв инcтpyмeнт нa oбщнocттa eвpoпeйcкa инвecтициoннa бaнкa
зaeм oт oбщнocттa зaeм нa eвpoпeйcкa инвecтициoннa бaнкa инициaтивa зa eвpoпeйcки
pacтeж зaeм нa eвpoпeйcкo oбeдинeниe зa въглищa и cтoмaнa пoмoщ нa eвpoпeйcкo
oбeдинeниe зa въглищa и cтoмaнa зaeм нa EBPATOM EBPATOM зaeм, пoлyчeн oт EC
мeждyнapoдeн зaeм

This translation of the expanded query is used for retrieval of the appropriate
documents from the full text search system.
The approach for query expansion is based on the intuition that when someone
searches for a concept they are interested in all subconcepts of the given one as well as
related concepts with step one from the initial concept because the related concepts that
are far from the initial concept could introduce too much unrelated content. In order to
provide more flexible control over the query expansion we have implemented the
following combinations:
NQE: No query expansion
QNA: Query expansion using transitive closure of the relation skos:narrower
QRE: Query expansion using the relation skos:related
QNR: Query expansion using both relations
The implementation provides a possibility for the user to select the translation
direction: Bulgarian-to-English, English-to-Bulgarian as well as the query expansion
approach: one of the above. As one can see, there are many ways for query expansion.
The best one would depend on the domain, the task and so on. After the evaluation of
this module we will improve it to achieve a better performance.

4 Statistical Machine Translation Module

As it was mentioned above, our task is to handle the out-of-vocabulary items for the
lexicons aligned to the ontologies, and also to provide a module for translation of user
queries based on statistical machine translation. User queries are mainly lists of key
words and phrases which we expect to be domain dependent. Thus the parallel corpora
on which Statistical Machine Translation system (SMT) are trained become very
important. As a system for statistical machine translation we selected Moses14. Moses
is a data-driven and state-of-the-art machine translation system. It provides three types
of translation model implementations: phrase-based models, where n-grams (“phrases”)
are the basic units of translation, hierarchical or syntax-based models, where infor-
mation about the structure of the parallel data can be incorporated, and factored
translation models, where additional linguistic information (e.g., lemma, part-of-speech
tag) can be integrated into the translation process.
Moses has two main components – a training pipeline and a decoder. The training
pipeline includes various tools for data pre-processing (e.g., for tokenisation, lower-
casing, removing very long sentences on the source or target side of the training data,

14
http://www.statmt.org/moses/.
280 K. Simov et al.

etc.), for accessing external tools for data alignment (GIZA++), language model
building (SRILM, KenLM, IRSTLM, RandLM), and implementations of popular
tuning algorithms. The Moses decoder tries to find the highest scoring sentence during
translation, or outputs a ranked list of translation candidates with additional informa-
tion. Standard tools for the evaluation of translations (e.g., BLEU scorer) are also
available. In addition to parallel data in the form of plain text, Moses can be used to
decode data represented as confusion networks or word lattices. In this way, ambiguous
input data, such as the output of an automatic speech recognizer, or a morphological
analyzer, can be processed to reduce erroneous hypothesis.
Two machine translation systems were created for the language pairs English-
Bulgarian and Bulgarian-English by the Moses open source toolkit (see [4]). Parallel
data from several sources was used to train factored translation models [3], which can
be viewed as an extension to standard phrase-based models, where more linguistic
information can be utilized in the translation process in addition to word forms.

4.1 Used Corpora

Parallel Corpora
We used several parallel corpora:
SETimes15 (154K sentences)
A parallel corpus of news articles in the Balkan languages, originally extracted from
http://www.setimes.com. Here we use the Bulgarian-English part which was cleaned
within the European EuroMatrixPlus project16. We manually checked the alignment for
more than 25000 sentences. The rest was automatically cleaned from sentence pairs
that were suspicious with respect to their translation.
Europarl17 (380K sentences)
The parallel texts were extracted from the Proceedings of the European Parliament.
Bulgarian-English BTB lexicon (9K word translations)
This is a lexicon created by professional lexicographers especially for machine
translation purposes.
JRC Acquis18 (364K sentences)
Parallel texts were extracted from the European Union legislation.
EAC-ECDC19 (7K sentences)
These sentences were extracted from translation memories published on the above
web page. The sentences were extracted manually because of the many alignment
discrepancies.
APIS Legal Corpus (3844 sentences)

15
http://opus.lingfil.uu.se/SETIMES.php.
16
http://www.bultreebank.org/EMP/.
17
http://www.statmt.org/europarl/.
18
http://optima.jrc.it/Acquis/index_2.2.html.
19
https://ec.europa.eu/jrc/en/language-technologies/ecdc-translation-memory.
 A Multilingual Access Module to Legal Texts 281

These sentences were extracted from parallel texts covering parts of Bulgarian
legislation.
The parallel data was cleaned semi-automatically. Non-translated sentences in the
Bulgarian data were detected and removed together with their equivalents in the
English data. Empty sentences or sentences longer than 80 words were removed with
the Moses script clean-corpus-n.perl. The parallel data was lowercased for training. In
the parallel data corpora we used mainly domain related corpora, but we also included
an out-of-domain corpus in order to cover more general language usage.
Monolingual Corpora
Additionally, the following data sets were used for the creation of suitable language
models: Bulgarian: National Reference Corpus (1.4M sentences) and the Bulgarian
data from the parallel corpora. English: Europarl (2M sentences) and the English data
(without Europarl) from the parallel corpora.

4.2 Data Analysis

The basic units of translation in Factored translation models are vectors of factors,
where a factor is usually a linguistic property attached to word forms, such as the
lemma, part-of-speech tag, or morphological properties associated with the word. In the
current settings, we make use of the linguistic analyses produced by the Bulgarian
pipeline btb-pipe (implemented partially within the EU-Cases project) and the English
system ixa-pipes (see [1]). In the preprocessing step, we perform sentence splitting,
tokenization, part-of-speech tagging, and lemmatization of the Bulgarian and English
parallel corpora.
The Bulgarian data was processed with the pipeline btb-pipe, which includes rule-
based, hybrid, and statistical components. It performs tokenization and sentence
splitting jointly. The tagging module assigns tags from a rich tagset [6], which encodes
detailed information about the morphosyntactic properties of words. In the experiment,
simplified POS tags were used as factors for translation to avoid data sparseness issues.
Lemmatization is based on rules, generated with the help of a morphological lexicon.
For the processing of the English part of the data we created a wrapper for the
modules of the ixa-pipes system [1] ixa-pipe-tok (version 1.7.0), and ixa-pipe-pos
(version 1.3.3). The wrapper includes an additional module which generates factored
output, suitable for use with the Moses factored system. The first module, ixa-pipe-tok,
takes plain text input, and carries out rule-based tokenization and sentence segmen-
tation. The next step in the pipeline, ixa-pipe-pos, includes POS tagging and lemma-
tization. For tagging we selected one of the provided POS models for English –
Perceptron [2]. The wrapper provides the option to preserve the number of lines in the
input and output English file. This option should be used in the processing of parallel
corpora to ensure that the resulting factored output can be aligned to its corresponding
Bulgarian file in case when the English file contains more than one sentence on certain
lines.
282 K. Simov et al.

4.3 Translation Models

During training, a mapping between the source and target language linguistic factors,
produced in the previous step, is established. The following lines describe the alter-
natives which were chosen for the two translation directions.
The SRI Language Modeling Toolkit (SRILM) [4] was used to build 5-gram
language models for the two translation systems. Two types of language models were
used by both systems – word-form-based and part-of-speech-tags-based ones. For the
word-based language models we used Kneser-Ney smoothing, while part-of-speech-
tag-based models were smoothed with the Witten-Bell method.
Both systems were tuned with minimum error rate training (MERT) (see [5])
implementation provided as part of the Moses toolkit. The English-to-Bulgarian
translation system uses two types of factors – wordform and POS tag. The wordform of
the source language is mapped to the wordform and POS tag factors on the target
language side. Additionally, the proposed wordform and POS tag sequences are
evaluated with the corresponding language models.
The Bulgarian-to-English system uses three types of factors – word-form, lemma,
and part-of-speech tag. In this translation scenario a source word form is again mapped
to the target word form and tag, but if no translation is found for the word, its lemma is
used instead. This greatly helps with data sparseness issues caused by the rich mor-
phology of Bulgarian. Once again two language models were employed, one for tags
and one for wordforms. The translation process can be divided into the following steps:
1. The input is processed with the corresponding pipeline, and factors are generated.
2. The factored data is lower-cased (Moses script lower-case.perl).
3. The factored data is translated with the corresponding model.
4. The translation output is de-tokenized (Moses script de-tokenizer.perl).
The translation output is re-cased (a Moses re-caser model was trained for each
translation direction). The evaluation of the statistical model showed the following
BLEU scores, presented in Table 2.

Table 2. The BLUE scores for the statistical machine translation modules
BG-EN EN-BG
BLEU 23.39 23.70

These results show that only statistical models are not sufficient for the real
applications.

5 Integration and Evaluation of Both Models

Both translation modules have some drawbacks. The ontology-based translation module
is not able to disambiguate ambiguous terms because of the lack of annotated corpora
from where a statistical model to be learnt. It also cannot translate out-of-vocabulary
 A Multilingual Access Module to Legal Texts 283

text. Such out-of-vocabulary text might have new paraphrases of existing concepts in
the ontology or complete unrelated keywords or phrases. The statistical machine
translation module also might not be able to use enough contexts in the user search query
to translate some keywords or phrases in the right way. Also, the SMT module is not
able to translate words that are not mentioned in the training corpus. For that reason
sometimes the translation contains words in the source language. In order to handle
these drawbacks of both modules and to gain from their complementary performance,
we ingrate them in the following way.
First, we translate the user query Q by each of the modules. The results are Qot and
Qsmt. Each of them could contain substrings from the source language. We delete
them from the two translated queries. Then we concatenate the two strings. The result is
Qtrans. This result is used for the full text search in a document database in the target
language.
For the actual tests we have selected three queries from the list of queries provided
by APIS20. The goal of the selected queries is to test the translation service on different
levels of specificity of the queries. Because the translation service is based on the
EuroVoc thesaurus, two of the queries are selected to correspond to terms in EuroVoc
and one is not corresponding to a term in EuroVoc. The queries were originally in
Bulgarian. First they were translated to English by a professional translator working
with legal texts. We consider these translations as original queries in English. Then the
queries were translated in both directions by people that know English, but are not
professional translators. In this way, we simulate the case in which the users of the
service will translate their queries by themselves. When we have the original queries
and their human translations, we perform translation of each query by the service in
three modes: simple translation, translation with query expansion using the relation
skos:narrower (Query Expansion 1), and translation with query expansion using
the relation skos:narrower and skos:related (Query Expansion 2). The fol-
lowing table summarises the variants of the three queries:
Q1 English Original (EO1): “consumer protection”
Q1 English Human Translation (EH1): “consumer protection”
Q1 English Simple Translation (ES1): “consumer protection”
Q1 English Query Expansion 1 (E11): “consumer protection” “consumer informa-
tion” “european consumer centres network” “product quality” “product life”
“product designation” “consumer movement” “product safety” “defective product”
Q1 English Query Expansion 2 (E21): “consumer protection” “housing law” “ad-
vertising” “advertising malpractice” “food safety” “producer’s liability” “public
health” “restriction on competition” “consumer information” “labelling” “social
labelling” “european consumer centres network” “product quality” “competitive-
ness” “designation of origin” “industrial manufacturing” “quality control of agri-
cultural products” “quality standard” “quality control circle” “product life” “product
designation” “consumer movement” “product safety” “safety standard” “defective
product”
Q1 Bulgarian Original (BO1): “зaщитa нa пoтpeбитeлитe”

20
http://apis.bg/en/.
284 K. Simov et al.

Q1 Bulgarian Human Translation (BH1): “зaщитa нa пoтpeбитeлитe”

Q1 Bulgarian Simple Translation (BS1): “зaщитa нa пoтpeбитeлитe”
Q1 Bulgarian Query Expansion 1 (B11): “зaщитa нa пoтpeбитeля” “пoтpeбитeлcкa
инфopмaция” “мpeжaтa oт eвpoпeйcки пoтpeбитeлcки цeнтpoвe” “кaчecтвo нa
пpoдyктa” “живoт нa пpoдyктa” “нaимeнoвaниe нa пpoдyкт” “движeниe нa
пoтpeбитeли” “бeзoпacнocт нa пpoдyктa” “дeфeктeн пpoдyкт”
Q1 Bulgarian Query Expansion 2 (B21): “зaщитa нa пoтpeбитeля” “жилищнo
пpaвo” “peклaмa” “нeзaкoнни peклaмни пpaктики/нeдoбpocъвecтнa peклaмa”
“xpaнитeлнa бeзoпacнocт” “oтгoвopнocт нa пpoизвoдитeля” “oбщecтвeнo
здpaвe” “oгpaничeниe нa кoнкypeнциятa” “пoтpeбитeлcкa инфopмaция” “eти-
кeтиpaнe” “coциaлнo клacиpaнe” “мpeжaтa oт eвpoпeйcки пoтpeбитeлcки
цeнтpoвe” “кaчecтвo нa пpoдyктa” “кoнкypeнтнocт” “oбoзнaчeниe зa пpoиз-
xoд” “пpoмишлeнo пpoизвoдcтвo” “кaчecтвeн кoнтpoл нa зeмeдeлcкa пpo-
дyкция” “cтaндapт зa кaчecтвo” “кpъг зa кoнтpoл нa кaчecтвo” “живoт нa
пpoдyктa” “нaимeнoвaниe нa пpoдyкт” “движeниe нa пoтpeбитeли” “бeзo-
пacнocт нa пpoдyктa” “cтaндapт зa бeзoпacнocт” “дeфeктeн пpoдyкт”
Q2 English Original (EO2): “green area”
Q2 Bulgarian Original (BO2): “зeлeни плoщи”…21
Q3 English Original (EO3): “child support”
Q3 Bulgarian Original (BO3): “издpъжкa нa дeтe”…22
The evaluation of the queries is with the operator AND for the words within the
quotes and operator OR between different quotes. Each query was evaluated with
respect to the search system provided by APIS and each returned document was
evaluated by an independent subject knowing only the original query (and never
having seen the translations). Each document was evaluated by assigning one of the
following three categories: exact match, non-exact match and no match.23 The non-
exact match is necessary because in many cases the document is relevant to the original
query, but the content of the query itself is not mentioned in the document. The
documents returned by the queries in group of queries for each original query were put
together and each of the documents was evaluated with respect to the three categories.
Because the returned documents for some of the queries overlap, the actual number of
the evaluated documents is smaller than the expected 400 documents. The following
table represents the number of all documents in the group. See Table 3.
On the basis of these results we can draw some conclusions. For general queries
from a thesaurus (Query 1 in Table 3) the performance slightly decreases in the cases
of query expansion (B11, B21, E11, E21 in Table 4). For specific queries from a

21
Some of the queries were skipped due to the space limits.
22
The English and the Bulgarian queries are very similar to each other.
23
The traditional evaluation in terms of precision and recall as well as in terms of their combinations is
not possible in our setup, because we do not have a large set of documents and appropriate set of
queries. Creation of such a test set is a time-consuming task which is outside of the work, reported
here.
 A Multilingual Access Module to Legal Texts 285

Table 3. The number of documents per query type.

Query group Number of documents Exact NonExact No
Q1 Bulgarian 272 182 56 34
Q1 English 244 194 18 32
Q2 Bulgarian 181 42 8 131
Q2 English 196 14 8 174
Q3 Bulgarian 264 146 8 110
Q3 English 164 78 12 74

thesaurus or for queries translated automatically without the usage of a thesaurus the
performance improves in all cases except one case – B23. In case of general queries the
number of non-exact match documents drastically increase in comparison to exact
matches. To sum up, in cases when the users are not sure in their own translation of the
query they are interested in, it is better to use the automatic translation service. If the
users look also for similar documents that contain no exact match, then they might
prefer to use the query expansion module.
The evaluation is given in the following Table 4.

Table 4. The manual evaluation of the documents per query.

Query Exact NonExact No Query Exact NonExact No
(BO1) 90 6 4 (EO1) 92 0 8
(BH1) 90 6 4 (EH1) 92 0 8
(BS1) 90 6 4 (ES1) 92 0 8
(B11) 60 22 18 (E11) 78 9 13
(B21) 50 28 22 (E21) 59 23 18
(BO2) 36 6 76 (EO2) 12 6 84
(BH2) 19 6 57 (EH2) 8 8 82
(BS2) 36 6 76 (ES2) 12 6 84
(B12) 36 6 76 (E12) 12 6 84
(B22) 36 6 76 (E22) 12 6 84
(BO3) 90 6 4 (EO3) 40 4 56
(BH3) 47 1 52 (EH3) 40 4 56
(BS3) 58 4 38 (ES3) 62 10 28
(B13) 58 4 38 (E13) 62 10 28
(B23) 44 4 52 (E23) 65 6 29

6 Conclusions

The paper reports on a multilingual access module to legal data, which combines two
approaches to translation of users’ queries – ontology-based and statistical.
286 K. Simov et al.

The experiments showed that the applicability of the module is better when both
modules are combined to work together. Both modules require the availability of
appropriate data – lexical resources and ontologies for the ontology-based one as well
as big high-quality corpora for the statistical one.
Our future work includes the following streamlines: testing the presented ideas on
other languages; evaluation on the combined module setting; qualitative evaluation of
the results with real users; query expansion with Linked Open Data reasoning.

References
1. Agerri, R., Bermudez, J., Rigau, G.: IXA pipeline: efficient and ready to use multilingual NLP
tools. In: Proceedings of the Ninth International Conference on Language Resources and
Evaluation (LREC 2014) (2014)
2. Collins, M.: Discriminative training methods for hidden Markov models. In: Proceedings of
the ACL-02 Conference on Empirical Methods in Natural Language Processing, vol. 10,
pp. 1–8 (2002)
3. Koehn, P., Hoang, H.: Factored translation models. In: Proceedings of the 2007 Joint
Conference on Empirical Methods in Natural Language Processing and Computational
Natural Language Learning, pp. 868–876 (2007)
4. Koehn, P., et al.: Moses: open source toolkit for statistical machine translation. In:
Proceedings of the 45th Annual Meeting of the ACL on Interactive Poster and Demonstration
Sessions, pp 177–180 (2007)
5. Och, F.J.: Minimum error rate training in statistical machine translation. In: Proceedings of
the 41st Annual Meeting on Association for Computational Linguistics, vol. 1, pp. 160–167
(2003)
6. Simov, K., Osenova, P., Slavcheva, M.: BTB-TR03: BulTreeBank morphosyntactic tagset
BTB-TS version 2.0 (2004)
7. Simov, K., Osenova, P.: Applying ontology-based lexicons to the semantic annotation of
learning objects. In: Proceedings from the Workshop on NLP and Knowledge Representation
for eLearning Environments, RANLP-2007, pp. 49–55 (2007)
8. Simov, K., Osenova, P.: Language resources and tools for ontology-based semantic
annotation. In: Oltramari, A., Prévot, L., Huang, C.-R., Buitelaar, P., Vossen, P. (eds.)
OntoLex 2008 Workshop at LREC 2008, pp. 9–13. Published by the European Language
Resource Association ELRA (2008)
9. Simov, K., Peev, Z., Kouylekov, M., Simov, A., Dimitrov, M., Kiryakov, A.: CLaRK - an
XML-based System for Corpora Development. In: Proceedings of the Corpus Linguistics
2001 Conference, 553–560 (2001)
 Combining Natural Language Processing
Approaches for Rule Extraction
from Legal Documents

Mauro Dragoni1(B) , Serena Villata2 , Williams Rizzi3 , and Guido Governatori4

1
Fondazione Bruno Kessler, Trento, Italy
dragoni@fbk.eu
2
CNRS, I3S Laboratory, Paris, France
villata@i3s.unice.fr
3
Universitá degli Studi di Trento, Trento, Italy
wrizzi@fbk.eu
4
NICTA Queensland, Brisbane, Australia
guido@governatori.net

Abstract. Legal texts express conditions in natural language describing

what is permitted, forbidden or mandatory in the context they regulate.
Despite the numerous approaches tackling the problem of moving from
a natural language legal text to the respective set of machine-readable
conditions, results are still unsatisfiable and it remains a major open chal-
lenge. In this paper, we propose a preliminary approach which combines
different Natural Language Processing techniques towards the extraction
of rules from legal documents. More precisely, we combine the linguistic
information provided by WordNet together with a syntax-based extrac-
tion of rules from legal texts, and a logic-based extraction of dependen-
cies between chunks of such texts. Such a combined approach leads to a
powerful solution towards the extraction of machine-readable rules from
legal documents. We evaluate the proposed approach over the Australian
“Telecommunications consumer protections code”.

1 Introduction
Applying deontic reasoning techniques to real world scenarios has to face the
challenge of processing natural language texts. On the one side, all codes and
legal documents of public institutions and companies are expressed in natu-
ral language, and it is very unlikely to have a structured (possibly machine-
processable) representation of the deontic conditions contained in such docu-
ments. On the other side, automated reasoning techniques need to process for-
mal conditions to infer further information, or to check whether the observed
behavior is compliant with such conditions, or whether a violation occurred. In
this kind of frameworks, the basic representation of a legal rule is as follows:
sup received complaint ⇒ [Obl]inform consumer process meaning that a sup-
plier has to inform the consumer of the complaint procedure upon reception of

c Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 287–300, 2018.
https://doi.org/10.1007/978-3-030-00178-0_19
288 M. Dragoni et al.

a complaint. Note that this kind of rules are not always clearly identifiable in
legal texts, and this task is difficult even for humans, becoming challenging for
an automated system. Defining systems able to tackle this task in an automated
way is a main challenge that received a lot of attention in the past years from the
legal information systems community, and heterogeneous approaches have been
proposed, e.g., [1,2]. This interest is due, not only to the difficulty for humans
to address such a task, but also to the fact that the task is extremely time con-
suming for humans, and (even partially) automating it to reduce the amount of
work demanded to humans would become a valuable support.
Despite the huge number of proposed approaches, the problem of extracting
rules or conditions from legal texts is still open. In this paper, we start from the
observation that, given the difficulty of the task, the adoption of a single Natural
Language Processing (NLP) approach to solve it would not lead to satisfiable
results, as witnessed by very limited adoption of the current frameworks. The
research question we answer in this paper is: How to combine different NLP
approaches to extract in an automated way a set of rules from natural language
legal texts? This question breaks down into the following subquestions: (1) How
to deal with the variability of natural language texts for the identification of
the deontic components of each rule?, and (2) How to combine a syntax-based
approach and a semantic-based one to identify the terms composing each rule,
and correctly assign them as being the antecedent/consequent of the rule?
To answer these questions, we adopt and combine a set of NLP techniques.
More precisely, our framework for automated rules generation exploits the Stan-
ford Parser to obtain the grammatical representation of the sentences, and Word-
Net1 to deal with the variability of the language in expressing the deontic com-
ponents in natural language legal texts. We combine this syntactic-based rules
extraction approach, relying on the well known Stanford Parser, together with
a logic-based approach, exploiting the Boxer framework [3] for the extraction
of logical dependencies between chunks of text. The results of the evaluation of
our combined framework on a section of the Australian “Telecommunications
consumer protections code” show the feasibility of the proposed approach, and
foster further research in this direction. The advantage of our approach is that
there is no need to learn how to extract the rules building a huge annotated data
set of legal documents as for machine learning approaches.
The remainder of this paper is as follows: Sect. 2 discusses the related liter-
ature and compares it to the proposed approach. Section 3 presents the overall
framework for automated rules extraction, and Sect. 4 describes the evaluation
setting.

2 Related Work
The automated processing of legal texts to extract some kind of information is
a challenge that received a lot of attention in the literature. [4] address an auto-
mated processing of legal texts exploiting NLP techniques: they aim at classifying
1
https://wordnet.princeton.edu/.
 Combining Natural Language Processing Approaches for Rule Extraction 289

law paragraphs according to their regulatory content and extracting text frag-
ments corresponding to specific semantic roles relevant for the regulatory con-
tent, while our goal is to extract rules with deontic modalities from legal texts. [5],
instead, propose an automated framework for the semantic annotation of provi-
sions to ease the retrieval process of norms, [6] present a knowledge extraction
framework from legal texts, and [7] present a tool for extracting requirements
from regulations where texts are annotated to identify fragments describing nor-
mative concepts, and then a semantic model is constructed from these annota-
tions and transformed into a set of requirements. Also in these cases, the goal
of the automated processing of legal texts is different. [1] present an automated
concept and norm extraction framework that adopts linguistic techniques. The
goal of this paper is the same as ours: an automated norm/rules extraction sys-
tem will help in saving knowledge analysts a lot of time, and it also contributes to
a more uniform knowledge representation of such formal norms/rules. However,
the adopted methodology is different: they exploit Juridical (Natural) Language
Constructs (JLC) that formalize legal knowledge using NLP by introducing a set
of predefined natural language constructs to define a subset of all possible legal
sentences. This kind of “patterns” is identified in the text thanks to the identifi-
cation of noun and verb phrases, and then they are translated into formal rules.
Similarly to them, we define “patterns” for detecting the deontic rules, but we
combine two approaches to lead to better results: we rely on the structured rep-
resentation of the sentence returned from the parser and its logical one returned
from Boxer. Finally, [1] do not consider the identification of deontic modalities
in rules, and no evaluation of their automated norms extraction framework is
provided thus results cannot be compared. [8] use machine learning for Dutch
regulations, [9,10] do the same for Italian ones. These approaches classify docu-
ments or sentences, differently from our methodology where rules are extracted
from the structural representation of legal texts. Finally, [2] present a linguistic-
based approach to extract deontic rules from regulations. As underlined by the
authors, Stanford parser has not been evaluated against legal sources, that is
the what we do in our own framework and they do as well. However, we do not
exploit the General Architecture for Text Engineering, and our approach does
not require to annotate the legal texts. To obtain satisfiable results, we combine
the result of the parser together with the logical dependencies between chunks of
text extracted from the document through Boxer. An experimental comparison
with the performances reported in these works is difficult as the data sets used
to evaluate them are not available. [11] present a framework to automatically
extract semantic knowledge from legislative texts. A similarity with our work
is that, instead of using pattern matching methods relying on lexico-syntactic
patterns, they propose to adopt syntactic dependencies between terms extracted
with a syntactic parser. The idea, on which the present paper is grounded as well,
is that syntactic information are more robust than pattern matching approaches
when facing length and complexity of the sentences. The difference consists in
the kind of information extracted, legal rules in our case, and three semantic
labels, namely active role, passive role, and involved object in their work.
290 M. Dragoni et al.

3 The Framework
The combined NLP approach implemented in this paper adopts several compo-
nents to automatically generate rules from natural language legal texts. In par-
ticular, it exploits the following elements described in more details later on in
this section: (i) a lightweight ontology describing the deontic linguistic elements
allowing for the identification of the obligations, permissions, and prohibitions in
legal texts; (ii) a lightweight ontology describing how the natural language text
is structured, and how punctuation can be interpreted for helping the extrac-
tion of rules2 ; (iii) a NLP library, namely, the Stanford Parser library3 , used for
parsing natural language sentences to retrieve their grammatical representation.
We decided to adopt Stanford Parser as it is the reference parser for parsing
natural language sentences in English; (iv) a Combinatory Categorial Grammar
(CCG) parser tool including the Boxer framework [3], used for extracting logical
dependencies between chunks of text from the document.
The resulting combined framework is an extension of the approach presented
in [12]. In particular, the following drawbacks have been addressed with respect
to [12]: (i) the deontic ontology has been extended by extracting from Word-
Net [13] all synsets related to the meaning of the Obligation, Permission, and
Prohibition concepts (as described in Subsect. 3.1). In this way, we are able to
improve the precision of the term annotation activities with the deontic labels;
(ii) the set of the patterns used for detecting deontic rules has been enriched;
(iii) a parallel branch integrating the functionalities of the CCG parser has been
integrated to analyze the text from a different perspective. The analysis results
obtained by the CCG parser are then merged with the output of the NLP-only
branch for extracting the final set of rules. Figure 1 shows the pipeline of the
proposed framework.
After the preliminary steps consisting in the extraction of the text from
source documents, and the composition of the separated sentences generated
by the extractor, the structured representation of the text follows two parallel
branches implementing two different analysis techniques. In the lower branch,
the modules of the Stanford NLP library are applied for tagging sentence con-
tent, and building the related tree for extracting the terms contained in each
sentence. Then, the deontic ontology is applied to annotate each term with the
appropriate label, i.e., obligation, permission, prohibition. Finally, the system
looks for patterns within the terms set of each sentence in order to compose the
rules.

2
Note that these ontologies are explicitly called lightweight ontologies as they are
not expected to be used to normalize the concepts of legal text by mapping the
legal terms into concepts in ontology, and obtain the meaning of the text by using
the ontology structure. They uniquely provide a support for detecting the deontic
components in legal texts and the structure of such texts, respectively.
3
http://nlp.stanford.edu/software/lex-parser.shtml.
 Combining Natural Language Processing Approaches for Rule Extraction 291

Fig. 1. The pipeline of the proposed framework.

In the upper branch, instead, the CCG parser is applied to the full sentence to
extract logic relationships between terms. Then, the output of the CCG parser is
used for confirming the rules extracted through the lower branch, and for discov-
ering new relationships between terms that have not been detected by applying
the patterns adopted by the NLP parser. Each component of the pipeline is now
detailed.

3.1 Deontic Lightweight Ontology

The deontic lightweight ontology, called normonto, has been designed to sup-
port the system in the automated identification of the normative component of
the rules. More precisely, this ontology is exploited to identify whether a term
expresses a prohibition, a permission, or an obligation. Even if several ontologies
have been proposed in the latest years to represent such a kind of knowledge
in different contexts, the aim of the normonto ontology is not to represent and
model legal concepts but to specify the lexicon used to express permissions, pro-
hibitions and obligations in natural language legal texts. For this reason, we spec-
ify the three main concepts called Obligation, Permission, and Prohibition.
The lexicon used to express the normative component in legal texts is represented
in the ontology as individuals of such subclasses. For instance, the individual
must identifies an obligation, thus it belongs to the class LexicalTermObl, and
the individual not be allowed identifies a prohibition, thus belonging to the
class LexicalTermPro. Note that this ontology is intended to be general pur-
pose and extensible, and differently from the text structure ontology we present
in the next section, it can be exploited by the system to extract the deontic
component of the rules from heterogeneous legal texts. Finally, the ontology is
intended to model the legal lexicon in English. Further extensions to cover mul-
tilingual rules extraction are considered for future research. The selection of the
keywords modeled as individuals in the ontology has been performed by starting
from a basic set of keywords related to the concepts of prohibition, permission,
292 M. Dragoni et al.

and obligation. Such a set has been used for querying WordNet to extract syn-
onyms, hypernyms, and hyponyms that are directly connected with each element
of the set of keywords mentioned above. The process has been run for three times
and, after each step, the content extracted for enriching the ontology has been
manually validated.

3.2 Text Structure Lightweight Ontology

In order to support the NLP algorithm in the analysis of different textual struc-
tures, a lightweight ontology, defining the main elements of the text organization,
has been modeled in order to effectively address our particular use case. Depend-
ing on the text structure that has to be analyzed, it might be necessary to model
different lightweight ontologies dedicated to those particular purposes.
Concerning the concepts definition, we modeled three main concepts: (i)
Document, defining the conceptualization of the entire text to analyze; (ii)
TextChunk, defining a single piece of text containing valuable information
(i.e. antecedent or consequent of the rule that has to be extracted); and (iii)
Punctuation, defining the meaning that specific punctuation signs may have in
the text from the computational point of view (for instance, the “;” sign may be
used for splitting sentences).
Concerning individuals, we modeled each block of the text as a new individual
instantiating the TextChunk object. This way, we are able to represent each
sentence of the text, or part of it, as a new element of the ontology in order
to allow the definition of their semantic relations used by the system for the
extraction of the rules.
Besides concepts and individuals, we define two object properties
(hasGeneralChunk and hasPart, the second one modeled as inverseOf of the
first one) and one data property (hasText). The two object properties are used
for modeling the hierarchical relationships between different TextChunk-objects;
while, the hasText data property allows to associate the natural language text
with the correspondent individual.

3.3 Extraction of Sentences

The analysis of the text starts with the extraction of sentences of interest that are
subsequently used for the text analysis. The extraction of such sentences is done
by exploiting the structured nature of the text that generally characterizes legal
documents where a bullet-based representation is used for describing normative
conditions. As first step, we map single text chunks contained in the bullet
representation of the document to the lightweight ontology. In this way, we are
able to manipulate a linked structure of the text easing the extraction of the
full sentences. By considering the structured representation of the text as a tree,
we reconstruct the set of full sentences to analyze by starting from the root of
the tree and by concatenating, for each possible path, the text chunks found
until the leaves are reached. Let us consider an excerpt of the document used
 Combining Natural Language Processing Approaches for Rule Extraction 293

as test case (Sect. 4) showing the structured representation of one of the norms
contained in the document:
(1) - Acknowledging a Complaint:
(2) --- immediately where the Complaint is made in
person or by telephone;
(3) --- within 2 Working Days of receipt where the
Complaint is made by:
(4) ----- email;
(5) ----- being logged via the Supplier’s website
or another website endorsed
by the Supplier for that purpose;
(6) ----- post; and
(7) ----- telephone and a message is recorded
without direct contact with a
staff member of the Supplier.

By performing the mapping between the text and the lightweight ontology,
the resulting assignments are the “Level 1” to the first chunk, “Level 2” to the
second and third ones, and “Level 3” to the others. By navigating through the
tree representation, the sentences extracted from the text are the concatenations
of the following text chunks (based on the ids written at left of each chunk):
“1-2”, “1-3-4”, “1-3-5”, “1-3-6”, “1-3-7”. As in Sect. 3.2, the punctuation ele-
ments are used as regulators for deciding where to split sentences in case of
complex structures. Sentences extracted at this step are then used for the extrac-
tion of the single terms.

3.4 The Use of the Stanford NLP Library

The extraction of rules from natural language legal texts requires the use of tools
able to provide a grammatical structure of the text that may be exploited for
inferring the different components of a logical rule. The facilities available for
having an effective representation of sentences are very limited. By analyzing the
state of the art, one of the most prominent library is the one provided by Stan-
ford. Such a library includes a Tree Parser able to produce a tree-based repre-
sentation of each sentence and to tag them with grammatical prefixes. Moreover,
the parser includes also a facility able to produce a set of grammatical relations
explained which dependency elapses between two terms. The role of the parser is
to work out the grammatical structure of sentences, for instance, which groups of
words go together and which words are the “subject” or the “object” of a verb.
The Stanford Tree Parser is a probabilistic parser using knowledge of language
gained from hand-parsed sentences to try to produce the most likely analysis of
new sentences. Even if statistical parsers still make some mistakes in exceptional
cases, they commonly work very well and, currently, they are the most suitable
solution for a preliminary text analysis. In the proposed approach, we decided
to use the Stanford NLP library for parsing the extracted sentences, and to use
the produced output as starting point for terms extraction. Let us consider the
following sentence: “Suppliers must demonstrate, fairness and courtesy, objec-
tivity, and efficiency by Acknowledging a Complaint within 2 Working Days of
294 M. Dragoni et al.

receipt where the Complaint is made by email.” By parsing this sentence, we

obtain the grammatical tree of the sentence shown below4 :
(ROOT
(S
(NP (NNS Suppliers))
(VP (MD must)
(VP (VB demonstrate) (, ,) (VB fairness)
(CC and)
(NP
(NP (NN courtesy) (, ,) (NN objectivity) (, ,)
(CC and)
(NN efficiency))
(PP (IN by)
(S
(VP (VBG Acknowledging)
(NP (DT a) (NN Complaint))
(PP (IN within)
(NP
(NP (CD 2) (JJ Working) (NNS Days))
(PP (IN of)
(NP (NN receipt)))))
(SBAR
(WHADVP (WRB where))
(S
(NP (DT the) (NNP Complaint))
(VP (VBZ is)
(VP (VBN made)
(PP (IN by)
(NP (NN email)))))))))))))
(. .)))

3.5 Extraction of Terms

Given the parsed version of each sentence, the next step consists to extract
relevant terms from them. With “term” we do not mean a single word (or com-
pound names) having a meaning in a vocabulary, but we mean a complex textual
expression representing an entire concept. The extraction of the terms follows
the identification of the subordinate sentences identified by the parser. In gen-
eral, we interpret the beginning of a new sentence (or a subordinate one) as the
beginning of a new term with some exceptions based on the content of the gener-
ated tree. Some examples are (i) if an extracted term starts with the expression
“to VERB”, the term is automatically concatenated with the previous one, and
(ii) if an extracted term contains only one token, such a token is directly con-
catenated to the succeeding one. This mainly happens when tokens like “where”,
“what”, etc. are parsed. Let us consider again the sample sentence of Sect. 3.4
where the analysis of the parsed representation leads to the identification of the
following terms:
-: Suppliers must demonstrate fairness, and courtesy,
objectivity and efficiency, by
a: Aknowledging a Complaint within 2 Working Days
of receipt
b: where the Complaint is made by email

4
For more details about the meaning of each tag and dependency clauses used by the
parser, please refer to the official Stanford documentation: http://nlp.stanford.edu/
software/dependencies manual.pdf.
 Combining Natural Language Processing Approaches for Rule Extraction 295

The first row is not marked as actual term but as “implicit” term. Indeed,
as it will be explained in Sect. 4 concerning the document used as test case,
some text chunks occur in many sentences. Such terms, independently by their
eventual deontic meaning, are marked only once; while, for the other sentences,
they are considered as “implicit” terms and they are not marked. The role of
the “implicit” terms is to appear as antecedent of rules when, in a sentence, no
terms are detected as antecedent, but consequent are identified. Two terms are
identified here.

3.6 Annotation of Terms with Deontic Tags

After the extraction of terms, they have to be annotated with the deontic tags
of Obligation, Permission, and Prohibition defined in the deontic lightweight
vocabulary. We assign the deontic tags by applying a text processing approach.
For each extracted term, we first verify if one of the lemmatized version of the
labels of the vocabulary is present in the sentence; if yes, the term is annotated
with the corresponding tag. A further check is performed to verify if, for example
in case of verb, the label and the “not” auxiliary have been split during the
term extraction in two consecutive terms. Indeed, if this happens, the identified
deontic tag has to be changed. For instance, for the labels “must” and “must not”
the deontic tags used are, respectively, the “Obligation” and the “Prohibition”
ones. In the example, the only term in which a deontic element is identified is
the implicit one that is annotated with the “Obligation” tag due to the label
“must”:
-: Suppliers must demonstrate fairness, and courtesy,
objectivity and efficiency, by [O]
a: Aknowledging a Complaint within 2 Working Days of
receipt
b: where the Complaint is made by email

3.7 Combination of Terms for Rule Definition

The last step consists in the definition of the rules obtained by combining the
extracted and annotated terms. For creating the rules, we apply a set of patterns
to the terms in order to detect what is the antecedent and the consequent of
each rule. Due to space reasons, we are not able to report all patterns defined
in the system, but only some of them:
[O] Term1
WHERE Term2 Rule: Term2 => [O] Term1

IF Term1
[O] THEN Term2 Rule: Term1 => [O] Term2

[O] Term1
UNLESS Term2 Rule: Term2 => [P] NOT Term1

[O] Term1
WHEN Term2
AFTER Term3 Rule: Term2 AND Term3 => [O] Term1
296 M. Dragoni et al.

It is important to highlight that, in case a deontic tag is used for anno-

tating an implicit term, such a tag is inherited by the first term following the
implicit one. This happens because implicit terms are not taken into account for
generating the rules.
Finally, by considering the annotated terms shown in Sect. 3.6 and by apply-
ing the first pattern due to the presence of the “where” label, the generated rule
is: b => [O] a.

3.8 The Use of the CCG Parser

The CCG parser has been integrated for performing a logic analysis of each
sentence in order to find relationships between the contained words. Indeed,
semantic representations produced by the CCG parser, known as Discourse Rep-
resentation Structures (DRSs), can be considered as ordinary first-order logic
formulas and can be exploited to find semantic connections between each word
extracted from the sentences. The aim of the integration of such a component
is to support the NLP pipeline described above in detecting the relationships
between the sentences from which the Stanford parser is not able to extract any
information through the application of the pattern-based mechanism (Sect. 3.7).
The exploitation of such logical relationships allows to improve the general effec-
tiveness of the rules extraction system.
Consider the following example of the output generated by the CCG parser
(Fig. 2) and of the linguistic graph that we build starting from the relationships
between words found by the parser (Fig. 3).

Fig. 2. Example of output generated by Fig. 3. Example of the linguistic graph

the CCG parser. built by starting from the relationships
between words found by the parser.

Graph’s connections are exploited for two purposes. First, for sentences where
deontic rules have been extracted by the NLP-only pipeline, we verify if the CCG
parser finds relationships between the terms involved in the rule (the effectiveness
of the pipeline by considering the different scores between these rules will be
discussed in Sect. 4). Second, for sentences where deontic rules are not detected
 Combining Natural Language Processing Approaches for Rule Extraction 297

by the NLP-only pipeline, in particular by the pattern-based mechanism, if the

CCG parser identifies logical relationships between the terms contained in the
sentence, a new deontic rule is created and stored in the rules set.

4 Evaluation
The evaluation is based on the novel Australian Telecommunications Consumer
Protections Code, TC628-2012 (TCPC) effective from September 1st, 2012, in
particular Sections 8.2.1(a)–8.2.1(c) pertaining Compliant Management. The
section describes the obligations a telecommunication service provider has to
comply with when they receive a complaint from customer or consumer (for the
purpose of TCPC, Section 2 Terms and Definitions customer or consumer are
treated as synonymous).
The text under analysis contains a single top level clause (8.2.1) which is then
divided in 3 subclauses. Furthermore, it contains 19 clauses at level 3, 16 clauses
at level 4, and 4 level 5 clauses/conditions. The structure of the document (i.e.,
the organization of the clauses and their subclauses) indicates that the section
contains 35 prima facie clauses.
For example, Section 8.2.1.a.(vii) states that:

advising the Consumer or former Customer of the proposed Resolution of

their Complaint within 15 Working Days from the date the Complaint is
received in accordance with clause 8.2.1 (a);

is mapped to the following prescriptive rule:

complaint => [O] inform_proposed_resolution_15_days

For the evaluation, we manually compared the rule-set manually generated by

an analyst, and the set of rules automatically extracted using the methodology
described in Sect. 3. Note that we cannot compare to a baseline as either none
of the related work provides the evaluation results or the task they addressed is
a different one thus results are not comparable.
For measuring the effectiveness of the system, we evaluated the following out-
puts and we compared them to the outputs contained in the manual created gold
standard: (i) the number of correct sentences extracted from the text, (ii) the
number of correct terms identified in the extracted sentences, (iii) the number of
correct deontic components annotations performed on the identified terms, (iv)
the number of correct rules generated from the extracted terms annotated with
the deontic component, and (v) the impact of the CCG parser in supporting the
detection of new patterns for generating rules.
By referring to the architecture described in Sect. 3, single blocks have been
evaluated separately enabling a more fine-grained analysis of potential pitfalls
in the effectiveness of system components. The evaluation of the lower branch of
the pipeline consisted in the computation of the following values: (i) the number
of sentences extracted correctly, (ii) the number of terms detected, (iii) the
298 M. Dragoni et al.

number of terms correctly annotated with the deontic tags, and (iv) the number
of rules that have been generated correctly. While, the evaluation of the upper
branch, consisted in measuring: (i) the agreement between the rules extracted
from the CCG output and the ones generated by the lower branch, and (ii) the
number of rules correctly extracted from the CCG output regarding sentences
for which the lower branch generated anything.

Lower Branch Evaluation. The extraction of the sentences is the first performed
task, the number of extracted sentences was 28 out of the same number of
sentences contained in the gold standard. Therefore, concerning the first output
the precision and recall of the system are 100%.
The second task is the identification of the terms within sentences. The gold
standard contains 65 terms extracted by the analysts; our system is able to
extract 59 terms whose 49 are correct. Therefore, the obtained recall is 90.78%
and the precision is 83.05%, with a F-Measure of 86.74%. Concerning the assign-
ment of the deontic annotation, 47 out of the 49 correctly identified terms have
been annotated with the proper deontic component, leading to a precision of
95.92%.
The last step consists in determining which of the 36 rules contained in the
gold standard have a counterpart in the automatically generated rule set com-
posed by 41 rules. A rule r in the automatically generated set has a counterpart
if there is a rule s in the manually generated set such that the proposition in the
right hand side (or consequent) of s is mapped to the consequent of r. The num-
ber of rules satisfying this condition is 33 out of 36 with a Precision of 80.49%
and a Recall of 91.67%.
Finally, the last operation is to determine which extracted rules have a
full correspondence with the manually generated rules: 24 of the automatically
extracted rules have a corresponding rule in the manually generated rule set.
This means that, as final assessment, we obtain a recall of Precision of 66.67%.

Upper Branch Evaluation. The first evaluation performed on the upper branch
of the pipeline was the measure of the agreement between the rules generated by
the lower branch and the ones inferred from the output of the CCG parser. By
starting from the output of the CCG parser, we firstly verify if words belonging
to different terms are directly related by one of the logical modifiers used by the
CCG parser for representing relationships between words. After the verification
of the all generated relationships, we computed how many of them exist also in
the set of the rules generated by the NLP parser used in the lower branch.
The set of relationships between terms extracted by the CCG parser con-
tains 51 relationships and, by transforming them in rules, 35 of them have a
counterpart (as defined in the previous paragraph) in the gold standard. With
respect to the lower branch, the CCG parser was able to find 2 new rules hav-
ing a counterpart in the gold standard. This means that the recall increased
to 97.22% (35 rules detected out of 36), but the precision decreased to 68.63%
due to the high number of relationships extracted by the CCG parser. Indeed,
the CCG parser works at a more fine-grained logical-linguistic level with respect
 Combining Natural Language Processing Approaches for Rule Extraction 299

to the NLP-only parser; therefore, the detection of relationships between terms

that are not actually a rule it is more easy.
Another interesting result is the number of deontic annotations that have
been found by the CCG parser. The number of such annotations increased from
47 to 49 out of the 49 annotations contained in the gold standard by reaching a
precision of 100.00%.
Finally, 29 out of the 35 rules found by the CCG parser have a full correspon-
dence with the manually generated rules. This means that the synergistic use of
the NLP-only and CCG parser led to an improvement of 5 rules by increasing
the precision of the system from 66.67% to 80.56%.

Discussion of the Results. As we mentioned above, the text we analyzed contains

35 prima facie clauses, and some of these rules require to be decomposed in
two sub-rules to fully capture the nuances of the conditions under which the
obligations hold. For example, as we have seen in Sect. 3, Section 8.2.1.a.(xiii)
splits in two clauses each requiring two rules. Furthermore, we would like to
point out that the number of rules required to capture a norm could depend on
the logical formalism used to reason with the rule. For example, if a condition
of activation of an obligation is disjunctive, it is represented by two rules in the
manually generated rule set. However, the disjunction could be represented by
a single proposition encoding both of them. Thus, the number of rules required
to model a normative clause depends on the underlying logic. This means that
we can take as reference for the computation of the recall not the actual number
of rules in the reference rule set, but the number of prima-facie clauses. In this
case, the extracted rules cover 31 out of the 35 prima facie clauses. Finally,
note that the examples used in this section (Section 8.2.1.a(vii) and 8.2.1a(xiii)
are correctly identified by our system. Concerning error analysis, in most cases
incorrect rules depend on the incorrect identification of the propositions in the
first step, or on the fact that the rules contain implicit terms in the left-hand
side to be derived from the right hand side. The correct treatment of these cases
is left as future research.

5 Concluding Remarks

In this paper, we have presented a combined framework for the automated extrac-
tion of rules from legal texts. The framework exploits both syntax-based pat-
terns extracted using the Stanford Parser, and logic-based ones extracted through
the Boxer framework. Several steps need to be addressed as future research to
improve the performance and the applicability of the system. First of all, we need
to construct a gold standard of legal rules extracted from different kinds of legal
texts in order to validate the proposed approach on a larger dataset, taking into
account the variability of the legal documents. Second, we need to capture the co-
reference links that are present in legal texts. For instance, consider a section of
the code that starts with Suppliers must provide Consumers with a Complaint
handling process that [...]. Then, in another part of the section, we have the
300 M. Dragoni et al.

following text A Supplier must take the following actions to enable this
outcome. How to recognize what is “this outcome”? We need to establish that a
co-reference occurred such that the outcome is to provide consumers with a com-
pliant handling that satisfies the certain requirements. Third, we need to align
the terms used in the legal text with the terms we want to use in the rules. As
shown in our evaluation, the difference between the hand-written rules and the
automated extracted ones is that different terms are used to constitute the same
rules.

References
1. van Engers, T., van Gog, R., Sayah, K.: A case study on automated norm extrac-
tion. In: Proceedings of JURIX, pp. 49–58 (2004)
2. Wyner, A., Peters, W.: On rule extraction from regulations. In: JURIX 2011, pp.
113–122 (2011)
3. Curran, J.R., Clark, S., Bos, J.: Linguistically motivated large-scale NLP with
c&c and boxer. In: Carroll, J.A., van den Bosch, A., Zaenen, A. (eds.) ACL 2007,
Proceedings of the 45th Annual Meeting of the Association for Computational
Linguistics, 23–30 June 2007, Prague, Czech Republic. The Association for Com-
putational Linguistics (2007)
4. Soria, C., Bartolini, R., Lenci, A., Montemagni, S., Pirrelli, V.: Automatic extrac-
tion of semantics in law documents. European Press Academic Publishing (2005)
5. Biagioli, C., Francesconi, E., Passerini, A., Montemagni, S., Soria, C.: Automatic
semantics extraction in law documents. In: ICAIL 2015, pp. 133–140 (2005)
6. de Araujo, D.A., Rigo, S., Muller, C., de Oliveira Chishman, R.L.: Automatic infor-
mation extraction from texts with inference and linguistic knowledge acquisition
rules. In: Web Intelligence/IAT Workshops, pp. 151–154 (2013)
7. Kiyavitskaya, N., et al.: Automating the extraction of rights and obligations for
regulatory compliance. In: Li, Q., Spaccapietra, S., Yu, E., Olivé, A. (eds.) ER
2008. LNCS, vol. 5231, pp. 154–168. Springer, Heidelberg (2008). https://doi.org/
10.1007/978-3-540-87877-3 13
8. de Maat, E., Winkels, R.: Suggesting model fragments for sentences in Dutch laws.
In: Proceedings of LOAIT, pp. 19–28 (2010)
9. Brighi, R., Palmirani, M.: Legal text analysis of the modification provisions: a
pattern oriented approach. In: ICAIL 2009, pp. 238–239 (2009)
10. Francesconi, E.: Legal rules learning based on a semantic model for legislation. In:
Proceedings of SPLeT Workshop (2010)
11. Boella, G., Di Caro, L., Robaldo, L.: Semantic relation extraction from legislative
text using generalized syntactic dependencies and support vector machines. In:
Morgenstern, L., Stefaneas, P., Lévy, F., Wyner, A., Paschke, A. (eds.) RuleML
2013. LNCS, vol. 8035, pp. 218–225. Springer, Heidelberg (2013). https://doi.org/
10.1007/978-3-642-39617-5 20
12. Dragoni, M., Governatori, G., Villata, S.: Automated rules generation from natural
language legal texts. In: ICAIL 2015 Workshop on Automated Detection, Extrac-
tion and Analysis of Semantic Information in Legal Texts (2015)
13. Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press, Cambridge
(1998)
 Analysis of Legal References in an Emergency
Legislative Setting

Monica Palmirani(&), Ilaria Bianchi(&), Luca Cervone(&),

and Francesco Draicchio

CIRSFID, University of Bologna, Bologna, Italy

{monica.palmirani,ilaria.bianchi,luca.cervone,
francesco.draicchio}@unibo.it

Keywords: NLP
URI
Legal XML
Modifications
Visualization

Network analysis

1 Introduction

Each legal system is a complex network of norms and the normative citations in the
texts are the legal method for referring to other parts of the same legal system
diachronically (dynamically over time) or synchronically (statically) (Palmirani and
Brighi 2006). Normative citations are the textual part of a legal document that refers to
another legal source in the same legal system (e.g., Sect. 3 of the Human Rights Act
1998 in the UK legal system) or also to other legal systems (e.g., European directives).
One of the most relevant legislative techniques1 uses citations to summarize the ver-
bosity of norms, create semantic relationships between different normative resources,
or amend the original text. We can classify citations using this taxonomy: (i) internal
and external to the same document; (ii) dynamic or static at a given time (e.g., London
Regional Transport Act 1984 and later London Regional Transport Act 1996);
(iii) citations that express semantic normative specification (extension or restriction; see
also interpretation); (iv) citations for referring to an already expressed piece of text
already without duplicating it (shortcut); (v) citations that semantically connect dif-
ferent documents under the same topic (clustering). In all these cases citations set up an
interesting apparatus for analyzing a country’s legislative approach. In particular, it is
possible to understand the legal drafting techniques adopted, as well as to detect
anomalies in order to increase the effectiveness of normative action.
On the basis of these arguments, this paper investigates the references of a legal
corpus of the ordinances issued by the Regional Commissioner for Emergency and
Reconstruction over the first 18 months after the 2012 earthquake in Emilia-Romagna,
Italy. The goal is to analyze the critical issues in the regulative strategy in emergency
situations in order to help the lawmaker act better in future disasters, extract infor-
mation concerning the number and the types of modifications produced, and support
the debate on a national law on emergency in the wake of natural disasters.

1
https://www.law.cornell.edu/citation/; http://filj.lawreviewnetwork.com/files/2011/10/EU_Citation_
Manual_2010-2011_for_Website.pdf; http://eur-lex.europa.eu/content/techleg/KB0213228ENN.pdf.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 301–313, 2018.
https://doi.org/10.1007/978-3-030-00178-0_20
302 M. Palmirani et al.

The outcomes here presented are developed as part of the Energie Sisma Emilia
research project, conducted by a consortium of universities in the Emilia-Romagna
Region.2,3 All the graphs are presented in dynamic visualization in web portal.4 This
makes it possible to cross-check groupings based on lexical-textual analysis and
groupings based on structural elements.

2 Methodology

The earthquake struck in Emilia-Romagna Region5 on 20 and 29 May 2012 and on 3

June 2012. The disaster area6 includes 33 municipalities and stretches over 3,173
square meters across the region. The Commissioner for Emergencies issued 350
ordinances deliberated over a three-year period (2012–2015), so as to support the
rebuilding, aid for the population, organization of the territory. The main research goal
is to investigate the corpus of legislative ordinances so as to make more effective the
legislative actions in emergency settings. Analyzing the legal citations and the corre-
spondent references we can discover some dysfunctional behaviour in the lawmaking
system, the concentration of some urgent topics, to detect weaknesses in normative
area, to orient a more coordinated legislative action at the national level. In the Italian
Senate there are several bills being debated on emergency situations. This analysis
could provide useful input to that legislative debate.
The research at the University of Bologna was concentrated over the first 18 months,
considering that we have a total of 252 ordinances were emitted between June 2012 to
the end of 2013. We have concentrated our attention only on the pick of the emergency
to depict urgent needs in the immediate disaster and to observe how legal drafting
techniques could support them. We are not interested to investigate the post-emergency
period (after 18 months) because other legal drafting mechanisms could be adopted
closer to the normal rule of procedure. The work was divided into four main steps.
1. One group from the University of Modena was focused on building a clustering of
the documents by thematic area. The clustering was produced with automatic text
analysis — developed by Pavone7 et al. — on the corpus of those ordinances, and
has provided a list of main categories and subtopics. Four main topics have been
singled out: (i) grant criteria and contributions; (ii) management of allocation of
resources; (iii) urgent works for municipalities, school and church buildings;
(iv) and interventions in support of the population.

2
http://www.capp.unimo.it/pubbl/cappapers/Capp_p120.pdf.
3
http://www.energie.unimore.it/.
4
http://137.204.21.115/sisma-2012.
5
http://www.regione.emilia-romagna.it/terremoto/gli-atti-per-la-ricostruzione.
6
http://www.regione.emilia-romagna.it/terremoto/sei-mesi-dal-sisma/approfondimenti/il-documento-
completo-della-regione-emilia-romagna.
7
http://www.energie.unimore.it/analisi-lessico-testuale-delle-ordinanze-commissariali-un-contributo-
alla-legge-nazionale-su-emergenza-e-ricostruzione/.
 Analysis of Legal References in an Emergency Legislative Setting 303

2. Using this list of topics, automatically extracted by the first research group, a second
group of legal experts at the University of Bologna manually built a lifecycle map
of the modifications that took place in the ordinance corpus. In meantime, the
experts also classified the material and the partitions that had been modified. The
goal was to investigate the lawmaker’s behavior in taking legislative actions.
3. A third group at the University of Bologna extracted the references using a parser
(see Sect. 1) for legislative documents (Palmirani and Brighi 2009; Palmirani and
Cervone 2009).
4. A fourth group performed legal analysis of the results and defined the visualization
using network-analysis graphs showing the correlation among the data extracted.
This methodology made it possible to evaluate the results of automatic extraction
through the expertise of legal scholars. From a pure methodological perspective, this
second method makes it possible to validate (a) the robustness of the cluster analysis
produced through automatic text analysis and (b) the network analysis produced with
Infomap (Pavone et al. 2016).
On 26 November 2015 the results of this analysis were presented to policymakers
in Emilia-Romagna Region and to representatives of the Italian Parliament in order to
provide inputs on which basis to enhance the legislative framework for emergency
situations now making its way through Parliament.

3 From Citations to Neutral References

The normative citations in the digital-legislative-document collections are modelled

using a regular grammar and identifier naming convention (e.g., URN:LEX,8 ELI,9
ECLI,10 Opijnen 2017 AKN,11 etc.) for transforming the narrative text into normalized
references: persistent, machine-readable, univocal, meaningful, language-independent
and human-understandable.12 The goal in this work is to produce a canonical citation
form independent of the different morphological variants used for the same conceptual
citation (e.g., in Italy, Act 20/2012, Act no. 20 of 2012, or Legge 20 del 2012). This
makes it possible to manage different versions the same citation over time. We adopt
the Akoma Ntoso (Palmirani 2011; Palmirani and Vitali 2011) naming convention for
translating textual citations into canonical form.
The transformation of the citations into neutral references is done using specific
NLP parsers for Italian legislation already used with success in several research pro-
jects (Palmirani and Brighi 2009), and in particular in a project for the High Court of
Cassation of Italy (about 50,000 documents of primary and secondary law). Different

8
https://tools.ietf.org/pdf/draft-spinosa-urn-lex-09.pdf.
9
http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52012XG1026%2801%29.
10
http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=URISERV:jl0056.
11
Akoma Ntoso Naming Convention verion 1.0, https://lists.oasis-open.org/archives/legaldocml/
201407/msg00014/Akoma_Ntoso_Naming_Convention_Version-2014-07-30-wd12.doc.
12
https://wiki.oasis-open.org/legalcitem/FundamentalRequirements3rd.
304 M. Palmirani et al.

approaches for reaching with this goal are possible in the state of the art, such as
(Bartolini et al. 2004; Palmirani et al. 2004; Biagioli et al. 2005; de Maat et al. 2006;
Francesconi and Passerini 2007; Lesmo et al. 2013; Winkels and Boer 2014; Waltl and
Florian 2014; Koniaris and Vassiliou 2014). However, the parser used by the authors is
based on a hybrid approach that first generates the XML document’s hierarchical
structure so as to then refine the normative references with precision. After detecting
the main parts (preface, preamble, body, conclusion), the parsers use about 6,000
patterns acquired from the High Court of Cassation XML corpus marked up by legal
experts using the Norma-Editor system (Palmirani and Benigni 2007). The patterns of
references are properly coded into regular expressions and then are used to find ref-
erences to other documents. By exploiting the document’s given XML structure, we
can also contextualize. For instance, if a date is detected in the conclusion of the
document, it is unlikely to be part of a reference, but if a date is detected in the
preamble of the document, the date is likely to be a part of a reference that, because it is
in the preamble, should be a static link. The same goes with the numbers of the articles.
If the parser finds them within an article contained in a quoted structure, then the parser
treats them as a part of the hierarchy of the cited document. The parser also uses
vocabularies of frequent citations in legislative acts and ordinances so as to build the
canonical form of the most frequent abbreviation (e.g., Constitutional Law, Civil Code,
etc.) (Palmirani and Cervone 2013).

3.1 Parser Outcome

We detected 252 ordinances and 12,727 references. The references are sorted into three
groups:
• modifications in the ordinance set extracted through the manual analysis of the legal
experts;
• semantic citations between the ordinances elaborated using the parser and filtered
with the modifications;
• semantic citations between national law and regional ordinances.

4 Modificatory Relationship Analysis

The modifications introduce complexity within a normative corpus, thus affecting the
certainty of law, the clarity of the text being updated, and the simplicity with which
norms can be applied. However, especially in a new emergency domain, modifications
are necessary to introduce details, specify the norms’ range of application, correct
clerical and substantive errors, and extend procedural deadlines. An analysis of mod-
ifications makes it possible to evaluate the effectiveness of normative actions in the
emergency situations and to provide instructions for the lawmaker so as to make norms
more effective.
 Analysis of Legal References in an Emergency Legislative Setting 305

The modifications in the first 18 months were 814, made to 88 ordinances (from 22
June 2012 to 17 October 2013); 80% in 2012 (7 months after the event); 52% sub-
stitutions; 32% insertions; 5% clerical errors; 8% repeals; and 3% prorogations (Fig. 1).
This work was conducted manually by legal experts, building the lifecycle of the
normative documents. We adopt the modification taxonomy developed as part of the
NormeInRete and Akoma Ntoso projects (Palmirani and Cervone 2014).

Fig. 1. Most frequent modifications using the Akoma Ntoso and NoremInRete taxonomy

The analysis highlights (a) the temporal period in which the modifications were
concentrated (August 2012, October 2012, December 2012) and (b) the domain in
which they acted (textual modifications in 2012 and prorogation of terms in 2013), and
(c) which ordinances were mostly affected by the changes.
The most affected ordinances by modifications are nos. 51/2012, 29/2012, 57/2012,
and 86/2012 (Fig. 2). Using the clustering analysis done by the University of Modena
and Reggio Emilia, we have manually checked the classifications of each ordinance,
finding that the most ammended ordinances belong to category 2 (Management of
Resources) and category 4 (Assistance to the population) (Fig. 3).

Fig. 2. Ordinances most affected by modifications

306 M. Palmirani et al.

Fig. 3. Topics most amended by modifications.

We have also created a matrix graph (Fig. 4) representing the distribution of the
different types of modifications over time. We have a concentration of textual modi-
fications, due also to clerical errors, a lack of legislative technique, and template
methodology. In the second year we have a concentration of prorogation of terms (blue
dots).

Fig. 4. Matrix of correlations between modified and modifier ordinances

 Analysis of Legal References in an Emergency Legislative Setting 307

Fig. 5. Graphs showing the impact of modifications on the ordinance corpus

The same information were also described by using a navigable bubble graph
(Fig. 5) on the Web that can be used to better analyse the data from a legal point of
view. The data are sorted by year and month. Each bubble in the ordinance is intended
to show the number of modifications within each type.
Another analysis was conducted for examining the evolution over time of modi-
fications of the most modified ordinance n. 57/2012. The Fig. 6 presents the articles
after 262, 305 and 535 days after the earthquake. The parallel lines pattern means that
the same articles were subject of recursive modifications. This is a critical indicator that
provides important information: those articles needed to be amended regularly due to
the inadeguacy of the normative effectiveness. The colour expresses the topic of the
modifcation.

5 Semantic References

The second goal was to analyse the semantic interconnection among ordinances in
order assess whether there were critical normative topics that need complex policy by
the decision-maker. Semantic web in legal domain is now quite developed (Casanovas
et al. 2016). Other works in the state of the art other inspired our research (Winkels
2012, 2013, 2014; Bommarito and Katz 2010, 2017; Boulet et al. 2010).
The following dataset was built using the result of the parser analysis and filtering
the normative references excluding the amendment link. The results (Fig. 7) stressed
the presence of four islands that are grouped around major topics: green means
“Assistance of the Population” norms; red means “Management of the Resources”; blu
means “Criteria for Rebuiliding and Funding”; purpule means “Urgent construction
works”.
308 M. Palmirani et al.

Fig. 6. Evolution of the modifications over time

The most cited ordinances are nos. 17/2012 (138), 57/2012(102), 82/2012 (72), and
19/2012 (69) (Fig. 8) and the topics that are most cited are “Assistance of the Popu-
lation” and “Management of the Resources”, the two main actions in the political
agenda of the regional government (Fig. 9).
Another interesting visualization is the following matrix graph (Fig. 10), making it
possible to see a regularity in the citation pattern. We can see parallel lines that indicate
the frequency of citations from one active act (X axis) and the cited act (Y axis). The
parallel lines mean that there is a template of citations (e.g., in the preamble) that all
refer to the same legislative basis. The avarage of the citations from an ordinance to
other ordinances is 6 references.
Another important analysis was intended to investigate the relationships among
regional ordinances and national laws (Fig. 11). This contributes to a better under-
standing of which legislative actions could be useful at the national level to improve the
regional legislative activities in emergency settings. In the Fig. 11 we could notice that
the decree n. 74/2012 is the most cited. It is the decree of the emergency declaration in
the geographic area. Other important primary law cited are: Act 122/2012, Act
340/2012, At 225/1992. The average number of citations from an ordinance to the
national law is 31 references.
 Analysis of Legal References in an Emergency Legislative Setting 309

Fig. 7. Network of citations among ordinances filtered by amendment links

Fig. 8. Frequency of citations (passive citations) without modifications

6 Future Work

Our next work going forward will be focused on two goals: (i) to compare and cross-
check results with the other team at the University of Modena and Reggio Emilia,
which used a different approach (fuzzy linguistic and lexicon analysis of the text); and
(ii) to find a more effective visualization for stressing the relationship between topics
and modifications; (iii) to make dynamic the visualization in the web portal using real-
time information in Akoma Ntoso. We aim to show the areas where the legislator was
forced to amend frequently. This suggests that we need to pay more attention to these
310 M. Palmirani et al.

Fig. 9. Topics of most cited (passive citation) ordinances without modifications

Fig. 10. Regularity in the active citations model.

critical topics in order to avoid modifications that undermine the effectiveness and
timeliness of actions. A greater amount of modifications means greater dispersion of
energy, effort, and resources, as well as problems having timely access to updated
versions of each regulation; it also means dealing with different interpretations by end-
users, making it difficult to give the process a single effective direction; and, finally, it
means a lack of simplification for the citizens affected by the emergency situation, as
well as corruption and mafia infiltration. We want also to improve the visualization of
the frequency of the citations among ordinances using coloured nodes according to the
prevalent topic (clusters) of the provisions. The more citations we have between
ordinances, the greater the likelihood of finding a topic that needs attention by
policymakers.
 Analysis of Legal References in an Emergency Legislative Setting 311

Fig. 11. Relationship between regional ordinances and national law

7 Conclusion

This research investigated the relationships between ordinances in the emergency

setting, with the aim of detecting potential dysfunctional situations, thus providing
helpful input for the regional and national lawmakers. Our analysis stressed the fol-
lowing points:
1. The ordinances do not follow a regular template, and this does not make it possible
to avoid textual mistakes, clerical errors, or a deluge of modifications. The legal
expert analysis stressed that the language is redundant and repetitive and can
therefore create distortions in the automatic classification.
2. The consolidation of modifications is not an easy task, considering that the template
of the ordinance is very flexible, unstructured, and variable, and the citations are not
accurate in their fragmentation of partitions.
3. Citations among ordinances reveal a concentration of effort on two topics:
(a) management of the rebuilding phase and (b) aid to the population. A national
law regulating these aspects could speed up the reconstruction and improve quality
of the life for the affected population.
4. Some topics are neglected in the first year: these include waste management and
transportation for workers and employees. There is no energy plan in the first year;
for example, no plan for managing toxic or radioactive waste coming from industry.
312 M. Palmirani et al.

5. The annexes are too technical and verbose. They are substituted at least three times
a year.
6. It is clear from Fig. 3 that there is no structural intervention plan and some actions
(e.g., school) are recursively represented without any clear plan.
7. It is difficult to classify ordinances manually by their title.
8. The policy did not adequately address the issue of transparency so as to avoid
corrupt behaviour and infiltration by the mafia.
A Web portal has been set up to illustrate the analysis of this work and to provide a
new methodology for monitoring the effectiveness of legislative actions, especially in
emergency situations where the timeliness is a key issue for decision-makers.

References
Bartolini, R., Lenci, A., Montemagni, S., Pirrelli, V., Soria, C.: Semantic mark-up of italian legal
texts through nlp-based techniques. Proceedings of LREC 2004, 795–798 (2004)
Biagioli, C., Francesconi, E., Passerini, A., Montemagni, S., Soria, C.: Automatic semantics
extraction in law documents. In: ICAIL 2005 Proceedings of the 10th International
Conference on Artificial Intelligence and Law, pp. 133–140. ACM, New York (2005)
Bommarito II, M.J., Katz, D.M.: A mathematical approach to the study of the united states code.
Phys. A Stat. Mech. Appl. 389(19), 4195–4200 (2010)
Bommarito II, M.J., Katz, D.M.: Measuring and Modeling the U.S. Regulatory Ecosystem.
J. Stat. Phys. 168, 1125–1135 (2017)
Boulet, R., Mazzega, P., Bourcier, D.: Network analysis of the French environmental code. In:
Casanovas, P., Pagallo, U., Sartor, G., Ajani, G. (eds.) AICOL -2009. LNCS (LNAI), vol.
6237, pp. 39–53. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-16524-5_4
Casanovas, P., Palmirani, M., Peroni, S., van Engers, T., Vitali, F.: Semantic web for the legal
domain: the next step. Semant. Web 7(3), 213–227 (2016)
Francesconi, E., Passerini, A.: Automatic classification of provisions in legislative texts. Artif.
Intell. Law 15(1), 1–17 (2007)
Koniaris, M., Vassiliou, I.A.Y.: Legislation as a complex network: modelling and analysis of
European Union legal sources. In: Hoekstra, R. (ed.) Legal Knowledge and Information
Systems. JURIX 2014: The Twenty-Seven International Conference. Frontiers in Artificial
Intelligence and Applications, vol. 260, pp. 143–152. IOS Press, Amsterdam (2014)
Lesmo, L., Mazzei, A., Palmirani, M., Radicioni, D.: TULSI: an NLP system for extracting legal
modificatory provisions. Artif. Intell. Law J. 2013(21), 139–172 (2013)
de Maat, E., Winkels, R., van Engers, T.: Automated detection of reference structures in law. In:
van Engers, T.M. (ed.) Legal Knowledge and Information Systems. Jurix 2006: The
Nineteenth Annual Conference, vol. 152, pp. 41–50. IOS Press, Amsterdam (2006)
van Opijnen, M., Palmirani M., Vitali, F., Agnoloni T.: Towards ECLI 2.0. In: 2017 International
Conference for E-Democracy and Open Government, P6082, pp. 1–9. IEEE, Los Alamitos
(2017). (atti di: 2017 International Conference for E-Democracy and Open Government,
Krems, Austria, 17–19 May 2017)
Palmirani, M., Benigni, F.: Norma-system: a legal information system for managing time. In:
Proceedings of the V Legislative XML Workshop, European Press Academic Publishing,
FIRENZE, pp. 205–224 (2007). (atti di: V Legislative XML Workshop, Fiesole, Firenze,
Italia, 14–16 Giugno 2007)
 Analysis of Legal References in an Emergency Legislative Setting 313

Palmirani, M., Brighi, R., Massini, M.: Processing normative references on the basis of natural
language questions. In: DEXA 2004 Proceedings of the Database and Expert Systems
Applications, 15th International Workshop, pp. 9–12. IEEE Computer Society (2004)
Palmirani M., Brighi R.: Legal text analysis of the modification provisions: a pattern oriented
approach. In: Proceedings of the International Conference on Artificial Intelligence and Law
(ICAIL) (2009)
Palmirani, M., Brighi, R.: Model regularity of legal language in active modifications. In:
Casanovas, P., Pagallo, U., Sartor, G., Ajani, G. (eds.) AICOL -2009. LNCS (LNAI), vol.
6237, pp. 54–73. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-16524-5_5
Palmirani, M., Brighi, R.: Time model for managing the dynamic of normative system. In:
Wimmer, M.A., Scholl, H.J., Grönlund, Å., Andersen, K.V. (eds.) EGOV 2006. LNCS, vol.
4084, pp. 207–218. Springer, Heidelberg (2006). https://doi.org/10.1007/11823100_19
Palmirani, M., Cervone, L.: Legal change management with a native XML repository. In:
Governatori, G. (ed.) Legal Knowledge and Information Systems. JURIX 2009. The Twenty-
Second Annual Conference, Rotterdam. 16th–18th December 2009, pp. 146–156. ISO Press,
Amsterdam (2009)
Palmirani, M., Cervone, L.: A multi-layer digital library for mediaeval legal manuscripts digital
libraries and archives. In: Communications in Computer and Information ScienceDigital
Libraries and Archives, Communications in Computer and Information Science 2013, vol.
354, pp. 81–92. Springer, Heidelberg, 9–10 February 2012. (atti di: IRCDL 2012, Bari)
Palmirani, M., Cervone, L.: Measuring the complexity of the legal order over time. In: AI
Approaches to the Complexity of Legal Systems, pp. 82–99. Springer, Heidelberg (2014)
Palmirani, M., Vitali, F.: Akoma-Ntoso for legal documents. In: Sartor, G., Palmirani, M.,
Francesconi, E., Biasiotti, M. (eds.) Legislative XML for the Semantic Web. Law,
Governance and Technology Series, vol. 4, pp. 75–100. Springer, Dordrecht (2011).
https://doi.org/10.1007/978-94-007-1887-6_6
Palmirani, M.: Legislative change management with Akoma-Ntoso. In: Sartor, G., Palmirani, M.,
Francesconi, E., Biasiotti, M. (eds.) Legislative XML for the Semantic Web. Law,
Governance and Technology Series, vol. 4, pp. 101–130. Springer, Dordrecht (2011).
https://doi.org/10.1007/978-94-007-1887-6_7
Pavone, P., Righi, R., Righi, S., Russo, M.: Text mining and network analysis to support
improvements in legislative action. In: The Case of the Earthquake in Emilia-Romagna,
Proceedings JADT2016, 7–10 giugno 2016, Nizza, Francia, pp. 237–247 (2016). ISBN 978-
2-7466-9067-7
Waltl, B., Florian, M.: Towards measures of complexity: applying structural and linguistic
metrics to german laws. In: Hoekstra, R. (ed.) Legal Knowledge and Information Systems.
JURIX 2014: The Twenty-Seven International Conference. Frontiers in Artificial Intelligence
and Applications, vol. 260, pp. 153–162. IOS Press, Amsterdam (2014)
Winkels R, Boer A.: Finding and visualizing dutch legislative context networks. In: Network
Analysis in Law. Diritto Scienza Tecnologia, pp. 157–182 (2014)
Winkels, R., Boer, A., Plantevin, I.: Creating context networks in Dutch legislation. In: Ashley,
K. (ed.) Legal Knowledge and Information Systems. JURIX 2013: The Twenty-Sixth
International Conference. Frontiers in Artificial Intelligence and Applications, vol. 259,
pp. 155–164. IOS Press, Amsterdam (2013)
Winkels, R., de Ruyter, J.: Survival of the fittest: network analysis of Dutch Supreme Court
Cases. In: Palmirani, M., Pagallo, U., Casanovas, P., Sartor, G. (eds.) AICOL 2011. LNCS
(LNAI), vol. 7639, pp. 106–115. Springer, Heidelberg (2012). https://doi.org/10.1007/978-3-
642-35731-2_7
Legal Ontologies and Semantic
Annotation
 Using Legal Ontologies with Rules
for Legal Textual Entailment

Biralatei Fawei, Adam Wyner(B) , Jeff Z. Pan, and Martin Kollingbaum

Department of Computing Science, University of Aberdeen, Aberdeen, UK

azwyner@abdn.ac.uk

Abstract. Law is an explicit system of rules to govern the behaviour

of people. Legal practitioners must learn to apply legal knowledge to
the facts at hand. The United States Multistate Bar Exam (MBE) is
a professional test of legal knowledge, where passing indicates that the
examinee understands how to apply the law. This paper describes an ini-
tial attempt to model and implement the automatic application of legal
knowledge using a rule-based approach. An NLP tool extracts informa-
tion (e.g. named entities and syntactic triples) to instantiate an ontology
relative to concepts and relations; ontological elements are associated
with legal rules written in SWRL to draw inferences to an exam ques-
tion. The preliminary results on a small sample are promising. However,
the main development is the methodology and identification of key issues
for future analysis.

1 Introduction
Law is an explicit system of rules to govern the behaviour of people. Legal prac-
titioners must learn to apply legal knowledge to the facts at hand. The United
States Multistate Bar Exam (MBE) is a professional test of legal knowledge,
where passing indicates that the examinee understands how to apply the law.
As such, the MBE provides a baseline for measuring the performance of a legal
question answering system. We are primarily interested in modeling legal rea-
soning, wherein we reason from rules and facts to conclusions as well as provide
explanations of reasoning. Such explicit modeling of legal reasoning is essential
to legal practice, e.g. for appeals or development of the law. This paper proposes
an automated question-answering mechanism (expert system) to answer legal
questions of the United States Multistate Bar Exam (MBE).
The questions we used in our work are actual MBE questions published
through the National Conference of Bar Examiners (NCBE). NCBE publishes
these questions, multiple choice answers, and the correct “Gold Standard”
answers. The example below shows question 7 from the July 1998 MBE, where
answer “b.” is correct.
The original MBE questions, which have four possible answers, were reorga-
nized so that the main body of the text represents all the “background” knowl-
edge from which we try to infer the answer1 . From original given answer, the
1
This is an abbreviated discussion. See [10] for what was done and why.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 317–324, 2018.
https://doi.org/10.1007/978-3-030-00178-0_21
318 B. Fawei et al.

part representing the rationale for either ‘acquittal’ or ‘conviction’ appears in the
background knowledge (see sentence in bold below); otherwise, the background
portions are the same. So, for example, from the background information in
(a.), we should not infer “Mel should be acquitted.”, while we should for (b.) we
should.
a. After being fired from his job, Mel drank almost a quart of vodka and decided
to ride the bus home. While on the bus, he saw a briefcase he mistakenly
thought was his own, and began struggling with the passenger carrying the
briefcase. Mel knocked the passenger to the floor, took the briefcase, and fled.
Mel was arrested and charged with robbery. He used no threats and was
intoxicated .
Mel should be acquitted.
b. After being fired from his job, Mel drank almost a quart of vodka and decided to
ride the bus home. While on the bus, he saw a briefcase he mistakenly thought
was his own, and began struggling with the passenger carrying the briefcase.
Mel knocked the passenger to the floor, took the briefcase, and fled. Mel was
arrested and charged with robbery. His mistake negated the required spe-
cific intent.
Mel should be acquitted.
The issue that we address is to extract information from the background and
answer textual passages, associate the information with our knowledge base, and
trigger rules to reason to the correct answer. Informally, we must understand
and apply: the commonsense implications of violent acts, body-part relations,
and possession; in addition, such facts must be tied to legal rules bearing on
forced transfer of possession as well as consent. While providing solutions to
such examples is difficult, we work with actual MBE questions to construct
constrained models, which can be incrementally developed.
The novelty of the paper, which is preliminary work, is the integration of
three main modules for modeling legal information: legal text annotation, legal
ontology instantiation, and the application of legal rules. Related work is briefly
reviewed in Sect. 2, followed by an outline of our approach (Sect. 3), results
(Sect. 4), and closes with some discussion (Sect. 5).

2 Related Work
Existing legal question answering tools retrieve articles, extract chunks of infor-
mation, and compare the retrieved information to the question in order to deter-
mine entailment [5]. A two layered approach for textual entailment was imple-
mented [2], which works with Japanese Bar Exam data. However, the questions
do not require legal reasoning and the approach is limited to handling issues
arising from complex constraints in statute conditions. Kim et al. [1] seek rele-
vant background information to facilitate inference in the questions by applying
TF-IDF and a SVM to retrieve texts t1 relevant to a query t2, then measuring
 Using Legal Ontologies with Rules for Legal Textual Entailment 319

the similarity between t1 and t2 using paraphrase features and word embedding.
However, the structure of the data used in these experiments is quite different
from the bar exam considered in this research. Do et al. [9] applied a combina-
tion of SVM ranking and Convolution Neural Network techniques for information
retrieval and legal question answering respectively.
There has been intensive research on textual entailment. Yet, these
approaches do not address legal question answering. A pairwise syntactic similar-
ity measure has been implemented in [8]. A dependency based paraphrasing was
adopted in [7] for textual entailment. Arya et al. [6] implemented a knowledge
base approach with different lexical resources to provide semantic and struc-
tural information for determining entailment. IBM DeepQA applies algorithms
to identify answers for questions from both structured and unstructured sources
of information [4].
Attention mechanism systems are based on sentence encoding [14,15] in
which the decoder determines the part of sentence to focus on. They apply
the Long Short-Term Memory (LSTM) networks to embed the premises and
hypothesis into same vector space. Though, this approach provides more inter-
play between the embedded sentences, it involves deep sentence modeling, which
require excessive training parameters and input of very long sequences [12]. These
approaches lack the sort of legal knowledge and reasoning required for answering
the MBE. Deciding entailment in this case requires the application of some legal
rules residing in a knowledge base and some contextual information from the
problem domain in answering such questions.

3 Approach
Broadly, in our approach, convergent tacks are taken to justify conclusions. For
the first tack, the tool first extracts semantic triples from the source text, which
are used to instantiate an OWL ontology [13] (see Subect. 3.1). Such instanti-
ations are used to ground predicates in the SWRL rules. Where a rule has all
premises grounded with respect to the ontology, we can forward-chain to draw
intermediate or final conclusions. For the second tack, keywords are used to
identify relevant rules, which are those where the keyword matches the predi-
cate of the rules conclusion (see Subsect. 3.2). Where a relevant rule is identi-
fied, backwards-chaining can be triggered, leading to searches for information
to ground the premises. The two tacks can be interleaved until a justified con-
clusion is attained (a prediction of an answer); performance is evaluated with
respect to the number of true positives versus false positives. Figure 1 shows the
workflow used for the instantiation of the ontology and application of rules. In
the following, we outline each of the subcomponents.

3.1 Text Processing, Keywords, Triples, and Ontology Instantiation

We apply standard NLP tools [3] of tokenisation, lemmatisation, and gazetteer

lookup to both the Bg and Q texts. The gazetteer is a list of relevant legal
320 B. Fawei et al.

Fig. 1. Instantiating ontology from text

terminology. If a token in the text matches against the gazetteer, then it is

extracted as a legal keyword. In our initial MBE example, Q contains two legal
keywords commit and robbery. These keywords are used to filter triples as below.
In addition, we do part-of-speech annotation and parsing, followed by triple
extraction, which are used to instantiate the OWL ontology, which consists
of about 90 classes and 130 properties, represented in the ALCHO descrip-
tion logic. Triples are a collection of three entities that represent a statement
derived from the textual information and which are expressed as s-p-o (Subject-
Predicate-Object). Semantic triples are important because they associate with
the concepts and object-properties in the ontology, where the Subject and Object
of the triples associate with concepts, and the predicate associates with an object-
property. Triple extraction is done using the Stanford CoreNLP API package [3].
For example, from the sentence “Mel knocked the passenger to the floor”, we
can extract a triple with s = Mel, p = knock, and o = passenger.
However, the tool over-generates triples, that is, triples are created which
are not relevant for the purpose of legal decision making. For example, other
triples that are extracted are: [s = Mel, p = be, o = fire], [s = Mel, p = knock,
o = passenger], [s = Mel, p = take, o = briefcase], [s = Mel; p = drink, o = quart],
[s = mistake, p = negate, o = intent] etc. The first triple [s = Mel, p = be, o = fire]
is not relevant. Hence, we remove irrelevant triples by comparing them to the
keywords as well as the to the classes and object/data properties of the ontology.
Where a match to a keyword or an instance or associated string is found, the
relation is identified as a relevant triple for populating the ontology. For example,
[s = Mel, p = take, o = briefcase] is a relevant triple because it matches the onto-
logical relation [class = Person, objectproperty = take property, class = Material],
where “Mel” is an instance of Person, “briefcase” is an instance of Material, and
“take” matches the string attribute of the object property take property.
Furthermore, as extracted terminology may vary with respect to the ontol-
ogy, we apply WordNet and NER, searching for relevant terminology that is
lexically related; e.g., “knock” appears in the text and the ontological term is
“forced”, then we search in WordNet whether there is a relevant path (measured
by similarity in the graph) connecting the terms. The threshold of similarity
measure used is between 0.7 and 1.0 on a scale between 0 and 1, using the
 Using Legal Ontologies with Rules for Legal Textual Entailment 321

Wu-Palmer path similarity (Wu and Palmer, 1994). Terminology that is similar
may be used to match against the ontology.

3.2 Semantic Web Rule Language, Legal Rule Acquisition,

Identification, and Execution
In this section, we briefly describe the Semantic Web Rule Language (SWRL)
that is used in conjunction with the ontology and give an overview of how the
rules are acquired, along with examples. We then explain how keywords are
used to identify potentially relevant rules to execute followed by description of
the execution of rules.
SWRL rules have a high-level abstract syntax to express Horn-like rules with
an antecedent and a consequent. The rules are expressed in terms of OWL classes,
properties, and individuals.
The legal SWRL rules that are developed in the approach are manually cap-
tured from criminal law and procedure rules found in bar examination prepara-
tory material and in consultation with domain experts. Starting from a statement
such as
Robbery is when property is taken from the person or presence of the owner
and the taking is accomplished by using force or putting the owner in fear.
[11, p. 192].
we derive a rule, where the variables such as ?x become grounded:
own property(?x, ?y) ∧ forced person(?z, ?x) ∧ take property(?z, ?y) ∧
differentF rom(?x, ?z) → has committedRobbery(?z)
The legal rules were manually created at this initial stage of development and,
currently, we have about 20 criminal law rules.
Rules are activated either by ontological terminology that has been instan-
tiated or by keywords. The search starts with legal keywords from the question
statement. If the search is not successful, then keywords from the background
information are combined with keywords in the question statement to search
for the rule. For example, from “Mel should be acquitted”, we extract the legal
keyword acquitted, which can be used to trigger a more general rule, where
should be acquitted is associated with the string acquitted.
has committedRobbery(?x)→has committed(?x, robbery) ∧ Robbery(robbery)
has committed(?x, ?y) ∧ did not intend(?x, ?y) ∧ Crime(?y)
→ should be acquitted of (?x, ?y)

Forward chaining takes place if the expression representing the antecedent

of a rule unifies with the terminological content of our legal ontology. Backward
chaining takes place if the conclusion part of the rule unifies with the legal
ontology. In that case, the antecedents are checked: if all antecedents also unify
with the ontology, then the conclusion is accepted. If, however, some of the
antecedents need to be checked, the tool returns to the source text to identify
322 B. Fawei et al.

possible further useful information – the tool incrementally and “interactively”

searches for an answer using as resources the MBE text itself, as well as the
ontology, and the extracted legal rules.
With this implementation, the system searches in the legal knowledge base
for criminal law rules that match legal keywords. Given instantiated object prop-
erties, a candidate legal rule can be:

own property(passenger, brief case), f orced person(M el, passenger),

take property(M el, brief case) and did not intend(M el, robbery)

and since Mel and passenger are distinct, the system infers that Mel has com-
mitted robbery, which is a crime, as we have Robbery  Crime in the ontology.
When this new knowledge of Mel has committed robbery is applied to the second
rule, the system infers that Mel should be acquitted because the action was not
an intended action. The axiom did not intend(?x, ?y) captured the information
(s = mistake, p = negate, o = intent) at the rule level.

4 Results
Given 15 MBE questions, each with four possible answers, we have 60 question–
answer pairs. Of these pairs, we found 15 true positives, 28 true negatives, and
17 false positives, giving a precision of 0.46, recall of 1, accuracy of 0.71, and an
F1 measure of 0.63.
Examining the false positives, error analysis highlighted several problems: (1)
polysemy, e.g. words such as own, hold, possess, are problematic; (2) We found
examples where named entities are not recognised. For example, a prosecutor is
a person. This problem could be solved using a legal named–entity recognizer;
(3) Complex Compound Nouns are syntactically complex elements which are
not recognised, e.g. wrist watch. Such compounds were made into sinlge lexical
items and saved in the gazetteer.

5 Discussion

This paper describes an initial implementation of a rule-based legal question

answering system to identify the correct answer to classic MBE exam ques-
tions. We used a rule-based approach and developed a question-answering legal
expert system consisting of three modules implementing the functionality for
legal text annotation, legal ontology instantiation, and the application of legal
rules. In the future, we plan to introduce more legal knowledge, augment an
NER tool adapted to the legal domain, and incorporate a large-scale common-
sense knowledge base such as, for example, YAGO. Our system is still in an early
development stage and our development approach is incremental. Nonetheless,
we can identify patterns of legal reasoning, which we believe can be abstracted
to facilitate further question-answering.
 Using Legal Ontologies with Rules for Legal Textual Entailment 323

References
1. Kim, M.Y., Xu, Y., Lu, Y., Goebel, R.: Legal question answering using paraphras-
ing and entailment analysis. In: Tenth International Workshop on Juris-Informatics
(JURISIN) (2016)
2. Kim, M., Goebel, R.: Two-step cascaded textual entailment for legal bar exam
question answering. In: Proceedings of the International Conference on Artificial
Intelligence and Law (2017)
3. Manning, C.D., Surdeanu, M., Bauer, J., Finkel, J.R., Bethard, S., McClosky,
D.: The Stanford CoreNLP natural language processing toolkit. In: ACL (System
Demonstrations), pp 55–60 (2014)
4. Ferrucci, D., Levas, A., Bagchi, S., Gondek, D., Mueller, E.T.: Watson: beyond
jeopardy!. Artif. Intell. 199–200, 93–105 (2013)
5. Monroy, A., Calvo, H., Gelbukh, A.: NLP for shallow question answering of legal
documents using graphs. In: Gelbukh, A. (ed.) CICLing 2009. LNCS, vol. 5449, pp.
498–508. Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-642-00382-
0 40
6. Arya, A., Yaligar, V., Prabhu, R.D., Reddy, R., Acharaya, R.: A knowledge based
approach for recognizing textual entailment for natural language inference using
data mining. Int. J. Comput. Sci. Eng. 2(06), 2133–2140 (2010)
7. Marsi, E., Krahmer, E., Bosma, W.: Dependency-based paraphrasing for recog-
nizing textual entailment. In: Proceedings of the ACL-PASCAL WS on Textual
Entailment and Paraphrasing, pp. 83–88 (2007)
8. Zanzotto, F.M., Moschitti, A., Pennacchiotti, M., Pazienza, M.T.: Learning textual
entailment from examples. In: Proceedings of the Second PASCAL Challenges
Workshop on Recognising Textual Entailment, vol. 6, no. 09, pp. 50–55 (2006)
9. Do, P.K., Nguyen, H.T., Tran, C.X., Nguyen, M.T., Nguyen, M.L.: Legal question
answering using ranking SVM and deep convolutional neural network. In: Tenth
International Workshop on Juris-Informatics (JURISIN) (2017)
10. Fawei, B., Wyner, A.Z., Pan, J.Z.: Passing a USA national bar exam: a first corpus
for experimentation. In: Tenth International Conference on Language Resources
and Evaluation, LREC 2016 (2016)
11. Emmanuel, S.L.: Strategies and Tactics for the MBE (Multistate Bar Exam).
Wolters Kluwer, New York (2011)
12. Liu, P., Qiu, X., Huang, X.: Modelling interaction of sentence pair with coupled-
LSTMs. arXiv preprint arXiv:1605.09090 (2016)
13. Pan, J.Z.: A flexible ontology reasoning architecture for the semantic web. IEEE
Trans. Knowl. Data Eng. 19(2), 246–260 (2007)
14. Chen, Q., Zhu, X., Ling, Z.H., Wei, S., Jiang, H., Inkpen, D.: Enhanced LSTM for
natural language inference. In: Proceedings of the Second Workshop on Evaluating
Vector Space Representations for NLP (RepEval 2017) (2017)
324 B. Fawei et al.

15. Kolawole John, A., Di Caro, L., Robaldo, L., Boella, G.: Textual inference with
tree-structured LSTM. In: Bosse, T., Bredeweg, B. (eds.) BNAIC 2016. CCIS, vol.
765, pp. 17–31. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-67468-
12
16. Wu, Z., Palmer, M.: Verbs semantics and lexical selection. In: Proceedings of the
32nd annual meeting on Association for Computational Linguistics, pp. 133–138
(1994)
 KR4IPLaw Judgment Miner - Case-Law
Mining for Legal Norm Annotation

Shashishekar Ramakrishna1(&), Łukasz Górski2, and Adrian Paschke1

1
Freie Universität Berlin, Berlin, Germany
shashi792@gmail.com, paschke@inf.fu-berlin.de
2
Nicolaus Copernicus University, Toruń, Poland
lgorski@mat.umk.pl

Abstract. The use of pragmatics in applying the law is hard to deal with for a
legal knowledge engineer who needs to model it in a precise KR for (semi-)
automated legal reasoning systems. The negative aspects of pragmatics is due to
the difficulty involved in separating their concerns. When representing a legal
norm for (semi-)automated reasoning, an important step/aspect is the annotation
of legal sections under consideration. Annotation in the context of this paper
refers to identification, segregation and thereafter representation of the content
and its associated context. In this paper we present an approach and provide a
proof-of-concept implementation for automatizing the process of identifying the
most relevant judgment pertaining to a legal section and further transforming
them into a formal representation format. The annotated legal section and its
related judgments can then be mapped into a decision model for further down
the line processing.

Keywords: LegalDocML
Case-law mining
Legal norms
Topic modeling

1 Introduction

In the domain of legal informatics there exists a necessity to ‘simplify’ the induced
vagueness in legal language. An approach to minimize the vagueness, or in other words
to simplify the language, is by annotating of a legal section with its associated con-
textual background knowledge. Aggregation of such associated background knowledge
removes many uncertainties during legal norm reasoning.
Annotation in the context of this paper refers to identification, segregation and
thereafter representation of the legal content and its associated context.
In this paper, we present an approach and provide a proof-of-concept implemen-
tation for automatizing the identification of most relevant judgment pertaining to a
particular legal section (in the domain of intellectual property) and further transforming
them into a formal representation format. Thereby, the annotated legal section and the
related judgments can be mapped into a decision model for further processing.
The paper is structured as follows: Sect. 2 provides a brief overview on research
works related to our proposed approach. Section 3 discusses the notion of pragmatics
and presents the important aspects of legal norm annotation. Sections 4 and 5 introduce
to the overall system and specifically the new proof-of-concept implementation module

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 325–336, 2018.
https://doi.org/10.1007/978-3-030-00178-0_22
326 S. Ramakrishna et al.

for mining judgments. Sections 6 and 7 provide a detailed discussion on building a

decision model and its further use. Section 8 evaluates the module based on well know
information retrieval evaluation approach based on gold standards. Final Sect. 9 pre-
sents the conclusion and future works.

2 Related Work

A number of legal document annotation methods exist, including LegalRuleML [1, 2],
LKIF [3] and LegalDocML [4], which stems from earlier Akoma Ntoso project [5].
LegalRuleML is an extension of RuleML [6] standard, aimed at enabling the exchange
of legal semantic information (defined in laws, contracts, judgments) between legal
documents, business rules and software applications. LKIF (Legal Knowledge Inter-
change Format) was designed to support the requirements of Semantic Web and
Knowledge Representation domains. It supports a layered architecture, with distinct
modules for terminological knowledge description (implemented with the use of Web
Ontology Language - OWL) and rules (the extension of SWRL with negation of
defeasible reasoning). The LegalDocML project describes a metadata format for doc-
umentation of parliamentary activities, such as debates, briefs, journals, as well as
courts’ decisions (opinions, judgments). It provides with a common data and metadata
model that supports the temporal evolution of legal documents.
The research into text retrieval techniques (that match a set of documents against a
user query) is vital for processing the increasing amount of electronic information [7].
There are, generally, two approaches to information retrieval in legal domain: manual
knowledge engineering and those based on automated text analysis (natural language
processing, NLP) [8]. Both techniques were used in the History Assistant project,
aimed at finding interconnected judicial opinions [9]. Text mining was employed, for
example, for legal argument extraction [10, 11]. Techniques similar to ours were used
to build a law report recommendation system in the case of Sri Lankan judiciary [12].
The SMILE project is another project concerned with NLP, as it allowed to classify
legal texts based on a set of classification concepts [13].

3 Legal Norm Annotation

3.1 Introduction- Pragmatics

3.1.1 What Is the Relationship Between Case-Law and Statute Law?
Even though the ‘clarity’ of a legal provision is an idealistic aim that legislators are
bound to achieve [14], it cannot be denied that the professional members of particular
legal community - through their training and practical experience - tend to share a
greater deal of particular contextual knowledge than ordinary layman [15]. Judgments
provide a body of contextual knowledge that enhances the understanding of legal
section under consideration. Therefore, when viewed in the lens of the theory of
pragmatics, they enhance the meaning of legal norms. The inclusion of pragmatics,
either directly from a legal section or from its interpretations provided via judgments
 KR4IPLaw Judgment Miner - Case-Law Mining 327

therefore plays a very important role in capturing the interpretation concerning legal
norms.
Despite the fact that both the courts and non-judiciary legal professionals share and
adhere to the same legal interpretation guidelines (and could have been expected to
easily arrive at the same result when interpreting legal provisions), the study of court
opinions is a vital part of professional legal training, both in common law and civil law
countries. The importance of contextual knowledge formalization has also been
acknowledged in the area of legal informatics, with the development of the Ontology of
Professional Judicial Knowledge (OPJK), a FAQ system for young judges [16]. Some
legal theorists have as well viewed the process of applying the law (e.g.
inventions/patent applications) as a non-verbal conversation between the legislature
and members of judicature and executive (e.g. patent officer/examiner), and further -
between a citizen and executive (e.g. the inventor and the patent office/examiner). Such
conversation is guided by a set of conversational maxims, which - while obeyed or
deliberately flouted - allow to decode the meaning of the utterance. The philosophical
aspects of application of the theory of pragmatics in the domain of law are out of the
scope of this paper. A philosophical framework addressing this concern was developed
in other works [15, 17]. In the case of a semi-formal legal norm representation, the
pragmatic aspect is also accounted for, at least in some of representation systems, e.g.
the SBVR (cf. Sect. 7). This linguistic-based standard acknowledges the separation of
expression and meaning: the semantic expression is enriched by relevant context, thus
producing meaning (cf. [18]).
In this paper, we acknowledge that the semi-automatic annotation of legal sections
facilitates the knowledge transfer even to people who are not as skilled in the domain of
law (e.g. knowledge modelers).

3.2 Norm Annotation

Akoma Ntoso or Legal Document Markup Language (LegalDocML), a standard under
OASIS for managing legal documents, provides a systematic mechanism for refer-
encing documents based on URIs through sound ontological approach by designing
metadata and relationships between documents and different versions of documents. Its
provides a possibility to define a common format for most parliamentary activities. Of
the possible document type handled by LegalDocML, few document types which are of
our interest in domain of IP Laws are, <legislativeDocs>, comprising of ‘acts’ and
‘bills’, <amendmentDocs> related to amendments to any acts/bills/sections etc.,
and <judgementDocs>, pertaining to judgments, amicus briefs and other documents
pertaining to any parliamentary activities concerning a law.
In addition to capturing (by annotation) the legal information from legal
sections/paragraphs, we also annotate the landmark decisions pertaining to this legal
section. Those supplementary annotations capture the pragmatic context in which such
a legal section has to be applied. I.o.w, it defines new pragmatics, in terms of under-
standing and commitment, encompassing the legal section.
For the purpose of this paper, we focus only on the elements <act> and <judg-
ment>. Further hierarchical subdivision of these elements into several classes is shown
in Fig. 1. The <meta> element is used to capture all the meta-information concerning a
328 S. Ramakrishna et al.

Fig. 1. ‘Judgement’ and ‘Act’ element hierarchical structure in LegalDocML

‘act’ or ‘judgment’, the <preface>, defines document type (<docType>), document

title (<docTitle>), document number (<docNumber>), date of assent (<docDate>),
table of contents (<toc>). The <preamble> element defines ‘citations’ i.e. list of other
legislation resources that empower the current document and recitals, that provide
the justification and motivation. Finally the <Conclusions> element describes the
‘Signature’, ‘date of signature’, ‘place of signature’ etc. The <judgment> element,
comprises of self descriptive sub-elements such as meta-information (<meta>), content
of the judgment (<judgementBody>), verdict (<conclusions>), attachments pertaining
to the judgment (<attachments>) etc. Identification and further annotation of such
judgments while being of utmost importance, is a very time consuming process and
requires digging through a sizable corpora of judgments. In pursue of automating the
process we provide a proof-of-concept implementation, which mines for the most
relevant judgments pertaining to the legal sections and further renders the selected
judgments in a required formal representation format (in here, as LegalDocML).

4 KR4IPLaw System

Knowledge Representation for Intellectual Property Law (KR4IPLaw),1 is an system

(ecosystem) with modules built to bridging the gap between legal practitioners and
knowledge modelers. The system is built on the KR4IPLaw framework, which in-turn
loosely coupled with the OMGs Model Driven Architecture [19]. The current systems
includes modules for representing legal knowledge at different level of granularity
(Computational Independent, Platform Independent and Platform Specific). The system
includes an inference engine Prova [20] to support reasoning on top of such represented
legal knowledge. Further sections of this paper describes the functionalities of a newly
added sub-module, the KR4IPLaw Judgment Miner. It is a tool to aimed at capturing
the a priori information needed to build the decision model thereafter.

1
http://kr4iplaw.wordpress.com.
 KR4IPLaw Judgment Miner - Case-Law Mining 329

5 KR4IPLaw Judgement Miner

The proof of concept system, as a simple I/O system, considers formally represented
legal sections as input and produces the most relevant judgment or a set of judgments
pertaining to the legal section. Whilst the current system is tailored to work with
LegalDocML files, it can be very easily extended to other formal representations.
The inner stack of the module comprises of few translators, a search engine, a topic
modeler, and query constructors. Translators help in extracting the required information
from a LegalDocML file and thereby translating it to a format necessary for further
processing and vice-versa. We integrate Apache Solr as a search engine - a high
performance, scalable search server built based on Apache Lucene. The advanced
caching and replication, the distributed search and easy integration of other modules
makes it a better choice than existing search servers such as Minion [21] or Sphinx2.
Case-law indexing is one of the important steps. CourtListner,3 provides as a bulk
download, the information pertaining to 361 jurisdictions of the United States courts.
The data on CourtListener is a combination of many partial sources such as court
websites, Public.Resource.Org and a donation from Law- box LLC, thus making it one
of the most covered dataset available. The bulk data from the CourtListner is indexed
by Solr.
For Solr scoring model, we use the Best Matching (BM25) algorithm [22] - a
probabilistic Information Retrieval (IR) model against the well known TF-IDF - a
vector space model for increased precision.
Solr provides the ability to create different user defined queries through its API.
Solr query support different search patterns such as term, field, wildcard, fuzzy,
proximity or range searches. Table 1 provides the list of search patterns available
within this module.
To increase the accuracy of the search results, an obvious step would be to con-
struct complex queries, which combine multiple patterns. One such pattern is the query
boosting parameter which utilizes certain keywords to boost the search query and
thereby altering its default ranking.
While such process of identifying keywords could be a manual operation, our
module uses the topic model approach to automate the process of extracting and
transforming the keywords, to be used as boosting terms. We integrate Mallets’4 sta-
tistical natural language processing (NLP) and topic modeling modules. Within the
NLP module, we integrate the Snowball stemmer [23] and provide legal dictionaries
(e.g. USPTO- glossary5) as an exclusion list. As to the Topic Modeler module, we use
the Parallel Topic Modeler [24], which realizes the Latent Dirichlet Allocation
(LDA) model to compute the topics. The input to the topic modeler is the content part
(i.e. legal paragraphs) of the LegalDocML file and we assume that the each document
is composed of only a few dominant topics and each topic is composed of only a few

2
http://sphinxsearch.com/.
3
http://www.courtlistener.com/.
4
http://mallet.cs.umass.edu/.
5
http://www.uspto.gov/main/glossary/.
330 S. Ramakrishna et al.

Table 1. KR4IPLaw search query patterns

Type Pattern Example
Term “text” “Supreme Court”
Field text:“text” court:“Supreme Court”
Boolean AND, OR, NOT “Supreme Court” OR “CAFC”
Wildcard ? or * Supre?me or Supreme*
Fuzzy * Supreme*
Range TO filed date: [20150601 TO 200150801]
Boosting ˆ “Supreme Court”ˆ10

dominant words. We set the hyper-parameters, a and b values to 0.01. The output from
the topic modeller is used to build the complex query described hereinbefore. Addi-
tionally, the most frequently occurring terms from particular domains of law (i.e. patent
in the case of patent law) were explicitly excluded from boosting parameters list, as
they are routinely used in judgments and therefore are of little use for ranking purposes.
The purpose built-in translator thereafter translates the obtained judgments from its
XML-based native CourtListener format to a required formal judgment representation
format (e.g. LegalDocML; in this case the conversion is realised with XSLT). With the
annotation of legal sections and its relevant judgments completed, the next step is to
integrate them into a decision model, so that further disaggregation and formal rep-
resentation and reasoning can be performed.

6 Decision Model

The LegalDocML files representing the substantive laws and its relevant judgments are
treated as a starting point for the decision model creation. The legal section is disag-
gregated from its vague substantive law semantics into a concrete procedural (norm)
semantics, to extract the elementary concerns from the compound concerns of the
statutory law.
Such procedural norms are usually provided in the forms of memos, instructions or
guidelines and contain supplementary material pertaining to the interpretation of sub-
stantive laws and outline the procedure that is used for substantive law implementation.
For example, in our earlier works, we have used the US patent law, as illustrative
material in this respect, as the United States Patent and Trademark Office (USPTO),
through its Manual for Patent Examination Procedure (MPEP), provides its examiners
with such information. We transform such procedures into legal decision models,
wherein, each decision point is a single procedure or a set of procedures to be carried
out. The system provides a decision model representation module that utilizes UML
activity diagram formalisms. Landmark decisions provide an additional means of
interpretation, further supplementing the aforementioned guidelines. Hence their
incorporation into decision model is of great importance.
 KR4IPLaw Judgment Miner - Case-Law Mining 331

Fig. 2. Decision model for a legal section under consideration (adapted from [25]).

Figure 26, serves as an illustrative example of the above described approach of

representing the procedural norms and their interpretations provided via judgments as
decision models.

7 Representation and Reasoning

Further, we use the easy to understand decision models as basis for writing the legal
norms and their elementary concerns in terms of constitutive vocabulary definitions and
prescriptive behavioral legal rules in SBVRs Structured English. We can classify the
mapping relationships as 1:1- wherein each decision is mapped into a single SBVR
rule, 1:M- where, a single decision is mapped into many SBVR rules or an M:M
relationship. Legal domain experts and trained formal knowledge engineers can work
together in this formalization process using Structured English as common computa-
tional independent knowledge representation language.

6
The textual content inside the decision model is left out on purpose to handle the space restrictions.
332 S. Ramakrishna et al.

Consult the following partial representation of legal (procedural) norm from the
decision point ‘D’ (from Fig. 2) for an illustratory example:
– It is obligatory thattheoffice action includes Paragraph 7 33 01,ifclaim
is rejected under essential subject matter requirement.

Single decision point represents only a small part of reasoning that is performed
when evaluating and applying legal norms. More complete representation is given by
the whole decision tree. Yet, obviously, it is applicable rather in standard, non-hard
cases. Further work might employ using case-based reasoning methods for precedent
representation (in line with HYPO/CATO line of research [26]), enhancing the flexi-
bility of the system and allowing for its use in a wider range of situations.
This example is complemented by a vocabulary defining used concepts, like
essential subject matter requirement noun concept. The semi-formal procedural
rules are built on legal facts, and legal facts are built on legal concepts7 which are
expressed by legal terms. Annotation of legal concept with the meta-data description
further enables identification and representation of associated context information.
A more in-depth discussions pertaining to SBVR and its use in legal domain has been
presented by us in [27, 28].
Further, for formal rule representation, we use KR4IPLaw - a patent norm repre-
sentation format, which seamlessly integrates into the existing rule representation
standards like RuleML [29], ReactionRuleML [30] and LegalRuleML [1]. Figure 3
depicts the general structure of KR4IPLaw.
The module <rulePragmatics> holds all the pragmatic information which includes
meta-information such as; <Sources> to identify the collection of legal resources rel-
evant to the Legal document, <References> to provide an isomorphic relationship

Fig. 3. KR4IPLaw: rule representation format.

7
Represented using green color.
 KR4IPLaw Judgment Miner - Case-Law Mining 333

between the formal rule and the legally binding statements to the rule, <Author-
ity>, <Jurisdictions> , <Associations> , <TimeInstants> etc. While the general struc-
ture provided here depicts the intended semantics to the formal representation formats,
the authors would like to refer to a series of publications for a detailed discussion on it
[31–33].
Further, for reasoning, the formal rule is transformed into a Platform Specific
Model (PSM) rule representation format, Prova [20] - a semantic web rule language
and a high expressive distributed rule engine- where the pragmatic information from
the PIM layer, associated to a norm are handled with by guards.

8 Evaluation

8.1 Gold Standard

The proposed approach is evaluated on the well known precise-recall metric of infor-
mation retrieval. An ordered set of judgments based on its relevance (i.e. on case-law
citations8) to a legal section under consideration was created (a gold-standard or ground
truth judgment of relevance)9. To evaluate the system for its cross domain performance,
the gold standard included a minimum of 5 legal sections, each from different Intel-
lectual Property (IP) Laws of the Code of Laws of the United States of America (U.S.C).
As shown in Table 2, the legal sections from US Patent Law, Copyright law and
Trademark law were considered. 20 most relevant, based on their number of citations
pertaining to each legal section were identified. Relevance based on a citation count
approach providing an easy-to-compute, context-free, and relatively accessible metric,
may, in fact, be better for finding high-quality documents than sophisticated human or
pattern recognition queries and models [34]. In total, the gold standard comprised of 300
judgments from courts of 361 jurisdictions within the US.

8.2 Precision and Recall

F-measure metric (i.e. the harmonic mean of the precision and recal) l is used for the
results depiction. The F-measure for the judgments retrieved from the judgment-miner
module is as shown below in Fig. 4. The F-measure is plotted for legal sections from
different domains of the intellectual property law.
While vocabularies from US patent, trademark and copyright laws were introduced
as an exception list for stemming, an observation made during this study was the
necessity to perform a proximity search while indexing. This has been due to the fact
that judges responsible for writing decisions use a slightly non-aligned nomenclature
such as 35 USC 112 (a), 35 USC 112 First Paragraph or 35 USC sec 112 while
referencing legal sections within a judgment.

8
While the authors understand that an ideal approach is to use a expert driven gold standard
construction, a common consensus on thus generated standards for relevance is debatable in the
context of case-laws.
9
http://github.com/shashi792/KR4IPLaw-Act2Judgement/KR4IPLaw_Gold_Standard.xlsx.
334 S. Ramakrishna et al.

Table 2. Legal sections considered for the gold standard

Sec 35 U.S.C (patent law) 17 U.S.C (copyright law) 15 U.S.C (trademark law)
I 101 (inventions patentable) 107 (fair use) 1052 (registrable trademarks)
II 102 (novelty) 117 (computer programs) 1060 (assignment)
III 103 (non-obviousness) 201 (ownership) 1063 (opposition)
IV 112 (specification) 501 (infringement) 1070 (appeals)
V 271 (infringement) 1101 (unauthorized fixation) 1091 (supplemental register)

Fig. 4. F-measure (legal sections under consideration are as listed in Table 2)

While the use of boosting keywords resulted in significant result improvement in

the domain of patent law, the outcome for copyright and trademark law remains
satisfactory. Further improvement of results could be achieved by fine-tuning the
proximity search method in case of those two domains.

9 Conclusion

The paper presented an approach and a proof of concept implementation for automa-
tizing the process of identifying the most relevant judgment pertaining to a legal section
and further transforming them into a formal representation format. To fine tune the
retrieved judgments sorting order, we proposed an approach wherein topic modeling
using LDA was used. Further these judgments were transformed into a formal repre-
sentation format using a purpose built translator. In effect, any legal decision support
system can facilitate the provided a priori information for its use in scenarios such as
in-court argumentation.
 KR4IPLaw Judgment Miner - Case-Law Mining 335

References
1. Palmirani, M., Governatori, G., Rotolo, A., Tabet, S., Boley, H., Paschke, A.: LegalRuleML:
XML-based rules and norms. In: Olken, F., Palmirani, M., Sottara, D. (eds.) RuleML 2011.
LNCS, vol. 7018, pp. 298–312. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-
642-24908-2_30
2. Athan, T., Boley, H., Governatori, G., Palmirani, M., Paschke, A., Wyner, A.: OASIS
LegalRuleML. In: Proceedings of the Fourteenth International Conference on Artificial
Intelligence and Law, pp. 3–12. ACM (2013)
3. Hoekstra, R., et al.: The LKIF core ontology of basic legal concepts. LOAIT 321, 43–63
(2007)
4. Palmirani, M., Vitali, F.: Akoma Ntoso an open document standard for parliaments (2014)
5. Vitali, F., Zeni, F.: Towards a country-independent data format: the Akoma Ntoso
experience. In: Proceedings of the V legislative XML Workshop, Florence, Italy, pp. 67–86.
European Press Academic Publishing (2007)
6. Boley, H., et al.: Design rationale for RuleML: a markup language for semantic web rules.
In: SWWS, vol. 1, pp. 381–401 (2001)
7. Lee, D.L., Chuang, H., Seamons, K.: Document ranking and the vector-space model. IEEE
Softw. 14(2), 67–75 (1997)
8. Maxwell, K.T., Schafer, B.: Concept and context in legal information retrieval. In: JURIX,
pp. 63–72 (2008)
9. Jackson, P., Al-Kofahi, K., Tyrrell, A., Vachher, A.: Information extraction from case law
and retrieval of prior cases. Artif. Intell. 150(1), 239–290 (2003)
10. Wyner, A., Mochales-Palau, R., Moens, M.-F., Milward, D.: Approaches to text mining
arguments from legal cases. In: Francesconi, E., Montemagni, S., Peters, W., Tiscornia, D.
(eds.) Semantic Processing of Legal Texts. LNCS (LNAI), vol. 6036, pp. 60–79. Springer,
Heidelberg (2010). https://doi.org/10.1007/978-3-642-12837-0_4
11. Ashley, K.D., Walker, V.R.: From information retrieval (IR) to argument retrieval (AR) for
legal cases: report on a baseline study. In: Legal Knowledge and Information Systems. IOS
Press (2013)
12. Firdhous, M.: Automating legal research through data mining. arXiv preprint arXiv:1211.
1861 (2012)
13. Ashley, K., Brninghaus, S.: Automatically classifying case texts and predicting outcomes.
Artif. Intell. Law 17(2), 125–165 (2009)
14. Fuller, L.L.: The Morality of Law, vol. 152. Yale University Press, New Haven (1977)
15. Marmor, A.: The pragmatics of legal language. Ratio Juris 21(4), 423–452 (2008)
16. Benjamins, V.R., Contreras, J., Casanovas, P., Ayuso, M., Becue, M., Lemus, L., Urios, C.:
Ontologies of professional legal knowledge as the basis for intelligent it support for judges.
Artif. Intell. Law 12(4), 359–378 (2004)
17. Ramakrishna, S., Gorski, L., Paschke, A.: The role of pragmatics in legal norm
representation. CoRR abs/1507.02086 (2015)
18. OMG: Semantics of Business Vocabulary and Business Rules (SBVR) v. 1.3 (2015)
19. Bézivin, J., Gerbé, O.: Towards a precise definition of the OMG/MDA framework. In:
Proceedings of 16th Annual International Conference on Automated Software Engineering,
2001 (ASE 2001), pp. 273–280. IEEE (2001)
20. Kozlenkov, A., Paschke, A.: Prova rule language version 3.0 user’s guide. http://prova.ws/
index.html (2010)
21. Jeff, A., Stephen, G.: The minion search engine: indexing, search, text similarity and tag
gardening. Technical report, Sun Microsystems, New York (2008)
336 S. Ramakrishna et al.

22. Robertson, S., Zaragoza, H.: The Probabilistic Relevance Framework: BM25 and Beyond.
Now Publishers Inc., Breda (2009)
23. Porter, M.F.: Snowball: a language for stemming algorithms (2001)
24. Newman, D., Asuncion, A., Smyth, P., Welling, M.: Distributed algorithms for topic models.
J. Mach. Learn. Res. 10, 1801–1828 (2009)
25. Ramakrishna, S.: First approaches on knowledge representation of elementary (patent)
pragmatics. In: Joint Proceedings of the 7th International Rule Challenge, the Special Track
on Human Language Technology and the 3rd RuleML Doctoral Consortium (2013)
26. Rissland, E.L., Ashley, K.D., Branting, L.K.: Case-based reasoning and law. Knowl. Eng.
Rev. 20(03), 293–298 (2005)
27. Ramakrishna, S., Paschke, A.: Bridging the gap between legal practitioners and knowledge
engineers using semi-formal KR. In: The 8th International Workshop on Value Modeling
and Business Ontology, VMBO, Berlin (2014)
28. Ramakrishna, S., Paschke, A.: Semi-automated vocabulary building for structured legal
english. In: Bikakis, A., Fodor, P., Roman, D. (eds.) RuleML 2014. LNCS, vol. 8620,
pp. 201–215. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-09870-8_15
29. Boley, H., Paschke, A., Shafiq, O.: RuleML 1.0: the overarching specification of web rules.
In: Dean, M., Hall, J., Rotolo, A., Tabet, S. (eds.) RuleML 2010. LNCS, vol. 6403, pp. 162–
178. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-16289-3_15
30. Paschke, A.: Reaction RuleML 1.0 for rules, events and actions in semantic complex event
processing. In: Bikakis, A., Fodor, P., Roman, D. (eds.) RuleML 2014. LNCS, vol. 8620,
pp. 1–21. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-09870-8_1
31. Ramakrishna, S., Paschke, A.: A process for knowledge transformation and knowledge
representation of patent law. In: Bikakis, A., Fodor, P., Roman, D. (eds.) RuleML 2014.
LNCS, vol. 8620, pp. 311–328. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-
09870-8_23
32. Paschke, A., Ramakrishna, S.: Legal RuleML Tutorial Use Case - LegalRuleML for Legal
Reasoning in Patent Law (2013)
33. Ramakrishna, S., Gorski, Ł., Paschke, A.: A dialogue between a lawyer and computer
scientist: the evaluation of knowledge transformation from legal text to computer-readable
format. Appl. Artif. Intell. 30(3), 216–232 (2016)
34. Bernstam, E.V., Herskovic, J.R., Aphinyanaphongs, Y., Aliferis, C.F., Sriram, M.G., Hersh,
W.R.: Using citation data to improve retrieval from MEDLINE. J. Am. Med. Inform. Assoc.
13(1), 96–105 (2006)
 Towards Annotation of Legal Documents
with Ontology Concepts

Kolawole John Adebayo(B) , Luigi Di Caro, and Guido Boella

Dipartimento di Informatica, Universita Di Torino,

Corso Svizzera 185, 10149 Turin, Italy
kolawolejohn.adebayo@unibo.it, {dicaro,guido.boella}@di.unito.it

Abstract. This paper describes a task of semantic labeling of docu-

ment segments. The idea exploits ontology in providing a fine-grained
conceptual document annotation. We describe a way of dividing a doc-
ument into its constituent semantically-coherent blocks. These blocks
are then used to perform conceptual tagging for efficient passage infor-
mation retrieval. The proposed task interfaces other application areas
such as intra-mapping of ontologies, text summarization and informa-
tion extraction. The system has been evaluated on a task of conceptual
tagging of documents and achieved a promising result.

Keywords: Conceptual tagging · Semantic annotation

Information retrieval · Passage retrieval · Text segmentation

1 Introduction

The deluge of electronically stored information (ESI), coupled with the data
explosion on the Internet has practically necessitated developing frameworks for
intelligent document processing. This is important since ESI is mostly unstruc-
tured. The Legal domain has also witnessed an unprecedented growth in the
amount of ESI produced e.g., in the law courts, government assemblies etc.,
necessitating for repositories like Eurlex1 etc. This comes with the responsibility
of developing scalable retrieval techniques that provide legal practitioner an easy
interface for retrieving information in a timely and efficient manner.
In this paper, we propose a passage retrieval system which works by seg-
menting a document into different semantically coherent parts based on the
meaning of its content. In addition, the proposed system introduces a topic-like
structure to an unstructured text since each of the segments relates to different
concept/topic. The highlight of the proposed approach is how it automatically
associates each segment of an input document to its respective concept from
an ontology. We formalize this process as a Semantic Annotation task [5]. The
rationale behind the proposed system is to provide practitioners and other end

1
https://eur-lex.europa.eu/homepage.html.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 337–349, 2018.
https://doi.org/10.1007/978-3-030-00178-0_23
338 K. J. Adebayo et al.

users with a fine-grained passage retrieval with a simple Natural Language Pro-
cessing (NLP) tool that allows users to specify query using a controlled list of
concept and the system retrieves not just the document related to the concept
but specific part(s) of the document that is semantically related to the concept.
There are motivations for the proposed system. First, users are freed from
the rigours associated with query formulation. This is important because many
people understand their information need but the problem is how to effectively
represent and present such information need to a retrieval algorithm. By provid-
ing a controlled list of descriptors, such problems are adequately taken care of.
Secondly, by providing a fine-grained retrieval, the problem of information over-
load is solved. Information overload occurs when a system retrieves more infor-
mation than its actually needed by a user, often requiring the user to manually
inspect the retrieved item in order to locate the specific information of inter-
est. This improves not just the precision but also the recall which is especially
important in the legal domain. Moreover, concept mapping can support semantic
query processing across disparate sources by expanding or rewriting the query
using the corresponding information in multiple ontologies. More importantly,
an Ontology can capture a domain knowledge in a machine understandable way,
thereby providing a solution for solving semantic heterogeneity problem2 .
The method that we present in this paper is different from existing work in
the way that we represent the meaning of a concept in an unsupervised way.
Moreover, it does not directly rely on any Machine Learning (ML) methods.
Instead, it is based on the expansion and enrichment of linguistic terms using
some natural language processing techniques. The remaining parts of the paper
are structured as follows. In Sect. 2, we review some related work. Section 3
contains a description of our algorithm and methods. Finally, we describe our
experiment, the evaluation, and the result obtained in Sect. 4.

2 Related Work

Semantic Annotation (SA) formalizes and structure a document with well-

defined semantics specifically linked to a defined ontology [23]. SA can be for-
malized as a 4-tuple (a, b, c, d), where a is the subject of the annotation,
b is the object of the annotation, c is the predicate which defines the type of
relationship between a and b, while d signifies the context in which the annota-
tion is made. An ontology is a formal conceptualization of the world, capturing
consensual knowledge [15]. It contains the concepts along with their properties
and the relationship that exists between them. This study uses the Eurovoc3
thesaurus as the ontology.
Generally, annotation can aid structured organization of documents for opti-
mized search. For instance, users may search information by well-defined general
concepts that describe the domain of information need rather than use keywords.

2
E.g homonyms and synonyms.
3
http://eurovoc.europa.eu/.
 Towards Annotation of Legal Documents with Ontology Concepts 339

There are existing works which share similarity to the method proposed here.
GATE [9] is a semi-automatic annotation system based on NLP. GoNTogle [5]
uses weighted k -Nearest Neighbor (kNN) classifier for document annotation and
retrieval. The authors in [24] developed a tool for ontology learning and popu-
lation in the Semantic Web. Their approach utilizes Discourse Representation
Theory and frame semantics for performing knowledge extraction. KIM [23]
assigns semantic descriptions to Named Entities (NEs) in a text. The system
is able to create hyperlinks to NEs in a text such that indexing and document
retrieval is performed with the NEs. Regular Expressions (RE) have also been
used to identify semantic elements in a text [16,17]. It works by mapping part
of a text related to semantic context and matching the subsequent sequence of
characters to create an instance of the concept. Another NE-based annotation
tool is GERBIL [27] which provides rapid but extensive evaluation scheme for
NE recognition tools for the semantic web. Application of these systems includes
document retrieval especially in the semantic web domain [10,12].
The authors in [21] performed semantic annotation on legal documents for
document categorization. Using Eurovoc concept descriptors on EurLex 4 , a ML
classifier was trained for multi-label classification. The authors’ work is a super-
vised concept-to-document mapping, i.e., a document categorization where the
concept(s) for a document is predicted by the ML classifier. In our work, we
learn the concept representation in a completely unsupervised way, furthermore,
instead of associating the concept to a whole document, our algorithm asso-
ciate the specific segment(s) of document(s) to a concept once the algorithm
determines that they have the same semantics.
Associating a concept to a document segment is only possible after finding
the points in a text where there is a semantic drift. In our work, we identify
these points by observing the distribution of topics in a text and how the top-
ics change from one part of the document to another. Researchers working on
Discourse structure have shown that a document is usually a mixture of topics
and sub-topics [11]. Each topic discusses a theme of the document. Furthermore,
topics sharing the same theme usually cohere, thus forming a form of segment
or semantically align units. We incorporate this idea of topical segmentation of
document such that concept(s) can be easily linked to any topical segment.
This part of our work follows the TextTiling algorithm [14] and a topic-
modeling based improvement [25] which uses the Latent Dirichlet Allocation
(LDA) [6]. These systems divide text into a contiguous, non-overlapping dis-
course units that correspond to the pattern of subtopics in a text. Some Semantic
Textual similarity [19] ideas have also been incorporated for improved accuracy.
The intuition here is that sentences that falls belonging to the same segment
must be semantically similar.
We take a segment as an information unit in a document, defined by dif-
ferent levels of granularity, i.e., sentence, paragraph or section holding many
paragraphs. The goal is to find a semantic correspondence between a concept

4
An online database of EU government documents.
340 K. J. Adebayo et al.

Fig. 1. The system architecture Fig. 2. A schematic representation of

the concept-document mapping

descriptor and each topical segment. A high level structure of the proposed task
is shown in Fig. 1.
The proposed system has diverse applications. As an example, legal doc-
uments usually contain many pages of information. A practitioner looking for
specific information of interest from such documents would have to read through
the pages before identifying the needed part even though the information that
is being searched is contained within just a few paragraphs. With the proposed
system, a user only has to specify a concept which represents the information
need, and the system identifies and retrieves the particular segment(s) which is
semantically related to the concept.

3 System Description

We use the Eurovoc thesaurus as the ontology and a collection from the EurLex
website as documents to be processed. EurLex documents are already labeled
with corresponding Eurovoc concept descriptors.
Consider a text in a collection of documents, where each document is asso-
ciated to a set of concepts taken from a thesaurus σ (i.e., a multi-labeled text
collection). In our work, instead of assuming a Universe of Concept (UoC) from
the ontology, we use sets of concept already classified with that EurLex docu-
ment.
We divide the task into 3 steps, i.e., document analysis, concept analysis and
concept-segment mapping. The document analysis part performs topical docu-
ment segmentation needed for the concept-segment mapping. The concept anal-
ysis step creates a profile for each concept, each profile is a unique symbolic sig-
nature capturing possible semantics of a concept. Concept profile can be viewed
as a form of virtual document containing all possible descriptive information
for that concept. The concept-Segment mapping part compare each profile and
topical segment for similarity. Based on some heuristics, the system assigns the
profile to the respective related segments. Figure 2 shows a pictorial representa-
tion of the details.
We now proceed to fully describe each step.
 Towards Annotation of Legal Documents with Ontology Concepts 341

3.1 Document Representation

Given a text ti and its sequence of sentences Seqi = <s1 , s2 , ..., sk >. The docu-
ment analysis part aggregates all the sentences or paragraphs of the text that
semantically align into a group. This semantically aligned group is called a seg-
ment, each of which can be directly mapped to the concept profile from an ontol-
ogy. Basically, a topic segmentation algorithm tries to discover topic boundaries
which symbolizes a semantic drift in the flow of topics within a document. These
boundaries are then used to separate the sentences that is about individual top-
ics, the group of aligned sentences must be coherent, hence a separate segment
or topical unit.
In our work, topical segmentation is done based on an improved topic-
modeling based Text Segmentation [25]. Our text segmentation implementation
not only rely on topics but further incorporates a Semantic Nets5 based similar-
ity in order to improve segmentation accuracy. The approach follows the general
template of the TextTiling [7,13] system. Generally, TextTiling algorithm uses
the amount of word overlaps among contiguous sentences. It relies on the obser-
vations that the structure of a text is a function of its constituents terms. Using
basic cues, it groups coherent sentences based on word repetition and term co-
occurrences within a small axis and the extent of word similarity. A text is broken
into pseudo-blocs, with each bloc comprising of k (3–5) adjacent sentences. A
measure of relatedness is then obtained with cosine similarity of their vectors.
n
wt,b1 wt,b2
cos(b1 , b2 ) =  t=1 n (1)
n
w 2
t=1 t,b1 w 2
t=1 t,b2

Where b1 and b2 are the two sentences compared. Highly similar blocs tend
to form peaks and in contrast dissimilar blocs forms valleys, hence a boundary.
Unlike TextTiling which uses only word repetition as evidence of coherence,
we also compute semantic similarity using a WordNet-based Sentence similarity
algorithm, this typically overcomes the problem of ambiguity in word usage 6 .
Following the works of [25], we compute similarity between the LDA [6] generated
topic vectors for each sentence in a candidate segment, this approach results in
an improved performance. The reader is referred to [2] for detailed description
of the text segmentation approach and algorithm.

3.2 Concept Representation

Once the topical segments of a document are identified, the next step is to
retrieve the concepts that a document is manually labelled with on EurLex and
then create a profile for each concept. A concept profile is essentially a signature
which by design incorporates all the descriptive information of and/or about the

5
We used WordNet concept distance to compute semantic similarity between adjacent
sentences of a candidate segment according to a chosen window.
6
E.g., synonymy and polysemy.
342 K. J. Adebayo et al.

concept. Concept descriptors could range from unigram, bigrams to n-grams.

Similarly to query expansion, if a concept descriptor has more than one word,
we check if it is included in the WordNet as a word, e.g., the word Jet-lag has
two joint terms (jet and lag). However, it is an entry in the WordNet as a single
term. On the other hand, the word public health is not included in WordNet as
a single term. In the second case, e.g., public health, we break the n-gram term
into its constituent words, e.g., public and health. We then enrich the constituent
term(s) in a process called lexical expansion.
The goal of lexical expansion is to retrieve semantically similar words such
as synonyms from WordNet etc. We also obtain a top-k 7 related terms from a
word embedding model obtained when the Word2vec [18] algorithm was trained
on the Wikipedia dump8 . The Word2Vec9 is an algorithm capable of inducing
semantic relatedness and similarity of words from a large corpus based on their
Distributional features [26]. Given a word, it learns to predict the word that
appears next to the given word, using past contextual information about that
word obtained from the corpus.
The Gensim10 implementation of the Word2vec algorithm uses Skip-Gram
with Hierarchical Softmax which is widely preferred over the Continous-Bag-of-
Word (CBOW) in terms of performance. The model obtained when the algo-
rithm is trained over a corpus can be used to obtain semantically related words
to a term. As an example, the model is able to understand that Paris is to France
as London is to England thus capturing country-capital relationships based on
knowledge extracted from the corpus when employed in a word analogy task.
The use of the word embedding model is important since some knowledge
cannot not be derived from the WordNet. Recall that a concept does not have
to appear in a text explicitly for humans to understand what the text is talking
about, e.g., say a concept soccer, a document that contains terms like player,
striker, goal, or even names like Manchester United, Ronaldo, Messi etc., is
an indicator that the document is about soccer even without the word ’soccer’
appearing in the document. Entries like Manchester United or Ronaldo which
cannot be found in WordNet are easily captured by a word embedding model
assuming that the related documents are included in the training set. Owing
to its characteristics, Researchers have found that it is particularly useful for
Information retrieval task [3,4,8,22].
To further enrich the profile, the ontology is traversed to obtain the neighours,
e.g., ancestors, siblings and the children of a concept. The idea is that siblings of
a concept are also semantically similar and the children could be more specific
and less general than the parent node. Most importantly, using this informa-
tion automatically disambiguates a concept. The profile of the neighours are
also extracted and included in the profile of the concept. We call this ontology

7
Where k is a chosen number and set as 3.
8
English wikipedia dump was downloaded on July 30, 2015.
9
https://code.google.com/p/word2vec/.
10
https://github.com/RaRe-Technologies/gensim. Training Parameters: Context
Window: 5, Neural Network layer size: 200, Minimum word count: 5.
 Towards Annotation of Legal Documents with Ontology Concepts 343

enrichment. We say that the combination of the synonym sets from the Word-
Net, the terms obtained from the ontology enrichment, and the related words
obtained with the word embedding model constitute the profile for a concept.
The profile prof ilecon , of a concept is defined as:
−−−−−−−−−→
prof ilecon (a) = {Synwn (a)} + {Rtermwe (a)} + {Ontosurround (a)} (2)

Where Synwn is the set of synsets obtained from WordNet, Rtermwe is the
set of semantically related terms obtained from the word embedding model and
Ontosurround are the surrounding or neighours of a concept in the ontology.
The concept profile can be unnecessarily large due to duplicity of terms
e.g. a synset obtained from WordNet may be returned by the word embedding
model as well. It is therefore necessary to remove this redundancy. A concept
profile is trimmed by removing any repeating term and those terms that are
less semantically similar to the other terms in the profile. First, some randomly
selected baseline terms whose similarity exceeds a threshold are selected. The
threshold is heuristically determined and it ensures that those terms are highly
semantically similar. The baseline terms are then used as a reference in measuring
the similarity of every other terms in the profile. Similarity of other terms in the
profile is calculated and a weight is assigned to each term based on the similarity
score [1]. The weights of all the terms are then ranked and the top n ranked
terms are selected while others are discarded. The result is that the concept
profile contains a dense but highly semantic signature of the original concept.
This reduces over-fitting of the profile with less useful terms. The trimmed profile
prof iletrimmed is defined by:
−−−−−−−−−−−−−→
prof iletrimmed (a) = {t | SIM (baseline, t) > k} (3)

Where baseline is the set of chosen terms used to compare other terms, the
SIM function returns a weight after computing similarity between a term in the
profile and each of the baseline terms, k is a threshold for ranking, terms with
weight less than k are discarded. For example, say a profile contain terms like
{airport, transport, flight, shuttle, metro, wind}, we may safely assume that the
word wind constitute mere noise to the profile as it is the least related to the
others. Next, the terms in the trimmed profile are merged; in order to generate a
sentential representation for the concept which is used for the concept-document
mapping process.

3.3 Concept-Document Mapping

Concept-Document mapping can be viewed as finding a semantic correspon-
dence between the semantic representation of a concept and the topical segment
of a document. Recall that the concept profile for a particular document is built
from the Eurovoc descriptors that the Eurlex document is already classified with.
We can formalize the maping task as 4-tuple (Senf orm , Topseg , Rel, Sim) where
Senf orm is the sentential representation of a concept, T opseg is the topical seg-
ment of a document, Rel is the semantic relationship between Senf orm , and
344 K. J. Adebayo et al.

T opseg which is represented by similarity metric Sim. Sim can be any vector
distance metric like the Euclidean or Cosine similarity.
Matching a given concept to a text segment in a text is reduced to a simple
semantic relatedness task between the sentential representation of each concept
and the sentences in each segment. Thus, we can view it as a form of semantic
similarity task. In order to compute similarity, we use two approaches. The first
uses a cosine similarity metric. The issue with this approach is that Senf orm
may not have many words in common with T opseg , such that even if the two
share words that are similar in meaning but with different lexicographic form,
the cosine similarity will be low. The similarity obtained here is called Sim1.
Secondly, we compare each word in Senf orm with the words in T opseg . How-
ever, some words in T opseg may not be useful, e.g., stop words. We perform
part-of-speech tagging using the Stanford POS tagger [20] in order to retain
only the verbs, adjectives and nouns. We call the resulting set of words the
T opseg profile. Next, word-word similarity is done between the terms in T opseg
and those in Senf orm .
The word-word similarity computation is accomplished using a WordNet
based similarity implementation. To derive similarity from WordNet, we used
both the path length between each word as well as the depth function. Usually,
longer path length between two concepts signifies lower similarity. The author in
[19] introduced the depth function with the intuition that the words at the upper
layer of a Semantic Net contains general semantics and less similarity while those
at the lower layers are more similar. Therefore, a similarity assessment should
be a function of both the depth as well as the path length distances between
concepts. If f1(h) is a function of the depth and f2(l) is a function of the length,
then the similarity between two word is given by:

S(w1 , w2 ) = f 1(h).f 2(l) (4)

The length function is a monotonically decreasing function with respect to the

path length l between two concepts. This is captured by introducing a constant
alpha.
f2(l) = e−∝l (5)
Likewise, the depth function is monotonically increasing with respect to the
depth h of concept in the hierarchy.

eβh − e−βh
f1(h) = (6)
eβh + e−βh
The similarity between two WordNet concepts is then calculated by:

eβh − e−βh
S(w1 , w2 ) = e−∝l . (7)
eβh + e−βh
The author [19] empirically discovered that for optimal performance in Word-
Net, α should be set to 0.2 and β set to 0.45. Since this approach considers only
the similarity between two words, it is important to find a way to combine all
 Towards Annotation of Legal Documents with Ontology Concepts 345

the similarity scores that exceed a particular threshold. The aggregation function
below is used: m,n
i,j |S(wi , wj > x)|
Sim = (8)
tCount
Where S(wi , wj ) is the similarity score for two words, tCount is the total number
of the set of similarity scores that exceeds the threshold and Sim is the aggre-
gating function combining all pairwise similarities. Past studies have shown that
by encoding query and document terms with vectors from a word embedding
model, combining the terms using for example a simple vector addition operator
and then performing vector averaging, a retrieval system that is robust to lan-
guage variability issues like synonymy and polysemy can be obtained [3,8]. We
encode the terms in T opseg and those in Senf orm with vectors from our trained
word embedding model. Next, we obtain the representation for the concept and
segment by performing vector addition and averaging of the terms in order to
yield a single vector. The similarity between the concept vector and a segment
vector is then computed with the Cosine similarity formula.
We call the WordNet based similarity score Sim2 and the similarity score
from the word embedding is Sim3. The final similarity is obtained by the mean
of the 3 similarities earlier calculated. This is given by the formula:
Sim1 + Sim2 + Sim3
cSim = (9)
n
Where n is the number of values combined, e.g., n = 3. Comparing each Senf orm
to each T opseg profile yields a vector of similarity scores for each Senf orm , e.g.
assuming that a document has 3 topical segments, then a profile Senf orm (con)
of a concept con having vector [0.8, 0.4, 0.2] means that its similarity score with
the first topical segment is 0.8, second is 0.4 and third is 0.2 etc. A concept
profile is said to be similar to the topical segment it has the highest similarity
score with if it is Z 11 times higher than all other values else, the highest and the
next highest sentences are both taken to be semantically close to the concept
profile. The Z parameter is computed with the formula:
a−b
Z= × 100 (10)
b
Where a and b are the highest and next highest values in the similarity scores
vector of concept profile. In the example scores vector above, a = 0.8 and b =
0.40. Calculating z gives 1.00 which is greater than the threshold value for z.
Thus the segment with score 0.8 is the only one tagged with that concept.

4 Evaluation
We selected 100 documents from EurLex website, 25 documents each from four
different categories. EurLex is an open and regularly updated online database
11
Ensures the highest similarity is at least z percent higher than the next highest value.
By default, z = 0.3.
346 K. J. Adebayo et al.

of over 3 million European Union documents covering EU treaties, regulations,

legislative proposals, case-law, international agreements, EFTA documents etc.
Documents are already classified using Eurovoc descriptors. We used the Eurovoc
thesaurus as the ontology. EurLex and Eurovoc are both multilingual. Cur-
rently, it is available in 26 European languages. The documents downloaded are
English versions from Consolidated Acts section of the website. Specifically, we
selected documents under Transport Policy category. The sub-categories include
{Transport Infrastructure, Inland Transport, Shipping, Air Transport}. We eval-
uated the system on a task of conceptual tagging.
Conceptual Tagging measures the performance of the system in correctly
labeling a text segment with a concept. We measure the performance of the
system against annotations from human judgment. To achieve this, all the doc-
uments were first automatically segmented into topical sections with our text
segmentation algorithm. Two volunteers were asked to read each document and
assign Eurovoc descriptors to the segments in the document as appropriate. The
descriptors chosen for each document are those the document were classified
with on EurLex website.
A segment may not be labeled with a concept descriptor if the volunteer
annotators believed that there is no semantic relationship between it and any
of the concept. Also, a segment can have more than one concept related to it.
A third volunteer compares annotations from the first two volunteers, and where
annotations do not correlate, decides the final annotation per document. The
agreements were rated based on individuals judgment in labeling a text segment
with a concept.
The same topical segments for each document were fed into the developed
system. The goal of the system is to quantify the meaning of these segments and
for each, select the concept that is most semantically related. Using the manual
annotation as Gold Standard, we compared the performance of the system with
that of manual annotation using the popular information retrieval metrics: Pre-
cision and Recall. The precision is the number of accurate tagging by the system
in comparison to that of human annotators. The recall, on the other hand is
the number of accurate tagging made by the system. Table 1 shows the results
obtained under different categories of documents. We can see that we obtained
the best precision and recall scores from ’Transport Infrastructure’ documents
while the worst results comes from ’Air Transport’ category. Manual exploration

Table 1. Precision and recall across document genre

Domain of documents No of documents Precision (%) Recall (%)

Transport infrastructure 25 0.74 0.77
Inland transport 25 0.71 0.73
Shipping 55 0.72 0.74
Air transport 25 0.68 0.71
Average score 100 0.71 0.73
 Towards Annotation of Legal Documents with Ontology Concepts 347

of the documents reveal that the documents under this category with the best
result are quite short (average of 3 pages) compared to those under Air Trans-
port where the average number of pages was double that of the former. Overall,
we obtained an average precision score of 71%, recall value of 73%, and an overall
F1 score of 72%.

5 Conclusion

The paper describes an approach to semantically link concepts to the parts of

the document that the concept is semantically related to. The proposed concept-
document mapping can aid conceptual information retrieval task such that users
have a more fine-grained results, e.g., specific document part that satisfies the
user query instead of the whole document. The Eurovoc thesaurus was used as
the ontology of choice since EurLex documents are pre-labeled with descriptors.
The task has useful applications in Information retrieval and extraction as well
as text segmentation. For instance, a user who is more interested in a particular
information in a document can easily specify such information need through a
concept that describes such need. The system is then able to extract the specific
portion containing the requested information need. Experimental result shows
that the idea is feasible with a promising result. Future work will include an
extended set of experiments by getting more volunteers to help with manual
annotations require for thorough evaluation.

Acknowledgments. Kolawole J. Adebayo has received funding from the Erasmus

Mundus Joint International Doctoral (Ph.D.) programme in Law, Science and Technol-
ogy. Luigi Di Caro and Guido Boella have received funding from the European Union’s
H2020 research and innovation programme under the grant agreement No 690974 for
the project “MIREL: MIning and REasoning with Legal texts”.

References
1. Adebayo, K., Di Caro, L., Boella, G.: NORMAS at SemEval-2016 task 1: SEMSIM:
a multi-feature approach to semantic text similarity. In: Proceedings of the 10th
International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2016,
San Diego, CA, USA, 16–17 June 2016, pp. 718–725 (2016)
2. Adebayo, K.J., Di Caro, L., Boella, G.: Text segmentation with topic modeling and
entity coherence. In: Abraham, A., Haqiq, A., Alimi, A.M., Mezzour, G., Rokbani,
N., Muda, A.K. (eds.) HIS 2016. AISC, vol. 552, pp. 175–185. Springer, Cham
(2017). https://doi.org/10.1007/978-3-319-52941-7 18
3. Adebayo, K.J., Di Caro, L., Boella, G., Bartolini, C.: An approach to information
retrieval and question answering in the legal domain, pp. 15–25 (2016)
4. Ai, Q., Yang, L., Guo, J., Croft, W.B.: Improving language estimation with the
paragraph vector model for ad-hoc retrieval. In: Proceedings of the 39th Inter-
national ACM SIGIR Conference on Research and Development in Information
Retrieval, pp. 869–872. ACM (2016)
348 K. J. Adebayo et al.

5. Bikakis, N., Giannopoulos, G., Dalamagas, T., Sellis, T.: Integrating keywords
and semantics on document annotation and search. In: Meersman, R., Dillon, T.,
Herrero, P. (eds.) OTM 2010. LNCS, vol. 6427, pp. 921–938. Springer, Heidelberg
(2010). https://doi.org/10.1007/978-3-642-16949-6 19
6. Blei, D.M., Lafferty, J.D.: Dynamic topic models. In: Proceedings of the 23rd Inter-
national Conference on Machine Learning, pp. 113–120. ACM (2006)
7. Choi, F.Y.Y.: Advances in domain independent linear text segmentation. In: Pro-
ceedings of the 1st North American Chapter of the Association for Computa-
tional Linguistics Conference, pp. 26–33. Association for Computational Linguis-
tics (2000)
8. Clinchant, S., Perronnin, F.: Aggregating continuous word embeddings for infor-
mation retrieval. In: Proceedings of the Workshop on Continuous Vector Space
Models and their Compositionality, pp. 100–109 (2013)
9. Cunningham, H., Maynard, D., Bontcheva, K., Tablan, V.: A framework and graph-
ical development environment for robust NLP tools and applications. In: ACL, pp.
168–175 (2002)
10. Dill, S., et al.: A case for automated large-scale semantic annotation. Web Semant.:
Sci. Serv. Agents World Wide Web 1(1), 115–132 (2003)
11. Halliday, M.A.K., Hasan, R.: Cohesion in English (1976)
12. Handschuh, S., Staab, S.: Authoring and annotation of web pages in cream. In:
Proceedings of the 11th International Conference on World Wide Web, pp. 462–
473. ACM (2002)
13. Hearst, M.A.: TextTiling: a quantitative approach to discourse segmentation. Tech-
nical report, Citeseer (1993)
14. Hearst, M.A.: TextTiling: segmenting text into multi-paragraph subtopic passages.
Computational linguistics 23(1), 33–64 (1997)
15. Kiyavitskaya, N., Zeni, N., Mich, L., Cordy, J.R., Mylopoulos, J.: Text mining
through semi automatic semantic annotation. In: Reimer, U., Karagiannis, D.
(eds.) PAKM 2006. LNCS (LNAI), vol. 4333, pp. 143–154. Springer, Heidelberg
(2006). https://doi.org/10.1007/11944935 13
16. Laclavı́k, M., Ciglan, M., Seleng, M., Krajei, S.: Ontea: semi-automatic pattern
based text annotation empowered with information retrieval methods. In: Tools
for Acquisition, Organisation and Presenting of Information and Knowledge: Pro-
ceedings in Informatics and Information Technologies, Kosice, Vydavatelstvo STU,
Bratislava, part, vol. 2, pp. 119–129 (2007)
17. Laclavik, M., Seleng, M., Gatial, E., Balogh, Z., Hluchy, L.: Ontology based text
annotation-OnTeA, pp. 280–284 (2006)
18. Le, Q.V., Mikolov, T.: Distributed representations of sentences and documents.
arXiv preprint arXiv:1405.4053 (2014)
19. Li, Y., McLean, D., Bandar, Z.A., O’shea, J.D., Crockett, K.: Sentence similarity
based on semantic nets and corpus statistics. IEEE Trans. Knowl. Data Eng. 18(8),
1138–1150 (2006)
20. Manning, C.D., Surdeanu, M., Bauer, J., Finkel, J.R., Bethard, S., McClosky,
D.: The Stanford CoreNLP natural language processing toolkit. In: ACL (System
Demonstrations), pp. 55–60 (2014)
21. Loza Mencı́a, E., Fürnkranz, J.: Efficient multilabel classification algorithms for
large-scale problems in the legal domain. In: Francesconi, E., Montemagni, S.,
Peters, W., Tiscornia, D. (eds.) Semantic Processing of Legal Texts. LNCS (LNAI),
vol. 6036, pp. 192–215. Springer, Heidelberg (2010). https://doi.org/10.1007/978-
3-642-12837-0 11
 Towards Annotation of Legal Documents with Ontology Concepts 349

22. Mitra, B., Diaz, F., Craswell, N.: Learning to match using local and distributed
representations of text for web search. In: Proceedings of the 26th International
Conference on World Wide Web, pp. 1291–1299. International World Wide Web
Conferences Steering Committee (2017)
23. Popov, B., Kiryakov, A., Kirilov, A., Manov, D., Ognyanoff, D., Goranov, M.: KIM
– semantic annotation platform. In: Fensel, D., Sycara, K., Mylopoulos, J. (eds.)
ISWC 2003. LNCS, vol. 2870, pp. 834–849. Springer, Heidelberg (2003). https://
doi.org/10.1007/978-3-540-39718-2 53
24. Presutti, V., Draicchio, F., Gangemi, A.: Knowledge extraction based on discourse
representation theory and linguistic frames. In: ten Teije, A., et al. (eds.) EKAW
2012. LNCS (LNAI), vol. 7603, pp. 114–129. Springer, Heidelberg (2012). https://
doi.org/10.1007/978-3-642-33876-2 12
25. Riedl, M., Biemann, C.: Text segmentation with topic models. J. Lang. Technol.
Comput. Linguist. 27(1), 47–69 (2012)
26. Turney, P.D., Pantel, P., et al.: From frequency to meaning: vector space models
of semantics. J. Artif. Intell. Res. 37(1), 141–188 (2010)
27. Usbeck, R., et al.: GERBIL: general entity annotator benchmarking framework.
In: Proceedings of the 24th International Conference on World Wide Web, pp.
1133–1143. ACM (2015)
 Reuse and Reengineering of Non-ontological
Resources in the Legal Domain

Cristiana Santos1(&), Pompeu Casanovas1,4,

Víctor Rodríguez-Doncel2, and Leendert van der Torre3
1
Institute of Law and Technology (IDT-UAB), Barcelona, Spain
cristiana.teixeirasantos@gmail.com,
pompeu.casanovas@uab.cat
2
Ontology Engineering Group,
Universidad Politécnica de Madrid, Madrid, Spain
vrodriguez@fi.upm.es
3
University of Luxembourg, Luxembourg City, Luxembourg
leon.vandertorre@uni.lu
4
Data to Decisions Cooperative Research Centre,
La Trobe Law School, Melbourne, Australia
P.CasanovasRomeu@latrobe.edu.au

Abstract. Instead of custom-building a new ontology from scratch, knowledge

resources can be elicited, reused and engineered to develop legal ontologies with
the goal of promoting the application of good practices and speeding up the
ontology development process. This paper focuses on the specificities of non-
ontological resources in the legal domain, and provides some guidelines of how
these can be reused and engineered to enable heterogeneous resources inte-
gration within a legal ontology. The paper presents some examples of these
processes using a case-study in the consumer law domain.

Keywords: Ontology engineering
Non-ontological resources

Legal ontology
MeLOn methodology
NeOn methodology

1 Introduction

Instead of custom-building a new legal ontology from scratch, knowledge resources are
elicited from the legal domain, reused and engineered to develop legal ontologies1,
promoting the application of good practices.
Knowledge resources have been classified as ontological resources (ORs) or non-
ontological resources (henceforth named NORs) [1]. This division regards the level of
formalization. We will focus on the latter type. There is much literature for reusing and
reengineering ORs [2, 3] and also ontology design patterns, but little about extracting
knowledge from NORs in the legal domain, probably due to its specificities, delved in

1
Ontologies are the chosen artifact to support the integration of data from multiple, heterogeneous
legal sources, making information explicit and enabling the sharing of a common understanding of a
domain.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 350–364, 2018.
https://doi.org/10.1007/978-3-030-00178-0_24
 Reuse and Reengineering of Non-ontological Resources 351

this paper. This subject is relevant as it has consequences at different levels, from
knowledge acquisition, to ontology engineering.
There is a large amount of NORs that embody knowledge in the legal domain, that
represent some degree of consensus for the legal community and possess related
semantics that allows interpreting the knowledge contained therein. In fact, within this
domain, NORs may correspond to some legal sources which consist on legislation, but
also other relevant sources of e.g., case law, doctrinal interpretations, social rules; it is
essential to connect this existing legal material to the ontology, even if its majority is
not formalised, and hence not necessarily interoperable. NORs from this realm can be
embedded in different and scattered sources of hard and soft law, such as classification
schemes, thesauri, lexicons2, textual corpora, among others, in a “patchwork” of “lego”
pieces. The heterogeneity of the legal sources is observed at multiples levels: structural,
semantic, and syntactic. To integrate information from multiple and heterogeneous
knowledge sources, it is important to cope with the problem of legal knowledge rep-
resentation, that consists in the balance between consensus and authoritativeness3 [4]
or, from the socio-legal perspective, dialogue and bindingness [5].
On the one hand, domain legal experts lack competencies in data modeling, and
they often adopt technical tools (e.g., Protégé) without the necessary awareness of the
technical consequences [6]. On the other hand, ontology developers, besides the data
modeling perspective, should consider likewise compliance with the specificities of the
juristic nature of legal NORs and of expert knowledge. A balanced combination would
yield reliable actionable knowledge in a real world context, for a thorough under-
standing of the considered legal field is necessary to bring out explicit conceptual-
izations, to shape the design of the ontology and its population. Legal information
specificities [7] and ontology interplay, in both its theoretical and engineering
dimensions, are intrinsically connected. The interaction between legal concepts that
affect the utilization of information is significative [6]. Hence, an interdisciplinary
approach is essential towards representing machine-readable concepts and relationships
from NORs in the legal domain, through the due processes of reusing and reengi-
neering thereof. Hence, the research question of this paper is how to reuse non-
ontological resources in the legal domain and their reengineering into ontologies. For
such purpose we follow two complementary methodological approaches: (i) “Building
Ontology Networks by Reusing and Reengineering Non Ontological Resources”,
Scenario 2 from NeOn4 methodology framework (henceforward called NeOn) that
explains how to build ontologies by reusing and reengineering non-ontological
resources [1, 8–10]; and (ii) the Methodology for building Legal Ontology (henceforth
called MeLOn) [6], developed by Monica Palmirani.

2
A lexicon is the vocabulary of an individual person, an occupational group or a professional field,
Glossary of Terms for the Standardization of Geographical Names, United Nations Group of Experts
on Geographic Names, United Nations, New York, 2002.
3
Regarding authoritativeness and bindingness, knowledge representation in the legal domain entails
some peculiar features, because it is supposed that authority is somewhat embedded into the text.
4
http://www.neon-project.org.
352 C. Santos et al.

The observations held in this paper are built upon the construction of two legal
ontologies named Relevant legal information for consumer disputes (RIC) and RIC-
ATPI, referring to the relevant information in the domain of air transport passenger
incidents [11].
The remainder of the paper is structured as follows. Section 1 describes the
specificities of NORs in the legal domain. Section 2 refers to the main methodologies
to build ontologies with NORS; Sect. 3 explains the NORs reuse and reengineering
processes. Section 4 concludes the paper emphasizing the challenges and lessons
learned while reusing and reengineering NORs from the legal domain.

2 Specificities of NORs in the Legal Domain

In this section we define NOR, providing some examples and we discuss the speci-
ficities of possible inputs (knowledge resources available for reuse) for building pos-
sible outputs (ontologies).

2.1 Non-ontological Resources

Non-ontological resources consist in:
(i) knowledge resources that embody knowledge in the legal domain;
(ii) they represent some degree of consensus;
(iii) whose semantics have not been formalized by an ontology yet, but they possess
related semantics which allow interpreting the knowledge they hold. Sometimes
this semantics is explicitly specified in natural language on the document, thus
fostering its reuse; however, in other cases, the semantics is implicit and this lack
of formalization prevents us from using them as ontologies.
Using consensuated NORs portrays benefits: it favors interoperability of the used
vocabulary, makes faster the ontology development process, lessens the knowledge
acquisition bottleneck problems, reuse, browsing/searching, and follows good prac-
tices. NORs in the legal domain can be glossaries, classification schemes, dictionaries,
taxonomies, thesauri and text. Table 1 exemplifies possible NOR for the legal domain.

2.2 Specificities of NORs in the Legal Domain

Yet, even cognizant of these benefits and amount of NORs in the legal domain, there
are specificities to look upon: they present a complex multi-layered informational
structure that should be considered when building a legal ontology. Some of these
features are recursively evoked within any legal knowledge engineering process.
i. Validity of a legal source, bounded both in time and jurisdiction;
ii. Level of formalization, in terms of being expressed in a logic formal system, as
illustrated in Fig. 1, for they can possess weak or strong semantics, in the line of
McGuinness [20]. Given the primacy of OWL ontologies, a description logics-
centric discourse is justified.
 Reuse and Reengineering of Non-ontological Resources 353

Table 1. Examples of NORs in the legal domain.

Types of Definition Examples
NOR
Glossary Terminological dictionary that Customer complaint glossarya, the
contains designations and definitions Integrated Public Sector Vocabularyb
from one or more specific subject (IPSV), Open Legal Terminology
fields. The vocabulary may be Dataset: IATE termbase,c ECRISd as
monolingual, bilingual or a shared terminology for the criminal
multilingual. (ISO 1087:2000) domain
Classification Descriptive information for an The classification of consumer
scheme arrangement or division of objects complaints from the EU Commission
into groups based on characteristics (COM(2009) 346 final)
the objects have in common (ISO
2004)
Dictionary A dictionary is a structured collection European Legal Taxonomy
of lexical units with linguistic Syllabus [12]
information about each of them (ISO
1087-1:2000)
Taxonomy A taxonomy is the simplest variant of The List of Extraordinary
controlled vocabularies as it contains Circumstances by the NEBS is a
only terms that are organized into a taxonomic examplee
hierarchical structure
Text corpora Texts are among the strongest data Dataset of complaints, legislation
available to acquire knowledge
Thesauri Thesauri are controlled vocabularies Eurovocf
of terms in a particular domain with
hierarchical, associative, and
equivalence relations between terms
(ISO 25964-1:2011)
a
http://www.jarrar.info/CContology/
b
http://doc.esd.org.uk/IPSV
c
InterActive Terminology for Europe, now also available in TermBase eXchange (TBX) format.
d
https://joinup.ec.europa.eu/sites/default/files/ckeditor_files/files/CR01(1).pdf
e
https://www.ombudsman.europa.eu
f
EuroVoc is a multilingual, multidisciplinary thesaurus covering the activities of the EU. Also
available in XML and SKOS/RDF.

iii. Hierarchy of the legal authority contained in legal sources5. The legal domain
itself defines a hierarchy of authority. Whilst legislation constitutes a primary
source of law and it is binding, therefore, its authority is explicit, known soft law

5
Legal knowledge structures are constructed in a different way than scientific knowledge structures.
Whilst the natural sciences only deal with persuasive authority, meaning that the truth of a
proposition does not depend on who states it, but only if empirical data supports it and/or is internally
consistent, the law deals with binding authority, that is, statements from a particular source whose
truth depends on that source, and other formal aspects, such as the law having been promulgated or
statement being part of a verdict ratio decidendi.
354 C. Santos et al.

Fig. 1. Examples of knowledge resources distributed according to its level of formalization. On

the left bubble, they are non-formal (i.e. not expressed in terms of a logic formal system). On the
right bubble, they are formal. The lines shows that the more to the right, the more complex the
formal system behind is.

sources comprising binding norms with a soft dimension may not be so explicit.
A possible agreed-upon typology of legal sources relies on the legal hierarchy
authority, shown in an informal way in Table 2 (for comprehension reasons and
not for a discrete selection of the valid sources). Figure 2 exemplifies a hierarchy
of knowledge sources. As illustrated, legislation, contractual terms and case law
occupy the base of the pyramid. EU Commission Interpretative Communications
and Recommendations are policy documents serving the purpose of providing
legal certainty, for they facilitate a more homogenous application of the EU

Table 2. Classification of the primary and secondary sources of law

Primary sources of law
• Refer to legislation: rules of law created by a governmental body, e.g. constitutions, statutes
and codes; regulations (from administrative agencies)
• Case-law
• Contracts
• These sources of primary law are binding
Secondary sources of law
• Legal doctrine (art. 38 (c) ICJ), which concerns legal scholarly writings and materials by legal
scholars that explain, interpret or comment primary sources of law, such as: articles, legal
commentaries, treatises, textbooks; legal encyclopedias, legal dictionaries, monographs
• Contain persuasive authority, which means that the court is not required to follow the analysis
(non-binding)
• Soft law instruments, which are interpretative sources generally making open textured concepts
operational (generally non-binding)
 Reuse and Reengineering of Non-ontological Resources 355

Fig. 2. Example of a hierarchy of knowledge resources.

Regulations and Directives, but lack on bindingness; these guidelines are

intended to tackle the issues most frequently raised by national regulators and
industry representatives. Reports and expert studies commissioned by the EU
Commission, Eurobarometer, etc. help the preparation of texts and in decision-
making, representing sources of knowledge but are non-binding.
iv. Open textured concepts. Inside the sources, vague concepts are subject to
interpretation, e.g. reasonable measures, extraordinary circumstances, etc.
v. Deontic legal operators. Deontic legal terms, such as right, obligation, prohi-
bition, permission, and sanction a.s.o. occur within legal and other normative
documents dispersedly located.
vi. Conjoint heterogeneity and fragmentation of the legal sources. In the legal
domain, NORs cannot be found in a single place, not even within one legislative
text, but in a “patchwork” of “lego” pieces. Patchwork is the expression cur-
rently used to point at conjoint heterogeneity, e.g. privacy and data protection
previous EU directive and from the intellectual property perspective, respec-
tively [13, 14]. NORs “bricks” can be embedded in different and separate
resources of hard and soft law, further articulated in case law and legal schol-
arship, scattered in a complex way in large textual corpuses, and reused in many
different ways, depending on the area of law considered. They can be found in
the sources indicated in Table 1;
vii. Citations within and among sources;
viii. Closed, shared or open status of the resources. As a result of the Open Data
movement, legally backed by the PSI-Directive6, fundamental legal sources of
democratic societies, as legislation, court decisions and Parliamentary datasets,
are freely available for reuse, and most of them have Uniform Resource Iden-
tifiers (URIs), being converted in linked data. Also the Eurovoc thesaurus, the

6
Directive 2013/37/EU, CELEX:32013L0037.
356 C. Santos et al.

IATE database, EU authority tables;7 semantic interfacing between disparate

national terminology repositories;8 Identifiers such as the European Case Law
Identifier (ECLI);9 the European Legislation Identifier (ELI)10 are open building
blocks. Figure 3 depicts this status, from internal access to anyone.
ix. Heterogeneity on its:
• type: glossaries, dictionaries, lexicons, classification schemes and tax-
onomies, thesauri, textual corpora, etc.;
• format: only some possess machine-readable format, e.g. XML, PDF, HML,
RDF, and the majority is free text, which is hard to process;
• structure: unstructured way, e.g. narratives; semi-structured, e.g. folk-
sonomies11; and structured, e.g. databases, standards, catalogues, classifi-
cations, thesauri, lexicons, legal text, among others;
x. Semantics of NORs. NORS may possess explicit and implicit semantics:
• explicit: there are hierarchies, part-of relations and other structures explicitly
expressed in natural language on the content documents, e.g. exceptions
contained in legal text; and
• implicit: interpreting the knowledge they contain, e.g. recitals of legislation;
terminologies emanated from relevant institutions with explicit definitions.
These peculiarities should be taken into account while building a legal ontology,
transferring the legal material into a computational context.

3 Methodological Reuse-Based Approaches on NORs

Research on a reuse-based approach in ontology engineering methodologies presents a

wide set of methods and tools for the ontologization of NORs, but mainly specific to a
particular resource type, or to a particular resource implementation, developing ad-hoc
solutions to transforming available resources into ontologies.
NeOn methodology provides guidelines for building ontologies by reengineering
knowledge resources widely used within a particular community. Therefore we have
used this methodology in our work, inheriting the activities of “search, assessing and

7
The EU Metadata Registry: The Metadata Registry registers and maintains definition data (metadata
elements, named authority lists, schemas, etc.) used by the different European Institutions involved in
the legal decision making process gathered in the Interinstitutional Metadata Maintenance
Committee (IMMC) and by the Publications Office of the EU in its production and dissemination
process.
8
The Legivoc project, http://legivoc.org/.
9
Council conclusions inviting the introduction of the European Case Law Identifier (ECLI) and a
minimum set of uniform metadata for case law, CELEX:52011XG0429(01).
10
Council conclusions inviting the introduction of the European Legislation Identifier (ELI),
CELEX:52012XG1026(01).
11
A folksonomy is the result of personal free tagging of information and objects (anything with an
URI) for one’s own retrieval, T. Vander Wal. Folksonomy coinage and definition. 2007. http://
www.vanderwal.net/folksonomy.html.
 Reuse and Reengineering of Non-ontological Resources 357

Fig. 3. Closed, shared and open knowledge resources, from Openlaws (https://openlaws.com/).

selecting” in the reuse process; and “reverse engineering, transformation and forward
engineering” in the reengineering process, explained in Sects. 3.1 and 3.2 respectively.
Nevertheless, NeOn does not refer to the domain specificities of legal knowledge
encountered in Sect. 1.
There is relevant precedent work on ontology design within the legal domain, in
particular, the MeLOn methodology, already implemented by a few scholars and used
flexibly in ontology development projects within a diversity of use-cases in the legal
domain [6, 15, 16]. This methodology was created for building legal ontologies in
order to help legal experts modeling legal concepts using the principles of data mod-
elisation. It comprises ten prescriptive methodological guidelines for building legal
ontologies, from specification of requirements to implementatin and placing special
emphasis to a thorough conceptual analysis and ontology evaluation12 processes.
MeLOn regards NORs in its step 4 which entails the formation of a list of all the
relevant terminology and subsequent production of a glossary of its main legal con-
cepts. Accordingly, legislation, case law and other sets of legal norms should be
consulted for determining the specific legal terminology. A glossary of terminology
should have the form of a table with these column headings: term, definition by legal
source (citing legal source, license, document, case law or legal theory, or common
custom of the legal domain), link to normative/legal source, normalised definition
(definition of term, made by the author of the new ontology, simplified or extended
from a normative/legal source to fulfill the expectations of possible methodology
users). The normalised definition should be a natural language description of the legal
text using subject, predicate, object, with the aim to reuse the terms of the glossary as

12
Evaluation parameters consist in: (i) completeness of the legal concepts definition; (ii) correctness of
the explicit relationships between legal concepts; (iii) coherence of the legal concepts modelisation;
(iv) applicability to concrete use-case; (v) effectiveness for the goals; (vi) intuitiveness for the non-
legal experts; (vii) computational soundness of the logic and reasoning; (viii) reusability of the
ontology and mapping with other similar ontologies.
358 C. Santos et al.

much as possible and avoid duplicative or ambiguous terminology. In this way a legal
expert is forced to create triples that can be aggregated later on into more abstract
assertions (TBox or ABox).
Notwithstanding the significance of the pioneering work discussed above, it leaves
space for enhancement regarding the NORs reuse process, as it provides high-level
guidelines for ontology construction, but could provide an account of methodological
steps, details and techniques employed. Three activities from NeOn could be added to this
comprehensive methodology: criteria to search NORs, assess the set of candidates and the
selection of the most appropriate NORs in the legal domain. We envisage that these
granularity (provided with definitions of the resources, tables, examples of NORs) targets
ontology practitioners with different backgrounds, encompassing domain experts, but
also ontology engineers, final users, linguists, etc. which are lay to legal specificities.

4 NORs Reuse and Engineering Processes

In this section we present the NOR reuse and the NOR reengineering processes.

4.1 NORs Reuse Process

The NOR reuse process refers to the process of choosing the most suitable available
NORs for the development of ontologies that, to some extent, cover the domain of the
ontology being built and that normally reflect some degree of community consensus.
The reuse process entails three activities: search, assessment ad selection of NORs
explained below.
(i) Searching for NORs. This activity entails searching highly reliable websites,
domain-related sites, and resources within organizations for NORs, using the
terms included in the Ontology Requirement Specification Document, hence,
according to the requirements and use-cases of the ontology;
(ii) Assessing the set of candidate NOR, using three criteria: relevance, coverage and
consensus, pursuant to the specificities delved in Sect. 1: primary and secondary
sources of law, level of formalization, status of the resources, semantics and
heterogeneity;
(iii) Selecting the most appropriate NOR to be used to build the ontology;
The purposive criteria to select and assess NOR can rely on relevance dimensions
and consensus and coverage, provided below. For each of the resource and whenever
possible, both the purpose and the components stemmed thereof should be made
explicit.
• Domain Relevance (also denominated as “domain relevance or legal authority,
legal importance”) [7, 16, 17] is two-folded, requesting the most important, or
authoritative domain documents, within the specific legal domain, which the legal
community considers relevant13;

13
Cfr. Point (iii) in Sect. 1.2 of the paper.
 Reuse and Reengineering of Non-ontological Resources 359

• Cognitive Relevance: the resources convening the users’ cognitive and informa-
tional needs. Examples are conveyed in dataset of consumer’s complaints, studies
on user’s search behaviour, studies on information-seeking behaviour of the con-
sidered users, etc.;
• Situational Relevance: the resources unfolding the user’s problems or legal cases,
which are mostly reported in case-law, in dataset of consumer com-plaints, and in
domain reports;
• Consensus and Coverage: consensus among agreed-upon knowledge is a subjective
and not quantifiable criterion. However, the reused resources should contain ter-
minology already consensuated by the legal community, therefore the effort and
time spent in finding out precise labels for the ontology terms decreased. Besides
Eur-lex (where legislation and case-law can be retrieved), the EU Commission
website on the topical domain might configure the relevant sources.
It is often the case that legal NORs in different languages have to be reused. Besides
the challenges posed by multi-jurisdictional environments, the language issues become
a problem by themselves ˗ matching elements is hardened. These problems can be
mitigated if linguistic models are used to mediate between resources. These linguistic
models, like Ontolex14 represent language information differentiating between lexical
entries, senses and concepts, easing the task of integration of cross-language resources.

4.2 Non-ontological Resource Reengineering Process

Reuse of NORs process implies their reengineering into computational ontologies,
exploiting the expressiveness and reusability of the RDF/OWL semantic web standards
for knowledge representation. This process comprises two activities [9]. The definition
of such activities and some examples are shown below. However, it is important to
consider that not all NORs should be reengineered, like legislation, as its self-contained
authoritativeness and authenticity needs to be guaranteed in its textual grounding, with
a clear reference to the texts.
(i) NOR reverse engineering, whose goal is to identify NORs’ underlying compo-
nents and then create representations of the resources at different levels of
abstraction (design, requirements and conceptual model). As an example,
provision-types and their instances can be manually harvested from the selected
sources, in order to develop a representation of the resource, a conceptual
structure (e.g. a taxonomy) or instance data for the ontology.
(ii) NOR transformation, whose goal is to generate a conceptual model from each
selected NOR. NOR transformation may include the following:
(ii.i) TBox transformation: transforms the content of the resource into an ontology
schema (generating classes, relations, instances, as depicted in Fig. 4). Forms are
usually useful to extract information due to its inherent classification scheme. As an
example, the Air Passenger EU Complaint Form depicts domain incidents and their
definitions (Fig. 5) used as classes in the RIC-ATPI domain ontology. Moreover, legal

14
https://www.w3.org/community/ontolex/wiki/Final_Model_Specification.
360 C. Santos et al.

Fig. 4. NORs transformation activity. From the schema embedded in the resource, a conceptual
model can be built.

Fig. 5. Air passenger EU complaint form as an example of a classification scheme.

theory expresses the basic concepts (also called provision-types or systemic categories)
common to (almost) all legal systems [22], e.g., obligation, permission, right, liability,
sanction, legal act, cause, entitlement, etc. Legislative documents present (most of)
these concepts and their stipulative definitions. The excerpt of the EC Regulation
261/2004 shown in Fig. 6 illustrates the extraction of anchoring provisions-types
(requisite, right, exception, etc., that constitute classes of RIC ontology) that enable its
transformation into the T-Box. The LegalRuleML metamodel [22] provides primitives
 Reuse and Reengineering of Non-ontological Resources 361

Fig. 6. Extraction of provisions-types from the EU Reg. 261/2004.

Fig. 7. Air transport passenger consumer complaint in Portuguese containing actual entities.

and their definitions, such as Permission, Obligation, Prohibition that can give a tax-
onomic skeleton of a legal ontology.
(ii.ii) ABox transformation: converts the resource schema into an ontology schema,
and resource content into ontology instances (generates classes, relations, instances and
attributes);
(ii.iii) Population: transforms the content of the resource into instances of an
existing ontology, as depicted in Fig. 7, where actual entities in this document
(“Easyjet”, “Portugal”, “a denied boarding on 2011”) will be class-instances of an
ontology (as in RIC-ATPI ontology);
362 C. Santos et al.

5 Conclusion: Discussion and Lessons Learned

This paper focuses on the specificities of NORs in the legal domain, and provides
guidelines of how some knowledge resources may be reuse and engineered following
both MeLOn and NeOn methodologies, to enable heterogeneous resources integration
within a legal ontology, as they are highly heterogeneous in their data model and
contents. We followed a text-based bottom-up approach to ontology building, in which
conceptual and terminological knowledge is contained in legal document collections
[22], demanding an expert-based analysis.
While reusing and engineering NORs, some problems occurred whilst other lessons
were learned and are hereby described and discussed. We argue that this
reuse/reengineering process should not only rely on a legal positivistic path of selecting
and interpreting norms, for legal knowledge can be used for an amplitude of situated
contexts and cases; thus, “the representation of meaning becomes a multi-faceted web
of interactions between different components (methods, data, tools, places, time,
people, organizations, users, and so forth) and importantly, this meaning is in flux”.
Hence, “reordering of subjects and objects, or to truncate concepts to be simply
attributes can render current representations of knowledge in triple form rather
cumbersome to use” [21].
Verification by domain experts was complex; mainly a presentation of drafts was
made possible. But tools as Grafoo15 are claimed to be more intuitive for a non
ontologist; it is an open source tool that can be used to present the classes, properties
and restrictions within OWL ontologies, or sub-sections thereof, as easy-to-understand
diagrams. No specific management of the knowledge sources was followed, for they
were not integrated into an information system (without a version, type, etc.) due to the
fact that they were too many to be managed (legislation, case law, doctrine, etc.) and
also considering the absence of guidelines. Much as there is criteria to add an entity or
not to an ontology (through competency questions), there is no fixed criteria to manage
legal resources. Therefore we denote that the management of the resources requires a
breed of tools to store and manage them. We posit there was a limited reproducibility of
the processes: the annotation of documents (PDFs etc.) with standard tools does not
keep track of authorship, timestamp, etc. We are cognizant that annotations tools are
necessary for commenting on NORs as a preliminary stage before building the
ontology. Nevertheless, LIME editor16 aims to annotate and connect the classes to the
texts. In order to make explicit the hidden semantics of the resource constituents, we
noticed the need of a domain expert. Furthermore, we used ad-hoc object properties of
the resource components extracted directly from the text. We acknowledge the
guidelines from MeLOn methodology to make explicit the hidden semantics in the
relations of the NOR terminology, which depend mostly on domain experts and
interpretation. We observed the need of clear criteria to select or disregard NOR
resources from the legal domain. By adapting NeOn methodology, we have decided for
three criteria: consensus, coverage and relevance dimensions (domain, situational,

15
S. Peroni, “Grafoo,” http://www.essepuntato.it/graffoo/.
16
LIME editor, http://sinatra.cirsfid.unibo.it/demo-akn/.
 Reuse and Reengineering of Non-ontological Resources 363

cognitive), but others could be accustomed. Reengineering transformation approaches

for legal text are required regarding TBox transformation, ABox transformation,
Population, forward engineering, and reverse engineering.
We plan deepening on the epistemic grounding of this position paper in the
immediate future, taking into account cooperative expert sharing in knowledge-
acquisition and a user based evaluation on using the methods for reusing and reengi-
neering NORs into ontologies to gain evidence on whether the usage of such granular
guidelines leads to users being able to design ontologies faster and/or better quality
standards [18, 19]. It is noteworthy that “resources do not explicitly carry knowledge
with them of how they were made, nor of how they should be understood, or used. Yet
such knowledge is often vital to would-be consumers” [21], hence, we envision a way
to describe some of them in a machine-readable so that users can make informed
decisions about the suitability of resources for their tasks and locate them. We aim to
use Akoma Ntoso standard [23] to provide semantic information on top of selected
legal text.

References
1. Suárez-Figueroa, M.C., Gómez-Pérez, A., Motta, E., Gangemi, A. (eds.): Ontology
Engineering in a Networked World. Springer, Dordrecht (2012). https://doi.org/10.1007/
978-3-642-24794-1
2. Breuker, J., Valente, A., Winkels, R.: Use and reuse of legal ontologies in knowledge
engineering and information management. In: Benjamins, V.R., Casanovas, P., Breuker, J.,
Gangemi, A. (eds.) Law and the Semantic Web. LNCS (LNAI), vol. 3369, pp. 36–64.
Springer, Heidelberg (2005). https://doi.org/10.1007/978-3-540-32253-5_4
3. Gangemi, A., Sagri, M.-T., Tiscornia, D.: A constructive framework for legal ontologies. In:
Benjamins, V.R., Casanovas, P., Breuker, J., Gangemi, A. (eds.) Law and the Semantic
Web. LNCS (LNAI), vol. 3369, pp. 97–124. Springer, Heidelberg (2005). https://doi.org/10.
1007/978-3-540-32253-5_7
4. Francesconi, E.: Semantic model for legal resources: Annotation and reasoning over
normative provisions. Semant. Web Leg. Domain Semant. Web 7(3), 255–265 (2016)
5. Casanovas, P.: Semantic web regulatory models. Philos. Technol. 28(1), 33–55 (2015)
6. Mockus, M., Palmirani, M.: Legal ontology for open government data mashups, pp. 113–
124 (2017). https://doi.org/10.1109/CeDEM.2017.25
7. van Opijnen, M., Santos, C.: On the concept of relevance in legal information retrieval. Artif.
Intell. Law 2017(25), 65–87 (2017)
8. Villazón-Terrazas, B., Suárez-Figueroa, M.C., Gómez-Pérez, A.: A pattern-based method for
re-engineering nonontological resources into ontologies. Int. J. Semant. Web Inf. Syst. 6(4),
27–63 (2010)
9. Villazon-Terrazas, B.M.: Method for reusing and re-engineering non-ontological resources
for building ontologies. Ph.D. thesis, UPC (2012)
10. Suárez-Figueroa, M.-C., Gómez-Pérez, A., Fernández-López, M.: The NeOn methodology
framework: a scenario-based methodology for ontology development. Appl. Ontol. 10, 107–
145 (2015)
364 C. Santos et al.

11.
Santos, C., Rodriguez-Doncel, V., Casanovas, P., van der Torre, L.: Modeling relevant legal
information for consumer disputes. In: Kő, A., Francesconi, E. (eds.) EGOVIS 2016. LNCS,
vol. 9831, pp. 150–165. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-44159-7_
11
12. Gianmaria, A., Boella, G., et al.: European legal taxonomy syllabus: a multi-lingual, multi-
level ontology framework to untangle the web of European legal terminology. Appl. Ontol.
11(4), 325–375 (2017)
13. De Hert, P., Papakonstantinou, V.: The proposed data protection regulation replacing
directive 95/46/EC: a sound system for the protection of individuals. Comput. Law Secur.
Rev. 28(2), 130–142 (2012)
14. Hunter, D., Thomas, J.: Lego and the system of intellectual property, 1955–2015, 7 March
2016. SSRN: http://ssrn.com/abstract=2743140
15. Rahman, M.: Legal ontology for nexus: water, energy and food in EU regulations.
Dissertation thesis, Alma Mater Studiorum Università di Bologna. Dottorato di ricerca in
Law, science and technology, 28 Ciclo (2016)
16. Santos, C.: Ontologies for legal relevance and consumer complaints. A case study in the air
transport passenger domain. Dissertation thesis, Alma Mater Studiorum Università di
Bologna. Dottorato di ricerca in Law, science and technology, 29 Ciclo (2017)
17. van Opijnen, M.: A model for automated rating of case law. In: 2013 ICAIL, NY, pp. 140–
149 (2013)
18. Ramakrishna, S., Górski, Ł., Paschke, A.: A dialogue between a lawyer and computer
scientist: the evaluation of knowledge transformation from legal text to computer-readable
format. Appl. Artif. Intell. 30(3), 216–232 (2016)
19. Casanovas, P., Casellas, N., Tempich, C., Vrandečić, D., Benjamins, R.: OPJK and
DILIGENT: ontology modeling in a distributed environment. Artif. Intell. Law 15(2), 171–
186 (2007)
20. McGuinness, D.: Ontologies come of age. In: Fensel, D., Hendler, J., Lieberman, H.,
Wahlster, W. (eds.) Spinning the Semantic Web: Bringing the World Wide Web to Its Full
Potential. MIT Press, Cambridge (2003)
21. Gahegan, M., Luo, J., et al.: Comput. Geosci. 35, 836–854 (2009)
22. Francesconi, E., Montemagni, S., Peters, W., Tiscornia, D.: Integrating a bottom–up and
top–down methodology for building semantic resources for the multilingual legal domain.
In: Francesconi, E., Montemagni, S., Peters, W., Tiscornia, D. (eds.) Semantic Processing of
Legal Texts. LNCS (LNAI), vol. 6036, pp. 95–121. Springer, Heidelberg (2010). https://doi.
org/10.1007/978-3-642-12837-0_6
23. Athan, T., et al.: OASIS LegalRuleML. In: Proceedings of the Fourteenth International
Conference on Artificial Intelligence and Law. ACM (2013)
24. Barabucci, G., Cervone, L., Di Iorio, A., Palmirani, M., Peroni, S., Vitali, F.: Managing
semantics in XML vocabularies: an experience in the legal and legislative domain. In: 2009
Proceedings of Balisage (2010)
 Ontology Modeling for Criminal Law

Chiseung Soh , Seungtak Lim, Kihyun Hong,

and Young-Yik Rhim(&)

Intellicon Meta Lab., Seoul 06136, Korea

{soh,stlim,kihyun.hong,ceo}@intellicon.co.kr

Abstract. In the continental law system, more attention is paid to judicial

interpretation to judge legal facts or actions than judicial precedents. Therefore,
in the continental legal system, it is appropriate to express the law itself with
knowledge such as legal ontologies or logical rules. To construct legal
ontologies and rule-based methods, legal analysis by collaboration between
legal experts and knowledge engineers should be preceded. This paper proposes
a general model for designing criminal law ontologies and rules. First, we
introduce the super-domain ontology that contains the common characteristics
of criminal law. Then, we explain the rule design method of criminal law and
present the application of the anti-graft act in Korea as an example.

Keywords: Criminal law ontology
Legal ontology model
SWRL rules

Legal reasoning
Super domain ontology

1 Introduction

Artificial intelligence (AI) has been increasingly applied to researches and services in
specialized areas such as law and medicine. Most recent AI applications utilizes data-
driven machine learning methods. However, these approaches are not transparent AI
methods and are therefore not suitable for legal applications. AI applications in the
legal field are generally required to be logically explainable implementations. In other
words, the legal AI systems need to provide clear logics about their inferences and
conclusions for solving legal problems. Therefore, machine and human readable legal
knowledge representation and logical structure-based approaches are preferable to
implement legal AI applications. There were several approaches to build such AI
systems: Generalization of legal information [1], semantic web technology [2], legal
ontology and rule designs [3, 17].
This paper presents ontology-based legal knowledge representation and logic based
legal rule design. Especially, this paper focuses on the construction of criminal law
ontology. It introduces the super-domain ontology as a general-purpose criminal law
ontology with commonality of criminal law. It also explains how to construct judgment
rules using the features of general criminal law. In addition, the proposed ontology
model is applied to the Korean anti-graft act, which is one of criminal law.
Section 2 introduces related works on legal ontology. Section 3 describes the back-
ground of this paper. This section deals with the legal system, characteristics of
criminal law, and Korean anti-graft act. The Korean anti-graft act is described as an

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 365–379, 2018.
https://doi.org/10.1007/978-3-030-00178-0_25
366 C. Soh et al.

example of criminal law. Section 4 introduces a criminal law ontology design. This
section firstly reviews existing researches on legal and criminal ontologies, and on
recent developments for legal analysis systems and legal ontology learning. Then, it
presents the main idea of this paper, which is the design of an ontology for criminal
law. Section 5 describes the application process of the proposed design for the Korean
anti-graft act. Finally, this paper concludes with a discussion of the proposed design
concept in Sect. 6.

2 Related Works

Various studies have been conducted on the design and construction of crime
ontologies [3–5]. As part of the e-Court European project1, Breuker [3] introduced
ontology construction and reusability for the Leibniz Research Institute for law
(LRI) ontology and the domain ontology of the Dutch Criminal Act (OCL.NL). The
LRI-Core ontology consists of two parts, which are a concept and legal key element
ontologies. The concept ontology describes physical, mental and abstract concepts, and
the legal key element ontology describes the legal case, legal action, legal person, etc.
Bezzazi [4] developed an ontology to identify what article of criminal law is applied to
a cybercrime. Each legal article is defined as description logic and crime cases are
classified by the logic. Bak and Jedrzejek [5] suggested an ontology model of financial
fraud. They constructed the concept-wise ontology with a modular ontology and
invented an inference method with Web Ontology Language2 (OWL) and Semantic
Web Rule Language3 (SWRL).
Many studies on rule-based legal argumentation have been conducted [6–8].
Gordon [6] showed the syntax of the Legal Knowledge Interchange Format (LKIF) rule
language and argumentation-theoretic semantics which are developed in the Euro-
pean ESTRELLA project4. The rules for those legal arguments were roughly composed
of the provisions of the German family law. Contissa [7] introduced the legal rule-
based system based on the Italian Copyright law. This is a support system creating and
deploying rule-based knowledge models Ontologies are suitable for modeling legal
knowledge because these represent resource relationships with inferable expressions.
Governatori [8] applied a rule-based approach to the business process field.
The legal ontology is made up of ontology design based on the accurate legal
analysis. Especially, legal analysis needs to be reviewed by legal experts because it
requires interpretation based on the meaning of the legal texts and the purpose of the
law. Knowledge engineers have actively worked with legal experts to conduct
researches on legal analysis, information extraction and ontology generation. Many
researches are conducted to automatically analyze legal sentences and construct legal
knowledge bases, which are tagging semantic annotation for legal sentences with NLP

1
http://cordis.europa.eu/project/rcn/56906_en.html.
2
https://www.w3.org/TR/2012/REC-owl2-primer-20121211/.
3
https://www.w3.org/Submission/SWRL/.
4
http://www.estrellaproject.org/.
 Ontology Modeling for Criminal Law 367

tools [9, 10], extracting rules from legal sentences with text structure lightweight
ontologies or NLP parser [11, 12], and generating ontology from legal documents using
NLP and ontology learning tools [13–16].

3 Legal System and Criminal Law in Korea

3.1 The Differences Between Continental Law and Common Law

Legal systems around the world generally fall into one of two main categories: con-
tinental law (civil law) and common law systems. The main difference between the two
systems is that the continental law systems are codified, whereas the common law is
normally uncodified. The continental law system is based on the statutory law and the
Pandekten-system5, and the deficiency of the law is supplemented by the precedents.
The common law system is based on the customary law and the common law, and the
precedents play a pivotal role. However, in the modern age, even common law
countries adopt the statutory law form as a new legal system. In addition, the prece-
dents are becoming more important in the continental law system, and the difference
between the legal systems is getting smaller.
The Korean legal system follows the German legal system, which is based on the
Roman law; the system has a hierarchical structure of the Constitution, law, enforce-
ment ordinance and enforcement rules. The inside of each law follows the Pandekten-
system.

3.2 Characteristics of Criminal Law

Law Types. Law is categorized into several law types, i.e., civil law, criminal law,
etc., according to its characteristics. Figure 1 shows law types and their subordinate
statutes (acts). Criminal law is a law that regulates crime and punishment. It specifies
which punishments are imposed. In contrast, civil law is a general law of private law
that regulates the rights and obligations arising from the relations between private law
actors such as private persons and private juristic persons. Administrative law deter-
mines acts on the institutions, organizations, authorities and mutual relations of
administrative entities (national and public organizations, etc.). It is also the upper laws
of those regulating the legal relationships between the public entity and persons.
Nondeterministic Structure. As shown in Fig. 2, criminal law is described as having
a nondeterministic structure as abstractly covering various real facts. On the contrary,
civil law follows a deterministic structure because its provisions are concrete and
univocal. For example, an assault offense is defined as “A person who uses violence
against another shall be punished by imprisonment for not more than two years, a fine
not exceeding five million won, detention, or a minor fine”. It is difficult to determine
the constituent requirements (corpus-delicti) of “assault” in a unique sense. All forms

5
Pandekten-system begins with the general principles, followed by separate provisions governing
particular areas of law.
368 C. Soh et al.

Commercial Criminal Anti-Graft Administrative

Act Act Act Litigation Act

Fig. 1. Hierarchy of law and law types

Criminal
Law

Deterministic Administrative Nondeterministic

Civil Law
Law

Fig. 2. Characteristics of law: deterministic vs nondeterministic

of tactics, such as sprinkling water, throwing objects, or wielding a bat, are considered
“assault”. Also, since the effect varies according to the type of “assault”, “assault” acts
as a categorical variable. In the case of bribery, it takes a continuous variable type
constituent requirement (corpus-delicti), which varies depending on the amount of
money received. Criminal law is similar to the abstracted legal requirement (corpus-
delicti) and the triggered effect structure, but the premise of criminal ontology opti-
mization is to analyze the nondeterministic structure and express it well by rules.

3.3 Korean Anti-graft Act

There are the Improper Solicitation and Graft Act6 as an anti-corruption law in Korea,
which is a part of criminal law. Traditionally, the sanctions for the prevention of
corruption are governed mainly by bribery in criminal law. However, the effectiveness
has been limited. Therefore, the Improper Solicitation and Graft Act was enacted in
accordance with OECD Anti-Bribery Convention7 and the precedents of anti-
corruption legislation in some countries shown in Table 1.
The Improper Solicitation and Graft Act is divided into the action of improper
solicitation and the action of graft. The sanctions against improper solicitation were

6
http://elaw.klri.re.kr/kor_service/lawView.do?hseq=39287&lang=ENG.
7
http://www.oecd.org/corruption/oecdantibriberyconvention.htm.
 Ontology Modeling for Criminal Law 369

Table 1. An anti-graft act of other countries

Country or organization Act title
OECD Anti-Bribery Convention
USA Foreign Corrupt Practices Acta
USA Bribery, Graft, and Conflict of Interest Actb
UK Bribery Actc
Singapore Prevention of Corrupt Actd
Singapore Corruption Confiscation of Benefits Acte
a
https://www.justice.gov/criminal-fraud/foreign-corrupt-practices-act.
b
http://law.justia.com/codes/us/2011/title-18/part-i/chapter-11.
c
http://www.legislation.gov.uk/ukpga/2010/23/contents.
d
search “Prevention of Corrupt Act” at http://statutes.agc.gov.sg/.
e
search “CONFISCATION OF BENEFITS” at http://statutes.agc.gov.sg/.

intended to eradicate corrupt practices and to prohibit public officials from improper
solicitation by enabling criminal sanctions beyond a simple code of ethics.
The penalty for graft was introduced to allow criminal punishment if a public
official receives a certain amount of money or entertainment even unintentionally.
A public official receives more than 1,000,000 KRW8 in money or entertainment from
a person who is not directly related to duties of the public official, he or she can be
punished even though there is no intention for an immediate favor. If the person is
directly related to duties of the public official, penalties are imposed even if the price is
less than 1,000,000 KRW regardless of whether it is intending the return.

4 Approach of Criminal Law Ontology

4.1 Criminal Law Ontology Construction

The basic methodology of ontology design is to distinguish between a common part
and a specific part of target design object. The common part is the set of the concepts
shared by a domain or a sub-domain. Therefore, modularization and reuse of this part is
an effective way to design an ontology. On the other hand, specific parts refer to
domain-specific concepts. The LRI-core [3] and LKIF [17] ontologies are implemented
in accordance with the basic design methodology. This paper proposes an ontology
design that extends the LRI and LKIF ontology design by defining the common part of
criminal law as a super domain. As shown in Fig. 3, the proposed design has the
common and specific parts. Briefly, upper ontology, legal core ontology, and criminal
ontology as super domains are common parts and legal domain ontologies are specific
parts.
The upper ontology is independent of the legal domain and refers to the most basic
concept. This includes abstract, physical, and mental concepts. Legal core ontology

8
KRW is the Korean currency.
370 C. Soh et al.

Common part
Upper Ontology

Legal Core
Ontology

Criminal Law
Ontology
as Super Domain
Legal Domain
Ontology
- Criminal Act
- Criminal Procedure Act
- Anti-Graft Act
Specific part - etc.

Fig. 3. A structure of legal ontology

consists of the basic elements of the law and contains the common concept of all laws.
For example, legal act, legal situation, etc. Criminal law ontology is composed of
common elements of all sub-laws (acts) of criminal law, which is called the super-
domain. This includes basic elements of criminal law such as crime, punishment, and
so on. This legal domain ontology, which reuses the elements of criminal law ontology,
is built on laws belonging to criminal law. Criminal act, criminal procedure act, and the
anti-graft act are implemented.
The design of a legal ontology is a representation of a relationship among the
elements (such as a legal object, legal actions, legal effects, etc.) required for legal
argument, the hierarchical structure between concepts, and concept description. Fur-
thermore, the ontologies distinguish between common and specific parts to facilitate
reuse and expansion of the knowledge base.

4.2 Judgment Rules of Criminal Law

This section describes the design process of judgment rules created by legal experts and
ontologists. Because criminal law follows nondeterministic structure, it is easier to
design judgment rules than other laws.
Tables 2 and 3 show the legal codes for private document counterfeit in some
countries. The underlined passages in the articles indicate corpus-delicti and effects.
The counterfeiting laws in each country are similar.
 Ontology Modeling for Criminal Law 371

Figure 4 shows a diagram for corpus-delicti (constituent requirements) of the

Article 231 of Korean Criminal Act9 on counterfeiting. The Greek alphabets in Fig. 4
indicate the related terms of the articles in Table 2 and 3, and the numbers 1, 2, 3 are
for the additional elements of ‘counterfeiting’, which are derived from legal theories
and precedents.

Fig. 4. A Diagram of constituent requirements and effects of counterfeiting in Korea

Table 2. Article on counterfeiting in Korea

9
http://elaw.klri.re.kr/kor_service/lawView.do?hseq=38891&lang=ENG.
372 C. Soh et al.

Table 3. Articles on counterfeiting in some countries

 Ontology Modeling for Criminal Law 373

The counterfeiting action in Fig. 4 can be described using SWRL expression as

follows:

hasSubjectð?action; ?xÞ ^ Personð?xÞ ^ hasprivilegeð?x; falseÞ

^ isIllegalUseOtherNameð?x; trueÞ
^ createDocumentð?x; trueÞ ! Counterfeitingð?actionÞ

The subject of the document forgery, that is, the person who forged the document
becomes a principal offender of this crime. The SWRL of the logic is expressed as
follows:

Counterfeitingð?actionÞ ^ Personð?xÞ ^ hasSubjectð?action; ?xÞ

^ hasObjectð?action; ?documentÞ
^ hasPurposeð?action; ?purposeÞ
! isprincipalOffender ð?action; ?xÞ

For example, in the case that nurse B wrote a false diagnosis in the name of doctor C
for friend A, nurse B did not qualify to write the diagnosis and he or she made a false
diagnosis by stealing the doctor’s name. Nurse B becomes a principal offender of this
crime. As shown in the case of document counterfeiting, legal arguments consist of
corpus-delicti. These forms of argument are identical in all laws, especially criminal law.

5 Application to Anti-graft Act

This chapter presents an application of the proposed criminal law ontology to anti-graft
act, which is one of criminal law. It also introduces the rules designed by analyzing the
conditional articles of this Act.

5.1 Legal Domain Ontology Construction

The anti-graft act is a sub-law of criminal law. Therefore, the anti-graft act ontology
reuses the elements of criminal law ontology and defines the elements used only in this
law. Figure 5 shows the anti-graft act ontology. The lower layers are components of the
anti-graft act ontology, which are constructed by reusing the elements of the upper,
legal-core, criminal law ontology as super domains.
Crime or Penalty Action concept (same as class of ontology) in criminal law
ontology is an upper concept of Improper Solicitation Action and Graft Action in the
anti-graft act. Public Official indicates a subject of Legal Action in the anti-graft act and
is a lower concept of Legal Person. In addition, Case Artifact is applied to express an
object for requisites of Legal Action such as in the form of food and drink, congrat-
ulatory or condolence money, gift, etc. Case Artifact such as a gift can have a value that
used in the numerical condition. Legal Case (situation) represents a set of various
actions, for example, a combination action that a public official receives gifts and is
treated to a meal simultaneously. In the anti-graft act ontology, principal entities are an
374 C. Soh et al.

Thing
Physical
Concept
Physical Physical
Process Object

Artifact
Case Action Agent Document

Legal Legal Legal Legal Case

Case Action Person Code Artifact
Penalty
Crime
Action

Graft Public
Act Gift
Action Official

Fig. 5. Layers of ontologies illustrated by relations between some typical concepts

action, a subject, an object, and a value of the object. And a variety of rules can be
expressed by the combination of these entities.
Figure 6 briefly expresses a hierarchical structure of Action Classes. The Receiv-
ingGift class is an action that has at least one person with hasReceiver and hasProvider
and at least one Gift with hasObject.

ReceivingGift  Action ^ 9 hasProvider:Person^

9hasReceiver:Person ^ 9hasObjectGift

A hasReceiver or hasProvider is a sub predicate of hasSubject, and the actions that

are restricted by this act are two subjects, the person giving (or asking) and the person
receiving (or listening to) a something (or the request).

Legal Action
Graft Action
Receiving Gift

Fig. 6. A brief hierarchy of graft action class

 Ontology Modeling for Criminal Law 375

5.2 Rules on Anti-graft Act Domain Ontology

The Korean anti-graft act has two main parts, which are the improper solicitation and
graft parts. And the act has an additional section, which regulates external lectures of
public officials. Furthermore, it describes various computational cases with numerical
limitation. Therefore, numerical calculation rules are needed. As shown Table 4, this
law allows for certain activities (such as giving food, general gifts, gifts for wedding
and funerals) to be less than a certain amount.

Table 4. Acceptable limits for foods, general gifts, and weddings & funerals gifts (Article 17 of
the Enforcement Decree)
Graft action Details Limited
money
Food Meals, dessert, alcoholic beverages, drink, etc. which KRW
the provider and public officials, etc. share 30,000
General gifts Money and other valuables except for food KRW
50,000
Gifts for weddings Congratulatory or condolence money, flowers or other KRW
and funerals equivalents 100,000

When other sub-laws of criminal law require numerical calculation rules for money
and age, the rules must be designed to that law. However, the most important thing in
any rule design is accurate legal analysis. The numbers of rules in the anti-graft act
ontology are given in Table 5.

Table 5. Number of SWRL rule per category

Category # of SWRL rule
Graft 174
Improper solicitation 7
External lecture 9
Graft-numerical calculation 38
Total 228

Some of the designed rules for the improper solicitation action and the graft action
are shown in Table 6. The first rule describes the case that a stakeholder improperly
solicits a public official through a third party; it is against the anti-graft act without an
exception. In this situation, the stakeholder is charged a fine (up to 10,000,000 KRW)
for negligence, and disciplinary and criminal punishments (not more than two years or
an administrative fine not exceeding 20,000,000 KRW) are imposed on the public
official.
376 C. Soh et al.

Table 6. The SWRL rule examples in the anti-graft act ontology

The second rule explains the graft-numerical calculator rule. When a public official
receives a present exceeding 50,000 KRW from a private person, a disciplinary pun-
ishment and a fine (2 to 5 times the price of the present) are imposed on the public
official and the private person receives a criminal punishment (not more than three
years or an administrative fine not exceeding 30,000,000 KRW).
Figure 7 describes a fictive example to demonstrate the graft-numerical calculation.
A public official “Person A” is instantiated as a subject for a gift action as a receiver.
The Receiving-Gifts has another subject “Person B” as a provider and has a gift object
with a value of more than 60,000 KRW. The receiver “Person A” has a relation to the
provider “Person B” regarding his or her duties. In this example, it is inferred that the
gift action is a penalty action when the rule for the graft-numerical calculation is
applied.
 Ontology Modeling for Criminal Law 377

Public Receiving Penalty

Official Gift Action

isA isA isA

hasReceiver hasObject
gift
Person A action
hasValue

hasDutyRelation hasProvider 60,000 KRW

Person B

Fig. 7. Reasoning on the anti-graft act ontology

6 Conclusion and Future Work

This paper has presented ontology design and make decision rules for criminal law.
These ontologies and rules were created for the legal argument, which should be
demonstrated on a clear logical basis. This paper proposed the design of the super-
domain ontology based criminal law ontology, which is composed of common ele-
ments of the laws in criminal law category. Criminal law ontology was added to the
existing European law ontology model (LKIF). This extended legal ontology model
enables to reuse the common characteristics of criminal law when construct domain
ontologies of the laws in criminal law. Also, this paper introduced the judgment rules
of criminal law, which were constructed with the general characteristics of criminal
law. Finally, this paper applied the proposed ontology structure and judgment rules to
Korea’s the anti-graft act, which is one of criminal law.
As a future work, we will extend the usage of the super-domain ontology concept to
other law types such as civil law, administrative law, industrial law, customs law, etc.
To apply the super-domain concept, careful analysis is required to extract common
features of the target law as presented in criminal law ontology design. We also
research on ontology learning, ontology population, and automatic judgment rule
design to construct domain or application ontology auto or semi automatically from
legal texts.
378 C. Soh et al.

References
1. Benjamins, V.Richard, Casanovas, P., Breuker, J., Gangemi, A.: Law and the semantic web,
an introduction. In: Benjamins, V.Richard, Casanovas, P., Breuker, J., Gangemi, A. (eds.)
Law and the Semantic Web. LNCS (LNAI), vol. 3369, pp. 1–17. Springer, Heidelberg
(2005). https://doi.org/10.1007/978-3-540-32253-5_1
2. Aguiló-Regla, J.: Introduction: legal informatics and the conceptions of the law. In:
Benjamins, V.Richard, Casanovas, P., Breuker, J., Gangemi, A. (eds.) Law and the Semantic
Web. LNCS (LNAI), vol. 3369, pp. 18–24. Springer, Heidelberg (2005). https://doi.org/10.
1007/978-3-540-32253-5_2
3. Breuker, J.: The construction and use of ontologies of criminal law in the eCourt European
project. In: Proceedings of Means of Electronic Communication in Court Administration,
pp. 15–40 (2003)
4. Bezzazi, E.-H.: Building an ontology that helps identify criminal law articles that apply to a
cybercrime case. In: ICSOFT (PL/DPS/KE/MUSE), pp. 179–185. INSTICC Press (2009)
5. Bak, J., Jedrzejek, C.: Application of an ontology-based model to a selected fraudulent
disbursement economic crime. In: Casanovas, P., Pagallo, U., Sartor, G., Ajani, G. (eds.)
AICOL -2009. LNCS (LNAI), vol. 6237, pp. 113–132. Springer, Heidelberg (2010). https://
doi.org/10.1007/978-3-642-16524-5_8
6. Gordon, T.F.: Constructing legal arguments with rules in the legal knowledge interchange
format (LKIF). In: Casanovas, P., Sartor, G., Casellas, N., Rubino, R. (eds.) Computable
Models of the Law. LNCS (LNAI), vol. 4884, pp. 162–184. Springer, Heidelberg (2008).
https://doi.org/10.1007/978-3-540-85569-9_11
7. Contissa, G.: Rulebase Technology and Legal Knowledge Representation. In: Casanovas, P.,
Sartor, G., Casellas, N., Rubino, R. (eds.) Computable Models of the Law. LNCS (LNAI),
vol. 4884, pp. 254–262. Springer, Heidelberg (2008). https://doi.org/10.1007/978-3-540-
85569-9_16
8. Governatori, G., Hashmi, M., Lam, H.-P., Villata, S., Palmirani, M.: Semantic business
process regulatory compliance checking using LegalRuleML. In: Blomqvist, E., Ciancarini,
P., Poggi, F., Vitali, F. (eds.) EKAW 2016. LNCS (LNAI), vol. 10024, pp. 746–761.
Springer, Cham (2016). https://doi.org/10.1007/978-3-319-49004-5_48
9. Biagioli, C., Francesconi, E., Passerini, A., Montemagni, S., Soria, C.: Automatic semantics
extraction in law documents. In: Proceedings of the 10th International Conference on
Artificial Intelligence and Law, pp. 133–140. ACM, New York (2005)
10. Soria, C., Bartolini, R., Lenci, A., Montemagni, S., Pirrelli, V.: Automatic extraction of
semantics in law documents. In: Proceedings of the V Legislative XML Workshop, pp. 253–
266 (2007)
11. Dragoni, M., Governatori, G., Villata, S.: Automated rules generation from natural language
legal texts. In: Workshop on Automated Detection, Extraction and Analysis of Semantic
Information in Legal Texts, San Diego, USA, pp. 1–6 (2015)
12. Dragoni, M., Villata, S., Rizzi, W., Governatori, G.: Combining NLP approaches for rule
extraction from legal documents. In: Proceedings of the Workshop on ‘MIning and
REasoning with Legal texts’ collocated at the 29th International Conference on Legal
Knowledge and Information Systems (2016)
13. Saias, J., Quaresma, P.: A methodology to create legal ontologies in a logic programming
information retrieval system. In: Benjamins, V.R., Casanovas, P., Breuker, J., Gangemi, A.
(eds.) Law and the Semantic Web. LNCS (LNAI), vol. 3369, pp. 185–200. Springer,
Heidelberg (2005). https://doi.org/10.1007/978-3-540-32253-5_12
 Ontology Modeling for Criminal Law 379

14. Lenci, A., Montemagni, S., Pirrelli, V., Venturi, G.: NLP-based ontology learning from legal
texts. a case study. In: LOAIT, pp. 113–129. CEUR-WS.org (2008)
15. Völker, J., Fernandez Langa, S., Sure, Y.: Supporting the construction of Spanish legal
ontologies with Text2Onto. In: Casanovas, P., Sartor, G., Casellas, N., Rubino, R. (eds.)
Computable Models of the Law. LNCS (LNAI), vol. 4884, pp. 105–112. Springer,
Heidelberg (2008). https://doi.org/10.1007/978-3-540-85569-9_7
16. Saias, J., Quaresma, P.: Using NLP techniques to create legal ontologies in a logic
programming based web information retrieval system (2003)
17. Hoekstra, R., Breuker, J., Di Bello, M., Boer, A., et al.: The LKIF core ontology of basic
legal concepts. In: LOAIT, pp. 43–63 (2007)
 ContrattiPubblici.org, a Semantic
Knowledge Graph on Public Procurement
Information

Giuseppe Futia1(B) , Federico Morando2 , Alessio Melandri2 , Lorenzo Canova1 ,

and Francesco Ruggiero1
1
Department of Control and Computer Engineering,
Nexa Center for Internet and Society, Politecnico di Torino, Turin, Italy
{giuseppe.futia,lorenzo.canova,francesco.ruggiero}@polito.it
2
Synapta Srl, Turin, Italy
{federico.morando,alessio.melandri}@synapta.it,
https://nexa.polito.it, https://synapta.it/

Abstract. The Italian anti-corruption Act (law n. 190/2012) requires all

public administrations to spread procurement information as open data.
Each body is obliged to yearly release standardized XML files, on its
public website, containing data that describes all issued public contracts.
Though this information is currently available on a machine-readable for-
mat, the data is fragmented and published in different files on different
websites, without a unified and human-readable view of the information.
The ContrattiPubblici.org project aims at developing a semantic knowl-
edge graph based on linked data principles in order to overcome the
fragmentation of existent datasets, to allow easy analysis, and to enable
the reuse of information. The objectives are to increase public awareness
about public spending, to improve transparency on the public procure-
ment chain, and to help companies to retrieve useful knowledge for their
business activities.

Keywords: Public procurement · Linked data · Knowledge graph

1 Introduction
In recent years the amount and variety of open data released by public bod-
ies has been factually growing1 , simultaneously with the increase of political
awareness on the topic2 . Public Sector Information (PSI)3 , in the form of open
1
See the Tracking the state of open government data report available at: http://index.
okfn.org/. Last visited July 2016.
2
See national roadmaps and technical guidelines, as well the revised of the EU Direc-
tive on Public Sector Information reuse in 2013 guidelines.
3
Public Sector Information includes “any content whatever its medium (written on
paper or stored in electronic form or as a sound, visual or audiovisual record-
ing)” when produced by a public sector body within its mandate. See more
details on Directive 2003/98/EC: http://eur-lex.europa.eu/LexUriServ/LexUriServ.
do?uri=OJ:L:2003:345:0090:0096:EN:PDF.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 380–393, 2018.
https://doi.org/10.1007/978-3-030-00178-0_26
 ContrattiPubblici.org 381

data, leads to a noticeable value for diverse actors and for different purposes,
from transparency on public spending to useful knowledge for business activities.
Open data is therefore a toolbox to improve relationships among governments,
citizens and companies by directly enabling informed decisions. Nevertheless,
access and reuse of data to build useful knowledge is extremely limited, mainly
because of the fragmentation in different data sources and websites, which cur-
rently characterizes the publication of PSI.
As defined by a World Wide Web Consortium (W3C) issue proposed by Bern-
ers Lee [2] on the publication of government data, linked data principles can be
a modular and scalable solution to overthrow the fragmentation of informa-
tion: “Linked data can be combined (mashed-up) with any other piece of linked
data. For example, government data on health care expenditures for a given geo-
graphical area can be combined with other data about the characteristics of the
population of that region in order to assess effectiveness of the government pro-
grams. No advance planning is required to integrate these data sources as long as
they both use linked data standards.” As stressed by Berners Lee, according to
these precepts linked data serves to: (1) increase citizen awareness of government
functions to enable greater accountability; (2) contribute valuable information
about the world; and (3) enable the government, the country, and the world to
function more efficiently.
Public procurement is an area of the PSI that could largely benefit from
linked data technologies. As argued by Svátek [10], an interesting aspect of
public contracts from the point of view of linked data is the fact that “they
unify two different spheres: that of public needs and that of commercial offers.
They thus represent an ideal meeting place for data models, methodologies and
information sources that have been (often) independently designed within the
two sectors.” At the same time, linked data is beneficial in the public contracts
domain since it gives ample space for applying diverse methods of data analytics,
performing complex alignments of entities in a knowledge graph and developing
data driven applications.
The contribution is structured as follows. Section 2 presents related works in
the field of public procurement and spending information published according
to linked data principles. Section 3 describes the Italian context and gives an
overall view of public procurement data spread by public sector bodies. Section 4
illustrates the data processing pipeline to improve the quality of data source and
to create the ContrattiPubblici.org knowledge graph. Section 5 shows results
and potential use of the information structured in the graph. The last section
describes conclusions and future advancements of the work.

2 Related Works
In this section, we report contributions in which public procurement and spend-
ing data is transformed and published as knowledge graphs, following the linked
data principles.
Public procurement domain has already been addressed by several works and
projects developed in the linked data field. One of the most notable is the LOD2
382 G. Futia et al.

project, since it systematically addressed many phases of procurement linked

data processing [10]. Such project exploits the Public Procurement Ontology
PPROC4 .
There are several other initiatives: the TWC Data-Gov Corpus [4], Public-
spending.gr [8], The FTS (Financial Transparency System) project [7], Linked
Spending [6], LOTED [11] and MOLDEAS [1].
In particular, the TWC Data-Gov Corpus gathers linked government data
on US financial transactions from the Data.gov project5 . This project exploits
a semantic-based approach, in order to incrementally generate data, support-
ing low-cost and extensible publishing processes, and adopting technologies to
incrementally enhance such data via crowdsourcing. Publicspending.gr has the
objective of interconnecting and visualizing Greek public expenditure with linked
data to promote clarity and increase citizen awareness through easily-consumed
visualization diagrams. The FTS (Financial Transparency System) project of
the European Commission contains information about grants for EU projects
starting from 2007 to 2011, and publishes such data according to the RDF6
data model. Exploring such dataset, users are able to get an overview on EU
funding, including data on beneficiaries as well as the amount and type of expen-
diture. Linked Spending is a project for the RDF conversion of data published
by the OpenSpending.org, an open platform that releases public finance infor-
mation from governments around the world. The project uses the RDF Dat-
aCube vocabulary7 to model data in order to represent multidimensional sta-
tistical observations. LOTED8 is focused on extracting data from procurement
acts, aggregating it over a SPARQL9 endpoint. Finally, MOLDEAS, between the
other things, presents some methods to expand user queries to retrieve public
procurement notices in the e-Procurement sector using linked open data.

3 Context and Data Source

The Italian Legislative Decree n. 33/2013 (DL33/2013) of March 14th, 201310
re-ordered obligations of disclosure, transparency and dissemination of informa-
tion by public administrations. According to specific requirements defined by
the decree (clause no.9 - DL33/2013), each body is required to create a spe-
cific section on its website called “Amministrazione Trasparente” (Transparent

4
More details on Public Procurement Ontology PPROC available at: http://contsem.
unizar.es/def/sector-publico/pproc.html.
5
Data.gov project website: https://www.data.gov/.
6
RDF (Resource Description Framework) is a standard model for data interchange
on the Web. It represents a common format to achieve and create linked data.
7
DataCube vocabulary information: https://www.w3.org/TR/vocab-data-cube/.
8
LOTED project website: http://www.loted.eu/.
9
SPARQL (SPARQL Protocol and RDF Query Language) is a semantic query lan-
guage for databases, able to retrieve and manipulate data stored in RDF format.
10
http://www.decretotrasparenza.it/wp-content/uploads/2013/04/D.Lgs .-n.-
332013.pdf. Last visit on July 2017.
 ContrattiPubblici.org 383

Administration). In this section, public administrations provide details related

to public procurement, with particular emphasis on procedures for the award
and execution of public works, services and supplies (clause no. 37 - DL33/2013).
Such data is published on the basis of a precise XML Schema Definition11 (XSD)
provided by ANAC - Autorità Nazionale Anticorruzione (the Italian National
Anti-Corruption Authority)12 , which has supervisory duties. After the publica-
tion on their websites, administrations transmit information on public contracts,
in a digital format, to ANAC. ANAC then performs a preliminary check and
releases an index file containing details related to the availability of data13 .
Public bodies can publish and transmit to ANAC two types of XML files.
The first type contains the actual data on contracts until the publication date
(January 31st of each year). As mentioned before, in order to facilitate the
consistency of publications and the comparison of information, the structure of
the document is defined by a precise XSD Schema14 . The main structure of the
XML file includes a section with the metadata of the dataset, reported in Fig. 1,
and a section containing multiple contracts. The metadata section lists diverse
information including the first publication date and the last dataset update,
the business name of the public body that spreads the dataset, the url of the
dataset and the license. The section containing data on contracts includes: the
identification code of the tender notice (CIG that stands for Codice Identificativo
Gara), the description of the tender identified by the CIG, the procedure type
for the selection of the contractor, the identification code (VAT number) and the
business name of bidders (tender participants), the identification code and the
business name of the tender notice beneficiary, the award amount, the amount
paid, the date of commencement and completion of works (for more details of
each field see Figs. 2 and 3). In Sect. 3.2 we describe the ontology used to map
those fields in the linked data domain to build the semantic knowledge graph of
ContrattiPubblici.org.
The second type of XML is an index that collects links to other XML files
containing actual public procurement data (Fig. 4)15 . This data is available in
machine-readable format according to a well-defined schema; a semantic layer to
interconnect such data is necessary for the assessment of transparency, it is effec-
tive to raise awareness about public spending, and to provide useful information
for enterprises.

11
XSD is a W3C recommendation that specifies how to formally describe an XML
document.
12
http://www.anticorruzione.it/.
13
The index is available at https://dati.anticorruzione.it/#/l190, clicking on the
“Esporta” (Export) button.
14
The full representation of the XSD schema is available at http://dati.anticorruzione.
it/schema/datasetAppaltiL190.xsd.
15
A more clear representation of the XSD schema is available at http://dati.
anticorruzione.it/schema/datasetIndiceAppaltiL190.xsd.
384 G. Futia et al.

Fig. 1. XSD Schema of metadata

Fig. 2. XSD Schema of public procurement data - Part 1

3.1 Data Quality Issues

The quality of public contracts data is one of the most important issues to be
tackled in order to reduce fragmentation and build a semantic knowledge graph.
 ContrattiPubblici.org 385

Fig. 3. XSD Schema of public procurement data - Part 2

Let’s consider, for example, a company that has participated in two calls for
tender proposed by two different bodies. VAT number and business name of
this company reported in the two XML files should be identical. However, some
data errors may occur due to management processes and software: this prevents
to generate a unique entity (the company itself) within the knowledge graph.
For example, a VAT number can present a wrong character (accuracy issue),
or even the field itself could be absent (completeness issue)16 . Analyzing the
VAT number issues, we have observed that 62,466 contracts (1.08% of the total)
present accuracy problems like wrong characters, and 60,731 contracts (1.05%
of the total) do not present the VAT number field in the data (completeness
problems).
For this reasons, different checks must be implemented with the aim of cor-
recting, where possible, the wrong data [12]. Section 4 describes the process we
have implemented to tackle data quality issues in order to build a semantic graph
upon public contracts data.

3.2 Ontology

In order to model the data source to build the ContrattiPubblici.org knowledge

graph, we decided to use the Public Contracts Ontology (PCO) developed by
the Czech OpenData.cz initiative17 . The authors of this ontology are modeling
“information which is available in existing systems on the Web” and “which will
be usable for matching public contracts with potential suppliers” [5]. Therefore,
the goal is to model a public contract as a whole, but without going into details
of the public procurement domain.

16
Such data quality metrics are defined by the International Organization for Stan-
dardization: ISO/IEC 2501.
17
The Public Contracts Ontology is available on GitHub platform at: https://github.
com/opendatacz/public-contracts-ontology.
386 G. Futia et al.

Fig. 4. XSD Schema of public procurement dataset index

In the PCO, a contract notice is a call for tenders, which may be submitted
for the award of a public procurement contract. Therefore, we are able to map
XML fields described in Sect. 3 into entities, classes and relations provided by
the PCO. Figure 5 shows precisely the data model adopted for building the
ContrattiPubblici.org knowledge graph. Although there is a significant degree
of overlap between the XSD that describes the data model of Italian public
contracts and the PCO, we had to introduce measures to better describe our
domain. For instance, the concept of tender was not fully expressed in the data
model adopted in XML files, since there are only information about participants
(inclusive of VAT numbers and company names), but not information related to
offering services and prices. Nevertheless, the tender is one of the most important
entity in the PCO to link the bidders to the public contract. For these reasons,
during the conversion to linked data (Sect. 4), we decided to create tender entities
in our knowledge graph using as identifier the VAT number of the participant
and the CIG of the contract.

4 Data Processing
When data derives from legacy databases, the publication of linked data is not
always immediate [9]; data frequently comes from different sources and it needs
to be gathered in a single file before proceeding with the conversion/translation
into RDF triples (the so-called triplification) [3]. In the following section we show
the process we used to obtain linked data as final result (Fig. 6).
 ContrattiPubblici.org 387

Fig. 5. A scheme of the data model used to build the ContrattiPubblici.org knowledge
graph

Fig. 6. Linked data conversion pipeline

4.1 Harvesting

As explained in Sect. 3, ANAC releases an index file that provides URLs of avail-
able XMLs, which are published on public administrations websites. Based on
such index, the Download component tries to fetch data distinguishing between
two different cases. In the first case, the component downloads and locally stores
the XML containing public contracts data, with additional metadata related to
the download outcome. In the second case, if fetched XMLs are indexes to files
containing real data, the component is able to cross the links chain18 and apply
the download process shown in the first case. When the component is not able to
recognize the expected schema of an XML with data or an XML index, it saves
the file apart for a later manual check. In most cases, this means that either the

18
In some cases an index points to another index that finally might point to a file, or
to another index.
388 G. Futia et al.

URLs are wrong or the resource is not published according to an accepted for-
mat (e.g., it is a PDF file). In the worst cases, XML indexes are recursive, since
they contain URLs that references to the XML index itself. For these reasons, we
implemented some features in the Download component in order to manage this
critical issue that threatens to undermine the entire pipeline. Moreover, during
the download operation a lot of servers do not reply, for several reasons. We
collected more than 10 different HTTP responses, which reveal how the quality
of service over the 15000+ infrastructures of the Italian public administration
might not be reliable.

4.2 Cleaning
By analysing the collected XMLs, we noticed that the quality of data is frag-
mented: different people with different systems led to inconsistencies and errors,
as the example shown in Sect. 3.1. Therefore, we implemented a Cleaning and
Conversion module that tries to guess potential errors for each field and attempts
to correct data and define a standard format. In particular:
– dates are converted into the ISO 8601 format (YYYY-MM-DD);
– a digit check is performed on the CIGs (identifiers for each procurement) and
on VAT numbers for detecting errors and verifying the syntactical correctness;
– agreed prices and payments, which are intended to be euro values, are casted
to float number with two decimal digits;
– procedure types, which are fixed text categories, are checked with a function
that calculates the similarity between strings. Such function tries to attribute
unconventional values to one of those predefined categories.
Every value is analyzed and pushed in a result file serialized in JSON. If a
value is modified by the Cleaning and Conversion component, both values, the
original one and the guessed one, will be stored in this result file. Furthermore,
a reference to the original XML file (the authoritative data) is included in such
file.

4.3 Conversion to Linked Data

After the Cleaning stage, we are able to convert the data into RDF using the N-
Triples serialization19 . During this conversion procedure, a component matches
each field of the JSON file with the respective property (or relation) and data
type from the Public Contract Ontology. When necessary, it also creates the
needed entities (see the example of tender entities explained in Sect. 3.2).
One issue we faced during the graph creation is the companies’ labels man-
agement. While merging data from heterogeneous XML files, it is frequent to
find different labels referring to the same company20 . The creation of the labels
19
More information available at: https://www.w3.org/TR/n-triples/.
20
The differences may be minimal, as in presence of spelling errors, or even consider-
able.
 ContrattiPubblici.org 389

triples is therefore handled by an algorithm that chooses the most common label
referring to a company.
The last step in the conversion procedure is the so-called interlinking. Inter-
linking means declaring that an entity is same as another entity in another
dataset, by adding new links to external resources and generating the so called
knowledge graph. For this purpose, we chose the SPCData database21 , provided
by the Agenzia per l’Italia Digitale, that contains the index of Italian public
administrations. The knowledge graph is thus created by matching the VAT
numbers of the original graph with the ones in the SPCData database.
After this procedure, the completed RDF file is pushed into a Triple Store
that exposes data via a SPARQL endpoint.

5 Results

The semantic knowledge graph of ContrattiPubblici.org is published using the

Virtuoso Triple Store22 , one of the most adopted technological solution to process
and spread linked data. By using the SPARQL endpoint provided by Virtuoso,
advanced users and robots are able to satisfy complex information needs. Table 1
details the total amount of RDF triples, entities, contracts, public bodies, com-
panies, and external links to other datasets.

Table 1. Amount of public procurement information available on July 2017

RDF triples 168,961,163

Entities 22,436,784
Contracts 5,783,968
Public bodies 16,593
Companies 652,121
Links to external datasets 13,486

Due to data quality issues illustrated in Sect. 3.1, the number of companies
presented in Table 1 is slightly overestimated. As explained in the next session,
some future work will be dedicated to implementing further checks and expedi-
ents to merge different instances of the same company in a well-defined entity
within the knowledge graph.
Despite these problems, in the context of transparency and open spending,
it is possible to identify cases of public contracts with anomalies that require
more investigation. With the following SPARQL query23 , for instance, advanced
21
More information available at: http://spcdata.digitpa.gov.it/index.html.
22
More information available at: https://virtuoso.openlinksw.com/.
23
The endpoint to perform the query is available at: https://contrattipubblici.org/
sparql.
390 G. Futia et al.

users and robots are able to get a list of 100 contracts in which the payment by
the public body is more than doubled of the agreed price.

PREFIX pc: <http://purl.org/procurement/public-contracts#>

PREFIX payment: <http://reference.data.gov.uk/def/payment#>
SELECT ?contract ?amount ?agreedPrice WHERE {
?contract pc:agreedPrice ?agreedPrice.
?contract payment:payment ?payment.
?payment payment:netAmount ?amount.
FILTER (?amount > 2*?agreedPrice)
} LIMIT 100

Fig. 7. Total amount of money (in Euros) assigned to a single beneficiary

Fig. 8. Number of call for tenders won by a single beneficiary

In addition to data analysis via SPARQL queries, users can exploit fea-
tures of a human-consumption interface. For these reasons, we have developed
a Web application that is available at: http://public-contracts.nexacenter.org/.
Through a dedicated search form, users can enter the VAT number of a public
administration, obtaining a visualization that shows different information. For
example, a ranking of the top 10 beneficiaries on the basis of the total-allocated
 ContrattiPubblici.org 391

Fig. 9. Extent (in Euros) of the public call for tenders divided by year

amounts (Fig. 7), or a ranking on the basis of the total number of contracts
(Fig. 8) by means of histograms. Or even a view on contracts by means of a
bubble diagram, allowing to compare very clearly the size of contracts put out
to tender by the body (Fig. 9). Through this kind of visualizations, an inter-
ested company could obtain an overview of tenders and contracts size, acquiring
an increased knowledge on how to allocate its investments. Other features on
the Web interface, including the search for individual contracts on the basis of
keywords and new types of visualization, will be developed in the future.

6 Conclusions and Future Work

This contribution outlines opportunities in building a semantic knowledge graph

based on linked data principles in the context of the Italian public procurement
data. Despite the difficulties to tackle issues related to coverage and quality of
open data sources, handling public contracts information in linked data enables
novel avenues to evaluate the transparency of public administration and to create
new business opportunities.
As shown in the Results Sect. 3.1, ContrattiPubblici.org semantic graph fos-
ters the extraction of useful knowledge in different ways, ranging from SPARQL
queries for complex analyses on data to rich and interactive visualizations. Fur-
thermore, linked data principles lead to an enhanced interoperability across var-
ious data formats and Web applications, unleashing the full value of PSI.
Future work on ContrattiPubblici.org will most likely concentrate on the
interlinking of the knowledge graph with more external datasets. To support
this task, on the one hand, we plan to develop a component to automatize the
process of Entity Identification, in particular for business entities. On the other
392 G. Futia et al.

hand, we will include metadata from DCAT-AP IT24 ontology to increase the
semantic expressiveness of contracts data. Moreover, PCO described in Sect. 3.2
will be extended with more details respect to the period of time of the contract,
considering, for example, any interruption of the works, and the provenance of
data. Further improvements for the research project are related to the addition
of new data quality tests on procurement information, involving legal experts to
control and validate the data after the cleaning process, and the development of
new human-consumption interfaces.

References
1. Álvarez, J.M., Labra, J.E., Calmeau, R., Marı́n, Á., Marı́n, J.L.: Query expansion
methods and performance evaluation for reusing linking open data of the European
public procurement notices. In: Lozano, J.A., Gámez, J.A., Moreno, J.A. (eds.)
CAEPIA 2011. LNCS (LNAI), vol. 7023, pp. 494–503. Springer, Heidelberg (2011).
https://doi.org/10.1007/978-3-642-25274-7 50
2. Berners-Lee, T.: Putting government data online (2009)
3. Canova, L., Basso, S., Iemma, R., Morando, F.: Collaborative open data ver-
sioning: a pragmatic approach using linked data. In: CeDEM15 - Conference for
E-Democracy and Open Governement, pp. 171–183. Edition Donau-Universität
Krems, Krems (2015). http://porto.polito.it/2617308/
4. Ding, L., et al.: TWC data-gov corpus: incrementally generating linked government
data from data.gov. In: WWW (2010)
5. Distinto, I., dAquin, M., Motta, E.: Loted2: an ontology of European public pro-
curement notices. Semant. Web 7(3), 267–293 (2016)
6. Höffner, K., Martin, M., Lehmann, J.: LinkedSpending: openspending becomes
linked open data. Semant. Web J. (2015). http://www.semantic-web-journal.net/
system/files/swj923.pdf
7. Martin, M., Stadler, C., Frischmuth, P., Lehmann, J.: Increasing the financial
transparency of European commission project funding. Seman. Web J. Spec. Call
Linked Dataset Descr. 5(2), 157–164 (2013). http://www.semantic-web-journal.
net/system/files/swj435 0.pdf
8. Vafolopoulos, M., et al.: Publicspending. gr: interconnecting and visualizing Greek
public expenditure following linked open data directives, July 2012. http://www.
w3.org/2012/06/pmod/pmod2012 submission 32.pdf
9. Rowe, M., Ciravegna, F.: Data. dcs: Converting legacy data into linked data. In:
LDOW, p. 628 (2010)
10. Svátek, V., Mynarz, J., Wecel,
 K., Klı́mek, J., Knap, T., Nečaský, M.: Linked
open data for public procurement. In: Auer, S., Bryl, V., Tramp, S. (eds.) Linked
Open Data – Creating Knowledge Out of Interlinked Data. LNCS, vol. 8661, pp.
196–213. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-09846-3 10

24
More information on the Italian profile of the DCAT-AP defined in the context of ISA
(Interoperability solutions for public administrations, businesses and citizens) pro-
gram of the European Commission is available at: https://www.dati.gov.it/content/
dcat-ap-it-v10-profilo-italiano-dcat-ap-0.
 ContrattiPubblici.org 393

11. Valle, F., dAquin, M., Di Noia, T., Motta, E.: Loted: exploiting linked data in
analyzing European procurement notices. In: Proceedings of the 1st Workshop on
Knowledge Injection into and Extraction from Linked Data - KIELD 2010 (2010).
http://sisinflab.poliba.it/sisinflab/publications/2010/VDDM10
12. Vetró, A., Canova, L., Torchiano, M., Minotas, C.O., Iemma, R., Morando,
F.: Open data quality measurement framework: definition and applica-
tion to open government data. Gov. Inf. Q. 33(2), 325–337 (2016).
http://www.sciencedirect.com/science/article/pii/S0740624X16300132
 Application of Ontology Modularization
for Building a Criminal Domain Ontology

Mirna El Ghosh1(&), Hala Naja2, Habib Abdulrab1,

and Mohamad Khalil3
1
LITIS, INSA, Rouen, France
{mirna.elghosh,habib.abdulrab}@insa-rouen.fr
2
Faculty of Sciences, Lebanese University, Tripoli, Lebanon
hala.naja70@gmail.com
3
Faculty of Engineering, CRSI Research Center,
Lebanese University, Tripoli, Lebanon
mohamad.khalil@ul.edu.lb

Abstract. The Ontology modularization is an essential field in the ontology

engineering domain helping to reduce the complexity and the difficulties of
building, reusing, managing and reasoning on domain ontologies either by
applying partitioning or composition approaches. This paper carries out a survey
on ontology modularization and presents a modular approach to build criminal
modular domain ontology (CriMOnto) for modelling the legal norms of the
Lebanese criminal system. CriMOnto, which will be used later for a legal rea-
soning system, is composed of four independent modules. The modules will be
combined together to compose the whole ontology.

Keywords: Ontology modularization
Modular ontology

Ontology composition
Ontology integration
Ontology reuse
Criminal domain
CriMOnto

1 Introduction

It is commonly known that ontologies aim to capture consensual knowledge of a given

domain in a generic and formal way, to be reused and shared across applications and by
groups of people [1]. The number of available ontologies has increased considerably in
various domains such as bioinformatics, genetics, medicine and law, among many
others, but they are also becoming larger and more complex to manage and reuse [2]. In
this context, interest in modularization techniques, as an ontology engineering prin-
ciple, has increased to resolve the problems of reusability, scalability and maintenance
of ontologies [3]. In addition to this, since the main use of ontologies is making the
intended meaning of a given domain available to all agents, an ontology conceptual
architecture is required to represent this meaning. According to [4], there is a need to
modularize the conceptual architecture to represent this intended meaning dealing with
the complexity of the domain such as heterogeneous knowledge with different levels of
detail of that knowledge. Therefore, the resulted designed ontologies are obtained with
a high quality.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 394–409, 2018.
https://doi.org/10.1007/978-3-030-00178-0_27
 Application of Ontology Modularization for Building a Criminal 395

Our motivation is to build a legal ontology-based system that performs reasoning

tasks in order to support the lawyers and judges in their decisions, specifically, in the
criminal domain. Such a system needs a legal domain ontology for modelling the legal
norms in order to improve the efficiency of reasoning that requires the representation of
the whole semantics of the criminal domain, including well-known properties such as
axioms [5]. In order to reduce the complexity of the building process of the criminal
domain ontology, to make it as much as possible reusable and able to perform efficient
reasoning, we propose to modularize it. According to [6], a modular ontology usually
contains a set of modules (component theories), implemented in same or different
languages, and a set of semantic relations among those modules.
These semantic relations are obtained by applying an integration process in order to
link the ontology modules together. At the end, a criminal modular ontology, named
CriMOnto, is obtained. Therefore, the resulting structure of CriMOnto is a Modular
Conceptual Architecture. The domain application of this research is the Lebanese
criminal system and the Lebanese criminal code is considered as the main textual
resource since it contains the legal norms of the criminal domain. The main goals of
this study are: survey the domain of ontology modularization for a better understanding
of this domain and then provide a modular conceptual architecture of CriMOnto as well
as the modular construction process. The global view of this architecture is seen as an
integration of different ontology modules developed independently.
The remainder of this paper is organized as follows: In Sect. 2, the ontology
modularization challenges are outlined. In Sect. 3, the motivation of the study is
introduced. In Sect. 4 the ontology modularization concept and approaches are
investigated. Section 5 defines the modular architecture of CriMOnto as well as the
integration process. The work is evaluated in Sect. 6. Finally, Sect. 7 concludes the
paper.

2 Ontology Modularization Challenges

Ontology modularization and the problem of formally characterizing a modular rep-

resentation for ontologies are great challenges in the ontological engineering domain.
Actually, this domain suffers from lack of the modularization theory which uncovers
number of unanswered questions [7, 8]:
• How complex ontologies can be built up from parts?
• In what ways can those parts be related (mapping, integration)?
• How the structure of a modular ontology can be represented?
• to create an ontology module with a certain purpose or use-case in mind:
• Which modularity type of module is defined?
These questions declare various aspects such as: define a modularization approach,
precise the modules types and define an integration process for combining modules. By
giving answers to these questions, a modular architecture for CriMOnto is introduced.
396 M. El Ghosh et al.

3 Study Motivation

According to [3], most of the existent ontologies, even if they implicitly relate several
sub-domains, are not structured in a modular way. In the other hand, several works tend
to combine different ontologies together, such as [9, 10], implicitly without the explicit
definition of ontology modularization concept. For [11], in any realistic application, it
is often desirable to integrate different ontologies, developed independently, into a
single, reconciled ontology. This would allow for the modular design of large
ontologies and would facilitate knowledge reuse tasks. Thus, there is no universal way
to modularize ontologies and that the choice of a particular technique should be guided
by the requirements of the considered application [3]. Therefore, an approach aims at
designing a modular architecture as well as an incremental process allowing a col-
laborative building of CriMOnto. The main features of CriMOnto are: it is composed of
four independently developed ontology modules. These modules are obtained from
different and heterogeneous sources (partial reuse of existent ontologies and semi-
automatic extraction mechanisms from textual resources). CriMOnto is empowered
with an integration process to combine the different modules.

4 Ontology Modularization

The main idea of modularization originates from the general notion of modular soft-
ware in the area of software engineering [12]. In software engineering domain, the
modularity is a well-established notion where it refers to a way of designing software in
a clear, well-structured way that supports maintenance and reusability [13]. However,
in the ontology engineering domain, the notion of modularization and the problem of
formally characterizing a modular representation for ontologies are not as well
understood [14], which causes suffer in the existing work and prevents further devel-
opment [3]. Despite this vagueness, ontology modularization is considered as a major
topic in the field of formal ontology developments and a way to facilitate and simplify
the ontology engineering process [7]. Moreover, ontology modularization has several
benefits where modular representations are easier to understand, reason with, extend
and reuse [13]. Therefore, using these representations tends to reduce the complexity of
designing and to facilitate the ontology reasoning, development, and integration.

4.1 What Is an Ontology Module?

Generally speaking, a module is a part of a complex system that functions indepen-
dently from this system [15]. In contrast to the software engineering domain, the notion
of ontology module is not clear or understood in the domain of ontological engineering
[16]. There is a need to formalize and define an ontology module, particularly in terms
of its requirements [17]. For [18], ontology module is considered as extractable part
that can be reused outside the context of the general ontology. More clearly, an
ontology module is defined by [16] as “An ontology module is a reusable component of
a larger or more complex ontology, which is self-contained but bears a definite
relationship to other ontology modules including the original ontology”. This
 Application of Ontology Modularization for Building a Criminal 397

definition implies that ontology modules can be reused either as they are, or by
extending them with new concepts, and relationships. Each ontology module is con-
sidered as ontology itself since it can be extended with new concepts and relationships.
Thereby, ontology modules are themselves ontologies [12].

4.2 Ontology Module Criteria

The criteria of ontology modules generally aim at characterizing modular ontologies in
order to evaluate the quality of modules [4]. Generally, inspired by the software
engineering domain, three main criteria a module should fulfill: self-contained, loose
coupling and high cohesion [19, 20]. Therefore, in the ontological engineering domain,
some studies, such as [21–23], believe that modularization criteria should be defined in
terms of the applications for which the modules are created. They defined some
ontology module criteria such as: Encapsulation, Independence and Domain coverage.

4.3 Ontology Module Classifications

In the literature, the authors of [24] proposed to analyze modules by taking as central
the following question: modules for what? Based on this question, various classifica-
tions of ontology modules are proposed such as [8, 25–27]. Meanwhile, the most useful
classification of ontology modules is based on their content where the ontologies are
classified into five main categories [28–30]:
• Generic, or top-level, ontologies: describe generic concepts independently of a
particular domain or problem.
• Core ontologies: in contrast to generic ontologies that span across many fields, core
ontologies describe the basic categories within a domain such as law.
• Domain ontologies: specialize a subset of generic ontologies in a domain or sub-
domain, e.g. criminal law.
• Application or domain-specific ontologies: developed for a specific application.

4.4 Approaches of Ontology Modularization

Generally, modularization denotes the possibility to perceive a large knowledge
repository as a set of modules, i.e. smaller repositories that, in some way, are parts of
and compose the whole knowledge [31]. Therefore, an ontological modularization
process is seen as a call for organizing ontologies into modules which could then be
reused and combined in novel ways [7]. In the literature, three different approaches of
ontology modularization are found [3, 12, 17]: ontology composition, ontology
decomposition and module extraction. In this work, the ontology composition approach
is used. Ontology composition aims to develop independently a set of ontology
modules and assemble them coherently and uniformly, by means of integrating and
mapping, to form a wider ontology. Examples of ontology composition approaches are,
among others, [32–35].
398 M. El Ghosh et al.

5 Application of Ontology Modularization in the Legal

Domain: CriMOnto

CriMOnto is a criminal modular domain ontology for modelling the norms of the
Lebanese criminal system as a domain application for this study. In previous work [36],
a middle-out approach is proposed for building CriMOnto where the ontology mod-
ularization techniques are discussed implicitly. In this work, the approach is enhanced
and explicit modularization techniques are applied for this purpose. The aim of this
approach is to show how ontology modularization can simplify and reduce the com-
plexity of ontology building processes. Therefore, a modular architecture of the
ontology is outlined by identifying the main modules, their number, type and criteria,
as well as the knowledge to be represented in each module. Moreover, ontology reuse
process, which is now one of the important research issues in the ontology field [37], is
recommended as a key factor to develop cost effective and high quality ontologies.
Actually, ontology reuse reduces the cost and the time required for building ontologies
from scratch [38–40]. Moreover, by reusing validated ontology components, the
quality of the newly implemented ontologies is increased. According to [41], there are
two main reuse processes: merge and integration. Merge is the process of building an
ontology in one subject reusing two or more different ontologies on that subject.
Meanwhile, integration is the process of building an ontology in one subject reusing
one or more ontologies in different subjects that are maybe related. In the current work,
the proposed conceptual architecture of CriMOnto is grounded on four different level
concepts: Upper ontology module (UOM), Core ontology module (COM), Domain
ontology module (DOM) and Domain-specific ontology module (DSOM). The pro-
posed approach is defined by developing the modules independently by using top-
down and bottom-up strategies and then combining them together to compose the
whole CriMOnto (see Fig. 1) [36]. From this perspective, the different modules are in
different subjects since they are in different conceptual levels. Therefore, an integration
process is performed to combine them. Inspired by the integration methodology of
[37], there are list of activities that precede the integration of modules into the resulting
ontology such as identify the knowledge to be represented in the different modules as
well as the candidate ontologies to be reused. Therefore, an analyzing and selection
process will take place in order to define the existent ontologies to be reused. In the
following, these activities are outlined.

Fig. 1. Middle-out approach for building CriMOnto.

 Application of Ontology Modularization for Building a Criminal 399

5.1 Upper Ontology Module

The UOM consists of abstract concepts and relations which are effectively independent
of any specific domain. For a well-founded building of this module, a partial reuse of
existent foundational, or top-level, ontologies can help. These ontologies are theoret-
ically well-founded domain independent systems of categories that have been suc-
cessfully used to improve the quality of conceptual models and semantic
interoperability [42]. In addition to this, partial reuse of foundational ontologies can
facilitate and speeding up the ontology development process by preventing to reinvent
the wheel concerning basic categories [43]. In the literature, several works seek for
reusing concepts from foundational ontologies in order to support in maintaining a
well-structured construction of domain ontologies that could serve as a future reusable
artifact [44]. Various foundational ontologies exist for reuse such as DOLCE [45] and
UFO [46]. In this context, a selection for an appropriate foundational ontology is a
crucial and difficult step since it depends on different elements such as: the purpose of
building the ontology and the applicability domain. Therefore, after studying the main
concepts of these foundational ontologies, UFO is the most convenient.
The unified foundational ontology UFO is a foundational ontology initially pro-
posed by Guizzardi and Wagner [47]. UFO is developed to support the activities of
both conceptual and organizational modeling [48]. Therefore, UFO permits the
building of an ontology reusing some generic concepts such as category, kind, subkind,
relator, role and role mixin where the ontologist does not need to rebuild these con-
cepts. UFO is divided into three layered sets: (1) UFO-A, ontology of objects, defines
terms related to endurants such as universal, relator, role, intrinsic moment; (2) UFO-
B, ontology of events, defines terms related to perdurants such as event, state, atomic
event, complex event; (3) UFO-C defines terms related to intentional and social entities
including linguistic aspects such as social agent, social object, social role and nor-
mative description [49]. In the current work, UFO-B and UFO-C are needed to ground
the criminal domain ontology for building UOM since they define some basic concepts
for the criminal domain such as Agent, Intentional_Moment, Action, Event, and Nor-
mative_Description. In order to make possible the activity of conceptual modeling via
UFO, a conceptual modeling language, named OntoUML [50], was proposed.
OntoUML uses the ontological constraints of UFO as modeling primitives and is
specified above the UML 2.0 meta-model [51]. To build, evaluate and implements
OntoUML models, a model-based environment is needed such as the standalone tool
OLED (OntoUML Lightweight Editor) [52] or the extension of UML production-grade
tool Enterprise Architect [53].
Reusing Concepts from UFO-C. There is list of essential concepts in UFO-C to reuse
for building the upper module, mainly those related to social entities such as Agents and
Objects [54] (see Fig. 2a). Agents can be physical (e.g. Person) or social (e.g. Orga-
nization). Objects are also categorized in physical (e.g. book) and social objects (e.g.
normative description). Normative_Description defines one or more rules/norms rec-
ognized by at least one Social_Agent. Regulations and constitutions are examples of
normative description. In Fig. 2a, a fragment of the conceptual model of the upper
module, which is represented in OntoUML language, is depicted.
400 M. El Ghosh et al.

Fig. 2. Fragment of the upper module in OntoUML (a) and Protégé (b).

Therefore, for knowledge representation and reasoning capabilities, there is a need

to transform the conceptual model of the upper module represented in OntoUML
language to a computational ontology language such as OWL. For this purpose, OLED
features defined automatic transformations of the OntoUML models to OWL files [55]
that can be managed using ontology editors such as Protégé (see Fig. 2b).
Reusing Concepts from UFO-B. The ontology of perdurants UFO-B defines Event,
which is a basic concept in the criminal domain (e.g. crime is an event), as a main
category [54]. In UFO-B, events can be atomic or complex [56] and an Action or
Participation (Fig. 3). Actions are performed by agents and considered as intentional
events caused by intentions [54]. Participation can be for agents and objects. Therefore,
participation of an agent can be intentional or unintentional.

5.2 Core Ontology Module

The COM consists of concepts and relations that are common across the domains of
law and can provide the basis for specialization into domain and domain-specific
concepts. The same perspective can be applied, as for upper module, for reusing
existent legal core ontology to build this module. Thus, it is not easy to define
appropriate legal core ontology among the existent (LKIF-Core, LRI-Core, and
FOLAW). Actually, LKIF-Core [57] is the most recent legal-core ontology and con-
tains essential legal concepts such as Medium, Document, Legal_Source, Legal_Doc-
ument, and Code (see Fig. 4).
 Application of Ontology Modularization for Building a Criminal 401

Fig. 3. Fragment of the upper module in OntoUML.

Fig. 4. Excerpt of core concepts reused from LKIF-Core.

5.3 Domain Ontology Module

The DOM is composed of categories that are related mainly to the criminal domain in
general such as Criminal_Act, Penalty, Misdemeanor, Violation, etc. In order to build
this module, two main strategies are applied: (1) specialize the concepts and relations of
the core module; (2) extract the knowledge from textual resources. The strategy of
specializing the concepts and relations of the core module is applied. Therefore,
concepts of the core module such as: Creation, Public_Act, Act_Of_Law, and Reaction
are specialized in domain concepts such as: Criminal_Act, Punishments and Defence
(see Fig. 5a). For the strategy of extracting the knowledge from textual resources, a
bottom-up approach is applied as an ontology learning process, with the support of
NLP techniques, for extracting the main elements of the domain ontology module from
the English version of the Lebanese criminal code (for more details refer to [58]).
Examples of extracted concepts are shown in the figure such as Defender, Offender,
Accessory, Accomplice, Instigator, Perpetrator and victim (see Fig. 5b).
402 M. El Ghosh et al.

Fig. 5. Excerpt of the domain concepts in Protégé.

5.4 Domain-Specific Ontology Module

The DSOM consists of concepts and relations of a specific subject domain such as the
Lebanese criminal system. Domain-specific ontologies are useful in systems involved
with reasoning. Thus, they should be at higher level of expressivity, in other words,
rich in axioms. To build this module, since it is related directly to the Lebanese criminal
system, an ontology learning process is applied on the criminal code in order to extract
semi-automatically the domain-specific ontology and part of the domain ontology
module [58]. The bottom-up strategy helped to generate semi-automatically an OWL
ontology including the basic elements: concepts, instances, taxonomies, relations and
axioms. Unfortunately, some of the generated results were inexpressive and thus
insufficient for practical use. The resulted ontology module is considered as lightweight
or semi-formal. Meanwhile, a more expressive ontology is required. For this reason and
inspired by the work of [59], a reengineering process is applied to correct, prune and
enrich the extracted ontology and make it more expressive by transforming it to
heavyweight or axiom-based ontology. In the Fig. 6, a fragment of the domain-specific
module concepts are depicted such as Fine_Contravention_Penalty and Imprison-
ment_Contravention_penalty, which are specific concepts in the Lebanese criminal
code, are considered as sub-concepts of the Contravention_Penalty concept of the
domain module.

Fig. 6. Excerpt of the domain-specific ontology module represented in Protégé.

 Application of Ontology Modularization for Building a Criminal 403

5.5 Evaluation and Integration of Ontology Modules

After building the ontology modules (upper, core, domain and domain-specific), there
is a need to evaluate them according to the modules criteria in order to compose the
resulting ontology CriMOnto. Actually, the independent building of the modules
makes them self-contained and enhances the encapsulation where any changes that
affect a module will not affect the others. Moreover, the independency leads to make
them reusable as well. Additionally, each module covers an adequate level of the given
domain where their integration will create a complete coverage of the Lebanese
criminal system. Furthermore, an integration process will be applied to aggregate,
combine, and assemble the modules. Subsequently, in the resulting ontology, regions
that were taken from the integrated modules can be identified. Moreover, it is essential
to consider heterogeneity resolution and related ontology matching or mapping
strategies to be an internal part of ontology integration [39]. In this context, list of
semantic mappings will be created among concepts of the different modules. Generally,
the mapping concept is defined by [60] as a morphism. Meanwhile, in this study a more
loosely definition is used based on some works in the literature such as [61] and claims
such as “simple mappings methods are sufficient and outperform more complex
methods” [62]. For ontology mappings, several studies, such as [63, 64], have proposed
a number of specialized semantics. Meanwhile, for other studies, such as [65], ontology
mappings are represented as OWL2 axioms of the form subClassOf, EquivalentClass
and DisjointClass [66]. In CriMOnto, the modules are located on vertical conceptual
levels from general (upper module) to specific (domain-specific module). For this
reason, the mappings will be based mainly on a parent-child, or subsumption, hierar-
chical relationship [67] and established manually as structural axiom of the form
subClassOf. Thus, the hierarchical relationship is established among the concepts of
modules. For this purpose, a linguistic-based matcher, such as WordNet, is used to deal
with ontology mapping for calculating the similarity values between concepts [68].
Then a domain expert, knowledgeable about the semantics of legal concepts, validates
the proposed mappings. Given two concepts Ci and Cj from the modules UOM and
COM respectively, if Cj is considered as a subclass of Ci, then the subClassOf axiom is
added between the two concepts in the resulting ontology. In CriMOnto, Event is a
general concept from the upper module UOM and Legal_Event is a legal concept from
the core module COM. Legal_Event is considered as a subclass of Event. Thus, the
subClassOf axiom is added between them (see Fig. 7).

Fig. 7. Hierarchical mapping in Protégé.

404 M. El Ghosh et al.

6 Evaluation

After building CriMOnto, there is a need to evaluate the characteristics and the validity
of the resulting ontology. According to [69], evaluation is required during the whole
life-cycle of an ontology in order to guarantee that what is built meets the requirements.
In the literature, various approaches to the evaluation of ontologies have been
considered [75]. They are classified mainly in different levels according to ontology
quality metrics [76]: lexical and concept/data [77], taxonomic and semantic relations
[78], context-level [79], application-based [80] and data-driven [81]. The kind and the
purpose of the ontology define the evaluation level. Generally, evaluation methods
consist of two parts: verification that ensures the ontology is constructed correctly, and
validation that ensures the ontology represents the real world [70]. For the verification,
Ontology taxonomy evaluation method is applied manually by domain experts while
integrating the ontology modules and by using a logic reasoner that checks the con-
sistency of the ontology such as Pellet. For the validation, the application-based
method is applied [71] by adding list of concrete individuals to the ontology. Actually,
CriMOnto is intended to be used for building a rule-based reasoning system composed
of set of formal logic rules. The output of this system will depend mainly on the quality
of CriMOnto.

7 Conclusion

In this work, a modular approach is described to build a criminal domain ontology

(CriMOnto). In the literature, several works applied implicitly the ontology modular-
ization such as [72] for the automation domain, [73] for ill-defined domains and [74]
for the enterprise architecture domain. To the best of our knowledge, there is no works
in the legal domain that deal explicitly with the ontology modularization for building
legal domain ontologies. For this purpose, this direction is tracked. The aim of this
study is to prove how ontology modularization can simplify the development of well-
founded legal domain ontologies with the support of ontology reuse and integration.
Furthermore, a modular architecture of CriMOnto is outlined. The architecture is
composed of four independent ontology modules. The upper and core modules are
developed by reusing existent foundational and legal core ontologies. Meanwhile, the
domain and domain-specific are constructed with the support of ontology learning
process from legal texts. Furthermore, an integration process, based on simple hier-
archical mappings, is applied to combine the modules in order to compose the whole
CriMOnto. Finally, this approach is considered as useful track for the ontologists who
seek to build domain ontologies based on existent valid ontologies and not totally from
scratch.

Acknowledgements. This work has been supported by the European Union with the European
Regional Development Fund (ERDF) under Grant Agreement no. HN0002134 in the project
CLASSE 2 (“Les Corridors Logistiques: Application a la Vallée de la Seine et son Environ-
nement”), Lebanese University and the National Support from the National Council for Scientific
Research in Lebanon (CNRS).
 Application of Ontology Modularization for Building a Criminal 405

References
1. Corcho, O., Fernández-López, M., Gómez-Pérez, A.: Ontological engineering: what are
ontologies and how can we build them? In: Jorge, C. (ed.) Semantic Web Services: Theory,
Tools and Applications, pp. 44–70. IGI Global, Hershey (2007)
2. Pathak, J., Johnson, T.M., Chute, C.G.: Modular ontology techniques and their applications
in the biomedical domain. Integr. Comput. Aid. Eng. 16(3), 225–242 (2009)
3. d’Aquin, M., Schlicht, A., Stuckenschmidt, H., Sabou, M.: Ontology modularization for
knowledge selection: experiments and evaluations. In: Wagner, R., Revell, N., Pernul, G.
(eds.) DEXA 2007. LNCS, vol. 4653, pp. 874–883. Springer, Heidelberg (2007). https://doi.
org/10.1007/978-3-540-74469-6_85
4. Gangemi, A., Catenacci, C., Battaglia, M.: Inflammation ontology design pattern: an
exercise in building a core biomedical ontology with descriptions and situations. Stud.
Health Technol. Inform. 102, 64–80 (2004)
5. Fürst, F., Trichet, F.: Integrating domain ontologies into knowledge-based systems. In:
FLAIRS Conference, pp. 826–827 (2005)
6. Wang, Y., Bao, J., Haase, P., Qi, G.: Evaluating formalisms for modular ontologies in
distributed information systems. In: marchiori, m, Pan, Jeff Z., Marie, C. (eds.) RR 2007.
LNCS, vol. 4524, pp. 178–193. Springer, Heidelberg (2007). https://doi.org/10.1007/978-3-
540-72982-2_13
7. Hois, J., Bhatt, M., Kutz, O.: Modular ontologies for architectural design. In: FOMI-09,
Frontiers in Artificial Intelligence and Applications, vol. 198. IOS Press, Vicenza (2009)
8. Khan, Z.C., Keet, C.M.: Toward a framework for ontology modularity. In: The 2015 Annual
Research Conference on South African Institute of Computer Scientists and Information
Technologists. SAICSIT 2015, Article No. 24 (2015)
9. Francesconi, E., Montemagni, S., Peters, W., Tiscornia, D.: Integrating a bottom–up and
top–down methodology for building semantic resources for the multilingual legal domain.
In: Francesconi, E., Montemagni, S., Peters, W., Tiscornia, D. (eds.) Semantic Processing of
Legal Texts. LNCS (LNAI), vol. 6036, pp. 95–121. Springer, Heidelberg (2010). https://doi.
org/10.1007/978-3-642-12837-0_6
10. Saias, J., Quaresma, P.: A methodology to create legal ontologies in a logic programming
information retrieval system. In: Benjamins, V.Richard, Casanovas, P., Breuker, J.,
Gangemi, A. (eds.) Law and the Semantic Web. LNCS (LNAI), vol. 3369, pp. 185–200.
Springer, Heidelberg (2005). https://doi.org/10.1007/978-3-540-32253-5_12
11. Turlapati, V.K.C., Puligundla, S.K.: Efficient module extraction for large ontologies. In:
Klinov, P., Mouromtsev, D. (eds.) KESW 2013. CCIS, vol. 394, pp. 162–176. Springer,
Heidelberg (2013). https://doi.org/10.1007/978-3-642-41360-5_13
12. Abbes, S.B., Scheuermann, A., Meilender, T., d’Aquin, M.: Characterizing modular
ontologies. In: 7th International Conference on Formal Ontologies in Information Systems
(FOIS), pp. 13–25 (2012)
13. Grau, B.C., Horrocks, I., Kazakov, Y., Sattler, U.: A logical framework for modularity of
ontologies. In: IJCAI 2007, pp. 298–303. AAAI Press (2007)
14. Grau, B.C., Kutz, O.: Modular ontology languages revisited. In: The IJCAI-2007 Workshop
on Semantic Web for Collaborative Knowledge Acquisition (2007)
15. Konev, B., Lutz, C., Walther, D., Wolter, F.: Formal properties of modularisation. In:
Stuckenschmidt, H., Parent, C., Spaccapietra, S. (eds.) Modular Ontologies. LNCS, vol.
5445, pp. 25–66. Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-642-01907-4_3
16. Doran, P.: Ontology reuse via ontology modularization. In: Proceedings of Knowledge Web
Ph.D. Symposium, pp. 1–6 (2006)
406 M. El Ghosh et al.

17. Bezerra, C., Freitas, F., Zimmermann, A., Euzenat, J.: ModOnto: a tool for modularizing
ontologies. In: WONTO-08, vol. 427 (2008). ceur-ws.org
18. Grau, B.C., Parsia, B., Sirin, E., Kalyanpur, A.: Modularity and web ontologies. In: KR,
pp. 198–209 (2006)
19. Stuckenschmidt, H., Klein, M.: Reasoning and change management in modular ontologies.
Data Knowl. Eng. 63(2), 200–223 (2007)
20. Stuckenschmidt, H., Klein, M.: Integrity and change in modular ontologies. In: 18th
International Joint Conference on Artificial Intelligence, pp. 900–905 (2003)
21. d’Aquin, M., Schlicht, A., Stuckenschmidt, H., Sabou, M.: Criteria and evaluation for
ontology modularization techniques. In: Stuckenschmidt, H., Parent, C., Spaccapietra, S.
(eds.) Modular Ontologies. LNCS, vol. 5445, pp. 67–89. Springer, Heidelberg (2009).
https://doi.org/10.1007/978-3-642-01907-4_4
22. Del Vescovo, C., Parsia, B., Sattler, U., Schneider, T.: The modular structure of an ontology:
an empirical study. Technical report, University of Manchester. http://www.cs.man.ac.uk/%
7Eschneidt/publ/modstrucreport.pdf
23. Del Vescovo, C., Parsia, B., Sattler, U., Schneider, T.: The modular structure of an ontology:
an empirical study. In: Haarslev, V, Toman, D., Weddell, G. (eds.), DL 2010, vol.
573 (2010). ceur-ws.org
24. Borgo, S.: Goals of modularity: a voice from the foundational viewpoint. In: Kutz, O.,
Schneider, T. (eds.) Fifth International Workshop on Modular Ontologies, Frontiers in
Artificial Intelligence and Applications, vol. 230, pp. 1–6. IOS Press (2011)
25. Studer, T.: Privacy preserving modules for ontologies. In: Pnueli, A., Virbitskaite, I.,
Voronkov, A. (eds.) PSI 2009. LNCS, vol. 5947, pp. 380–387. Springer, Heidelberg (2010).
https://doi.org/10.1007/978-3-642-11486-1_32
26. Del Vescovo, C., et al.: The modular structure of an ontology: atomic decomposition. In:
IJCAI Proceedings-International Joint Conference on Artificial Intelligence, vol. 22 (2011)
27. Cuenca Grau, B., Halaschek-Wiener, C., Kazakov, Y.: History matters: incremental
ontology reasoning using modules. In: Aberer, K., et al. (eds.) ASWC/ISWC -2007. LNCS,
vol. 4825, pp. 183–196. Springer, Heidelberg (2007). https://doi.org/10.1007/978-3-540-
76298-0_14
28. Guarino, N., Carrara, Giaretta, P.: An ontology of meta-level categories. In: Doyle, J.,
Sandewall, E., Torasso, P. (eds.) Principles of Knowledge Representation and Reasoning:
Proceedings of KR94. Morgan Kaufmannn, San Mateo (1994)
29. Guarino, N.: Understanding, building, and using ontologies. Int. J. Hum. Comput. Stud. 46
(2–3), 293–310 (1997)
30. Van Heijst, G., Schreiber, A., Th Wielinga, B.G.: Using explicit ontologies in KBS
development. Int. J. Hum. Comput. Stud. 46, 2–3 (1997)
31. Stuckenschmidt, H., Christine, P., Spaccapietra, S.: Modular Ontologies: Concepts, Theories
and Techniques for Knowledge Modularization. Springer, Berlin (2009). https://doi.org/10.
1007/978-3-642-01907-4
32. Ben Mustapha, N., Baazaoui-Zghal, H., Moreno, A., Ben Ghezala, H.: A dynamic
composition of ontology modules approach: application to web query reformulation. Int.
J. Metadata Semant. Ontol. 8(4), 309–321 (2013)
33. Bezerra, C., Freitas, F., Euzenat, J., Zimmermann, A.: An approach for ontology
modularization (2008)
34. Dmitrieva, J., Verbeek, F.J.: Creating a New Ontology: A Modular Approach. arXiv preprint
arXiv:1012.1658 (2010)
35. Steve, G., Gangemi, A., Pisanelli, D.: Integrating medical terminologies with onions
methodology (1998)
 Application of Ontology Modularization for Building a Criminal 407

36. El Ghosh, M., Naja, H., Abdulrab, H., Khalil, M.: Towards a middle-out approach for
building legal domain reference ontology. Int. J. Knowl. Eng. 2(3), 109–114 (2016)
37. Pinto, H., Martins, J.: Ontology integration: how to perform the process. In: The
International Joint Conference on Artificial Intelligence, pp. 71–80 (2001)
38. Bontas, E.P., Mochol, M., Tolksdorf, R.: Case studies on ontology reuse. In: IKNOW05
International Conference on Knowledge Management, vol. 74 (2005)
39. Caldarola, E.G., Picariello, A., Rinaldim A.M.: An approach to ontology integration for
ontology reuse in knowledge based digital ecosystems. In: 7th International Conference on
Management of computational and Collective intElligence in Digital EcoSystems, pp. 1–8.
ACM (2015)
40. Modoni, G., Caldarola, E., Terkaj, W., Sacco, M.: The knowledge reuse in an industrial
scenario: a case study. In: The Seventh International Conference on Information, Process,
and Knowledge Management eKNOW 2015, pp. 66–71 (2015)
41. Pinto, S.H., Gomez-Perez, A., Martins, J.P.: Some issues on ontology integration. In:
IJCAI99’s Workshop on Ontologies and Problem Solving Methods: Lessons Learned and
Future Trends (1999)
42. Guizzardi, G.: The role of foundational ontology for conceptual modeling and domain
ontology representation. In: 7th International Baltic Conference on Databases and
Information Systems, pp. 17–25 (2006)
43. Keet, M.: The use of foundational ontologies in ontology development: an empirical
assessment. In: 8th Extended Semantic Web Conference, Greece, vol. 6643, pp. 321–335
(2011)
44. Rosa, D.E., Carbonera, J.L., Torres, G.M., Abel, M.: Using events from UFO-B in an
ontology collaborative construction environment. CEUR-WSX 938, 278–283 (2012)
45. Masolo, C., Borgo, S., Gangemi, A., Guarino, N., Oltramari, A.: Wonderweb deliverable
D18 (ver. 1.0). Ontology Library (2003)
46. Guizzardi, G., Wagner, G.: A unified foundational ontology and some applications of it in
business modeling. In: CAiSE Workshops, vol. 3, pp. 129–143 (2004)
47. Guizzardi, G., Wagner, G.: Using UFO as a foundation for general conceptual modeling
languages. In: Poli, R., Healy, M., Kameas, A. (eds.) Theory and Applications of Ontology:
Computer Applications, pp. 175–196. Springer, Dordrecht (2010). https://doi.org/10.1007/
978-90-481-8847-5_8
48. Melo, S., Almeida, M.B.: Applying foundational ontologies in conceptual modeling: a case
study in a Brazilian public company. In: Meersman, R. (ed.) On the Move to Meaningful
Internet Systems: OTM 2014 Workshops, pp. 577–586. Springer, Heidelberg (2014). https://
doi.org/10.1007/978-3-662-45550-0_59
49. Guizzardi, G., Wagner, G.: Towards ontological foundations for agent modelling concepts
using the unified fundational ontology (UFO). In: Bresciani, P., Giorgini, P., Henderson-
Sellers, B., Low, G., Winikoff, M. (eds.) AOIS -2004. LNCS (LNAI), vol. 3508, pp. 110–
124. Springer, Heidelberg (2005). https://doi.org/10.1007/11426714_8
50. Guizzardi, G.: Ontological foundations for structural conceptual models. Ph.D. thesis.
Enschede, Telematica Institut, The Netherlands (2005)
51. Guerson, J., Sales, T.P., Guizzardi, G., Almeida, J.P.A.: OntoUML lightweight editor: a
model-based environment to build, evaluate and implement reference ontologies. In: IEEE
19th International Enterprise Distributed Object Computing Workshop (EDOCW), pp. 144–
147 (2015)
52. http://code.google.com/p/ontouml-lightweight-editor/
53. http://www.sparxsystems.com/products/ea/
408 M. El Ghosh et al.

54. Guizzardi, G., Falbo, R. A., Guizzardi, R.S.S.: Grounding software domain ontologies in the
unified foundational ontology (UFO): the case of the ODE software process ontology. In:
Proceedings of the Ibero American Workshop on Requirements Engineering and Software
Environments, pp. 244–251 (2008)
55. Barcelos, P.P.F., dos Santos, V.A., Silva, F.B., Monteiro, M.E., Garcia, A.S.: An automated
transformation from OntoUML to OWL and SWRL. In: ONTOBRAS 2013. CEUR
Workshop Proceedings, vol. 1041, CEUR-WS.org, pp. 130–141 (2013)
56. Guizzardi, G., Wagner, G., Falbo, A., Guizzardi, R.S.S., Almeida, J.P.A.: Towards
ontological foundations for the conceptual modeling of events. In: 32th International
Conference, ER 2013, pp. 327–341 (2013)
57. Hoekstra, R., Breuker, J., Bello, M.D., Boer, A.: The LKIF core ontology of basic legal
concepts. In: Workshop on Legal Ontologies and Artificial Intelligence Techniques, CEUR
Workshop Proceedings, vol. 321, pp. 43–63 (2007)
58. El Ghosh, M., Naja, H., Abdulrab, H., Khalil, M.: Ontology learning process as a bottom-up
strategy for building domain-specific ontology from legal texts. In: The 9th International
Conference on Agents and Artificial Intelligence, ICAART, vol. 2, pp. 473–480 (2017)
59. Gómez-Pérez, A., Rojas-Amaya, M.D.: Ontological reengineering for reuse. In: Fensel, D.,
Studer, R. (eds.) EKAW 1999. LNCS (LNAI), vol. 1621, pp. 139–156. Springer, Heidelberg
(1999). https://doi.org/10.1007/3-540-48775-1_9
60. Kalfoglou, Y., Schorlemmer, W.M.: Ontology mapping: the state of the art. In: Semantic
Interoperability and Integration (2005)
61. Dmitrieva, J., Verbeek, F.: Modular approach for a new ontology. In: 5th International
Workshop on Modular Ontologies WoMO (2011)
62. Ghazvinian, A., Noy, N.F., Musen, M.A.: Creating mappings for ontologies in biomedicine:
simple methods work. In: AMIA 2009 Symposium Proceedings (2009)
63. Euzenat, J.: Semantic precision and recall for ontology alignment evaluation. In: IJCAI,
pp. 348–353 (2007)
64. Borgida, A., Serani, L.: Distributed description logics: assimilating information from peer
sources. J. Data Semant. 1, 153–184 (2003)
65. Jimenez-Ruiz, E., Cuenca Grau, B., Horrocks, I., Berlanga, R.: Ontology integration using
mappings: towards getting the right logical consequences. Technical report, Universitat
Jaume, University of Oxford (2008)
66. Cuenca Grau, B., Horrocks, I., Motik, B., Parsia, B., Patel-Schneider, P., Sattler, U.: OWL 2:
the next step for OWL. J. Web Semant. 6(4), 309–322 (2008)
67. Wang, Y., Liu, W., Bell, D.: A concept hierarchy based ontology mapping approach. In:
KSEM, pp. 101–113 (2010)
68. Miller, G.A.: WordNet: a lexical database for English. Commun. ACM 38, 39–41 (1995)
69. Hartmann, J., et al.: Methods for ontology evaluation. In: Knowledge Web Deliverable
D1.2.3 (2004)
70. Gómez-Pérez, A., Fernandez-Lopez, A., Corcho, O.: Ontological Engineering. Springer,
London (2004). https://doi.org/10.1007/b97353
71. Gómez-Pérez, A.: Evaluation of ontologies. Int. J. Intell. Syst. 16, 391–409 (2011)
72. Legat, C.: Semantics to the shop floor: towards ontology modularization and reuse in the
automation domain. In: World Congress (2014)
73. Thakker, D., Dimitrova, V., Lau, L., Denaux, R., Karanasios, S., Yang Turner, F.: A priori
ontology modularisation in ill-defined domains. In: 7th International Conference on
Semantic Systems, I-Semantics 2011, pp. 167–170 (2011)
74. Bakhshandeh, M., Antunes, G., Mayer, R., Borbinha, J., Caetano, A.: A modular ontology
for the enterprise architecture domain. In: 17th IEEE International Enterprise Distributed
Object Computing Conference Workshops, EDOCW 2013, pp. 5–12 (2013)
 Application of Ontology Modularization for Building a Criminal 409

75. Brank, J., Grobelnik, M., Mladenic, D.: A survey of ontology evaluation techniques. In:
Conference on Data Mining and Data Warehouses (SiKDD) (2005)
76. Gangemi, A., Catenacci, C., Ciaramita, M., Lehmann, J.: Modelling ontology evaluation and
validation. In: Sure, Y., Domingue, J. (eds.) ESWC 2006. LNCS, vol. 4011, pp. 140–154.
Springer, Heidelberg (2006). https://doi.org/10.1007/11762256_13
77. Maedche, A., Staab, S.: Measuring similarity between ontologies. In: Gómez-Pérez, A.,
Benjamins, V.Richard (eds.) EKAW 2002. LNCS (LNAI), vol. 2473, pp. 251–263.
Springer, Heidelberg (2002). https://doi.org/10.1007/3-540-45810-7_24
78. Brewster, C., Alani, H., Dasmahapatra, S., Wilks, Y.: Data driven ontology evaluation. In:
International Conference on Language Resources and Evaluation, Lisbon (2004)
79. Ding, L., et al.: Swoogle: a search and metadata engine for the semantic web. In: CIKM,
pp. 652–659 (2004)
80. Porzel, R., Malaka, R.: A task-based approach for ontology evaluation. In: ECAI 2004
Workshop Ontology Learning and Population, Valencia, Spain, pp. 1–6 (2004)
81. Patel, C., Supekar, K., Lee, Y., Park, E.: OntoKhoj: a semantic web portal for ontology
searching, ranking and classification. In: Proceedings of the 5th ACM International
Workshop on Web Information and Data Management. ACM (2004)
 A Linked Data Terminology for Copyright
Based on Ontolex-Lemon

Víctor Rodriguez-Doncel1(&) , Cristiana Santos2,

Pompeu Casanovas2,3 , Asunción Gómez-Pérez1 ,
and Jorge Gracia1
1
Ontology Engineering Group, Universidad Politécnica de Madrid,
Madrid, Spain
vrodriguez@delicias.fi.upm.es
2
Institute of Law and Technology, Autonomous University of Barcelona,
Barcelona, Spain
3
La Trobe University, Melbourne, Australia

Abstract. Ontolex-lemon is the de facto standard to represent lexica relative to

ontologies and it can be used to encode term banks as RDF. A multi-lingual, multi-
jurisdictional term bank of copyright-related concepts has been published as
linked data based on the ontolex-lemon model. The terminology links information
from WIPO (concepts and definitions), IATE (multilingual terms, usage notes)
and other sources as Creative Commons (multilingual definitions) or DBpedia
(general concepts). The terms have been hierarchically arranged, spanning mul-
tiple languages and targeting different jurisdictions. The term bank has been
published as a TBX dump file and is publicly accessible as linked data. The term
bank has been used to annotate common licenses in the RDFLicense dataset.

Keywords: Term bank
Linked data
Copyright
Legal localization

Multilingualism

1 Introduction

Legal translations, namely the translations of texts within the field of law, are among
the most difficult types of translations. The legal system referred by the source text may
be different from the legal system referred by the target text, and the translation of the
parts with a specific legal significance must be particularly precise at ensuring the
correspondence of concepts at both sides. The mistranslation of a clause in a contract
can lead to lawsuits or loss of money.
A term bank (also known as term base or more informally as terminology) is a
database of concepts and terminological data related to a particular field. Terminologies
help keeping translations consistent and help choosing the most adequate term when
precision is required. Further, the localization of legal texts require of specialized
terminologies where the exact concept in a legal system must be invoked.
The work presented in this paper describes a terminology created in a half-
automated process, where terms and their definitions have been extracted and inte-
grated from different lexical sources and mapped in a supervised process.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 410–423, 2018.
https://doi.org/10.1007/978-3-030-00178-0_28
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 411

The resulting terminology has been published1 in the TBX format – ISO 30042 [1]
– which is the standard for the exchange of terminologies; and it has also been pub-
lished in Resource Description Format (RDF)2, according to the schema described by
Cimiano et al. [4]. The RDF version is especially suitable for establishing links with
other resources (like DBpedia3) and with other terminologies. IATE4, the inter-
institutional database of the European Union (EU), has been taken as the external
reference for some of the extracted terms.
Plain texts can be annotated, makin-g reference to concepts or terms in a term bank.
This work also presents the text of a license that has been annotated with the terms in
the copyright terminology here presented.
The use of a terminology of legal terms found in licenses is not exhausted with the
mere translation or localization. Once in a digital format, it can alleviate the task of
identifying the key elements in new licenses as in [5] or can help the study of com-
parative law.
The paper is organized as follows. Section 1 describes the motivation for having a
term bank of copyright-related terms published as linked data. Details on the followed
methodology and publication are given in Sects. 2 and 3 provides the related work and
finally Sect. 4 contains the conclusions and future work.

2 Motivation: Legal Term Banks as Linked Data

The representation of copyright and related rights constitutes a part of legal knowledge
currently at the limelight of European policy. Progress has been made in delivering
copyright-related actions identified in the Digital Agenda5, the Intellectual Property
Strategy6 and in the “Licences for Europe”7 . The European Commission has presented
legislative proposals to make sure that consumers and creators can make the most of the
digital world: the reviewed EU copyright rules consists of a regulation8 and a directive9
on copyright in the Digital Single Market.

1
The copyright terminology is online at: http://copyrighttermbank.linkeddata.es
2
http://www.w3.org/RDF/
3
http://dbpedia.org/
4
http://iate.europa.eu/
5
Communication on content in Digital Single Market (COM(2012) 789 final)
6
In order to modernise the EU copyright legislative framework, “A Single Market for Intellectual
Property Rights” (COM(2011) 287 final) was announced, which proposed series of measures to
promote an efficient copyright framework for the Digital Single Market that include short and long-
term key policy actions in various areas: patents, trademarks, geographical indications, multi-
territorial copyright licensing, digital libraries, IPR violations, and IPR enforcement by customs
7
As a premise for a cultural policy and from a structured stakeholder dialogue, industry-led solutions
were put forward by stakeholders as a contribution to improve the availability of copyright-protected
content online in the EU. Available at http://eur-lex.europa.eu/legal-content/EN/ALL/?uri=CELEX:
52012DC0789
8
COM(2016), CELEX:52016PC0594
9
COM(2016)593, CELEX:52016PC0593
412 V. Rodriguez-Doncel et al.

The complexity of the regulatory system in this field, together with the variety of
the corpus of copyright (patchwork of international and European sources, such as the
Berne Convention for the Protection of Literary and Artistic Works, the WIPO
Copyright Treaty, the Directive 2001/29/EC10 (Copyright Directive), amongst other
correlated sources11), poses difficulties to search, retrieve and understand the legal
information in this domain. Moreover, in a pluralistic legal order [10] the “[EU]
legislation is drafted in several languages and […] the different language versions are
all equally authentic. An interpretation of a provision of [EU] law thus involves a
comparison of the different language versions”12, in accordance with the principle of
linguistic equality13, which entails a “full multilingualism” [11]. Settled case-law refers
that “the need for a uniform interpretation of [EU] regulations makes it impossible for
the text of a provision to be considered in isolation but requires, on the contrary, that it
should be interpreted and applied in the light of the versions existing in the other
official languages […] [A]ll the language versions must, (…) be recognised as having
the same weight”.14
However, due to the factors that act as constraints in particular judgments, “limited
multilingualism” seems a more realistic approach [18]. Besides, identifiable hindrances
prevent cross-border access to legal information:
– Disclosure of open data makes difficult to retrieve relevant and useful information
due to its overload and oversupply (large assortments of data);
– Legal documents are published as plain text without hyperlinks to the official legal
resources, averting navigation and reasoning among documents; national and EU
websites are sometimes poorly interconnected or they use different identification
systems;
– Data is not always published in machine readable formats like XML or RDF for
Linked Open Data, but in heterogeneous, non standard formats;
– Ambiguity and polysemy of legal terms [6]: the terminological misalignment and
the conceptual misalignment [9] between the terminology used at the EU level from
that of the national level, even when implementing EU directives [7];

10
The purpose of Directive 2001/29/EC of the European Parliament and of the Council of 22 May
2001 on the harmonisation of certain aspects of copyright and related rights in the information
society (Copyright Directive 83), is to implement theWIPO Copyright Treaty and to harmonise
aspects of copyright law across Europe, such as copyright exceptions
11
Connected legal instruments: the Directive 2009/24/EC of the European Parliament and of the
Council of 23 April 2009 on the legal protection of computer programs, the Directive 96/9/EC of
the European Parliament and of the Council of 11 March 1996 on the legal protection of databases,
the WTO’s Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS)
12
Case 283/81 CILFIT e.a. [1982] ECR 3415, paragraph 18
13
See EEC Council: Regulation No 1 determining the languages to be used by the European
Economic Community, [1958] OJ L 17/385
14
See Case C-257/00 Givane and Others [2003] ECR I-345, para. 36 and C-152/01 Kyocera [2003]
ECR I – 13833, para. 32
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 413

– Context-specificity of legal terms: the meaning of terms is related to the context of

the legislation defining it (several context-specific definitions of legal terms with a
common thread)15;
– Cultural-specificity of legal terms: the meaning of terms is related to the context of
the legal and political culture to which these terms belong to (think not only of the
classical distinction between Common and Civil law countries, but of cultures with
ideographic languages such as Chinese and Japanese) [8].
The need for cross-border multilingual access to legislation is required for legal
practitioners, such as judges, lawyers, translators, legal drafters and scholars, but also to
other decision-makers, amongst enterprises, public administrations and citizens, subject
to regulatory compliance (even outside their own area of expertise and also jurisdic-
tion), in order to: (i) exploit legal (open) data and therefore produce new innovative
services for the legal information provision market; (ii) to predict the impact of
implementing the EU legislature in each member state by enriching [7] structurally the
documents (with navigable references along legal texts) and semantically (with con-
cepts from ontologies and annotations); (iii) enhance information retrieval, automatic
translation, automated reasoning; (iv) ensure the principle of legal certainty; (v) pos-
sibly strengthening the textual (or literal interpretation), and teleological interpretation
upon which the European Court of Justice (ECJ)’s reasoning primarily rests.
Our work integrates the Linked Open Legal Data [12] momentum that illustrates
“the accessibility and semantic interoperability of legal sources”16. Some of the
advantages for rendering multi-lingual, multi-jurisdictional legal term banks published
as linked data are:
– Clear separation and identification of concepts and terms, as data fits a formalized
model and every resource is identifiable in a permanent manner.
– Easy browsing from a term in one language to an equivalent term in another
language, although this makes only full sense when a preferred term is specified17.
– Easy browsing among general terms and the jurisdiction-specific terms, as concepts
can be hierarchically organized. This clarity helps towards the harmonization of
copyright terms in the EU, an explicit goal in the EU copyright roadmap18.
– Easy comparative analysis, as multiple sources are provided.
– Improved discovery and unequivocal identification of concepts and corresponding
terms at both European and national levels.
– Better organization of conceptual domain knowledge and its availability of inter-
related data sets on the Web in standard formats.

15
This point is illustrated by the ruling of the ECJ Case 283/81 CILFIT e.a. [1982] ECR 3415,
paragraph 19
16
European Council, Draft Strategy on European e-justice 2014–2018, 2013 (2013/C 376/06)
17
For example, IATE defines preferred term as: “a term which should be used instead of any other
(equally correct) synonym(s) present, for harmonisation purposes”.
18
For a explicit mention, see the “Draft Report on the implementation of Directive 2001/29/EC of the
European Parliament and of the Council of 22 May 2001 on the harmonisation of certain aspects of
copyright and related rights in the information society (2014/2256(INI))”
414 V. Rodriguez-Doncel et al.

3 Linked Resources and Methodology

3.1 Publication Format
In order to build the present linked term bank, several resources have been considered.
(a) WIPO (World Intellectual Property Organization) publishes documents which
include glossaries on copyright-related terms19. WIPO is an especially authori-
tative source as the custodian of the treaties on copyright signed by almost every
country. Given their almost-universal validity, the definitions provided by WIPO
are attributed thus to the most general concepts.
(b) IATE (Inter-Active Terminology for Europe) is the current EU’s inter-
institutional terminology resource database, created from several preexisting
databases like EURODICAUTOM (Commission), TIS (Council) and EUTERPE
(Parliament), among others. IATE is managed by representatives from different
institutions including the authoritative entities like the ECJ or the Translation
Centre for the Bodies of the European Union. IATE contains more than 8 million
terms in all official 24 EU languages. It has been recently published as a linked
data resource [4].
(c) Creative commons licenses are text documents published along with the referred
work, and usually symbolized by icons, summaries or hyperlink references.
Creative Commons licenses, massively adopted by the internet culture, have been
published in versions tweaked for up to 60 different jurisdictions and different
languages. These licenses commence with the definition of the key terms, which
typically address the ones used in the target jurisdiction. From version 4.020,
Creative Commons aimed at a neutral text, capable of fitting every legal system.
Consequently, these definitions have been added to the general concepts and not
to the jurisdiction-specific concepts.
(d) Other resources. Finally, the term bank can be linked to other linked resources to
make it a highly connected linked data resource. In particular, the term bank of
copyright-related resources has been linked to DBpedia, the linked data version of
Wikipedia, and Lexvo.org21, a dataset of entities about language. The use of these
resources was possible as they had been published under open licensing
modalities.
The methodology followed to create the term bank has been the following:
1. Collection of top concepts. Key copyright-related concepts have been extracted
along with a general definition from the WIPO glossaries.
2. Mapping to IATE. The linked data version of IATE version was systematically
queried in search of direct matches. From the different sources of IATE, the legal

19
http://www.wipo.int/edocs/pubdocs/en/copyright/891/wipo_pub_891.pdf
20
“version 4.0 international license […] is the most up-to-date version of our licenses, drafted after
broad consultation with our global network of affiliates, and it has been written to be
internationally valid.”, http://creativecommons.org/version4
21
http://www.lexvo.org/
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 415

one was preferred over others when more than one term matched. The resulting
links were verified and completed manually by inspecting the official IATE place22.
3. Addition Creative Commons terms. Over 100 creative commons terms have been
defined, including the different versions, different jurisdiction ports and different
languages. These resources are well classified in the RDFLicense dataset23 [3],
which also provides the links between license identifiers and legal texts. Creative
Commons issued versions of the same license adapted to different jurisdictions
before their version 4.0. Definitions from version 4.0 were added to the general
concepts. The publication style of Creative Common licenses favors its automatic
parsing and the formatting codes can be easily removed.
The publication of the dataset was made according to the linked data publication
guidelines24 and those specific for term bases [4].

3.2 The Copyright Term Bank

The information in term banks is usually arranged as depicted in Fig. 1, following the
principle of strict separation between abstract concepts and the terms referring to them.
More than one term is possible for the same concept, even in the same language.

Concept

Language1 Language2

Term1 Term2 Term3

Fig. 1. Concepts and terms in terminology databases.

Concepts are typically accompanied by definitions, whereas terms are sometimes

provided with additional information like the source, reliability, domain, additional
notes, comments and the context of use.
In order to build our term bank, the structure has been extended to tackle the multi-
jurisdiction information that is provided, and jurisdiction-specific concepts have been
arranged as subspecies of general concepts (Fig. 2).

22
http://iate.europa.eu/
23
http://rdflicense.appspot.com/
24
http://www.w3.org/TR/ld-bp/
416 V. Rodriguez-Doncel et al.

General
Concept
Language2
JurisdicƟon-specific
concept Language1

Term5 Term6
Language1 Language2
Term4

Term1 Term2 Term3

Fig. 2. Concepts in a legal terminology database.

3.3 Publication Format

TermBaseeXchange (TBX) is the industry XML standard language used to represent
terminology data, sometimes used as native format, sometimes as interchange format. It
is published by ISO as standard ISO30042 and by the Localization Industry Standards
Association (LISA). The following excerpt shows an example of the published ter-
minology in TBX:

<termEntry id="Derivativework (ES)>

<langSetxml:lang="es">
<tig>
<term>obra derivada</term>
<termNote type="termType">fullForm</termNote>
<descrip type="reliabilityCode">3</descrip>
</tig>
</langSet>
<langSetxml:lang="ca">
<tig>
<term>obra derivada</term>
<termNote type="termType">fullForm</termNote>
<descrip type="reliabilityCode">3</descrip>
</tig>
</langSet>
</termEntry>

Code excerpt 1. Fragment of a term bank in the industry-standard TBX format (ISO30042)

For an advanced format where linking to other resources is made more straight-
forward, the RDF data structure as in [4] has been chosen. In order to represent the
linguistic information, we have adopted the ontolex-lemon model [2, 3], whose rep-
resentative schema is shown in Fig. 3. OntoLex is based on the ISO Lexical Markup
Framework (LMF) and is an extension of the lemon model (LExicon Model for
ONtologies). The specification of ontolex-lemon is currently under finalization by the
W3C Ontolex Community Group25 but it is already a de facto standard to represent
ontology lexica.

25
https://www.w3.org/community/ontolex/wiki/Final_Model_Specification
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 417

Fig. 3. The Ontolex-lemon model. Boxes denote OWL classes. The upper part of the boxes
contains the class name and the lower part contains the name of datatype properties. Black arrows
denote object properties, the white arrow denotes derivation and the symbol ↳ denotes a
subproperty relationship.

The example in the excerpt that follows shows two concepts: the universal concept
of “derivative work” (lines 7–13) and the concept of “derivative work” in particular in
the Spanish jurisdiction (lines 14–20). “Derivative work” is a general concept (skos:
Concept) that can be linked to the corresponding IATE concept (74645) and even to
a DBpedia resource (“Derivative_work”). “Derivative work (ES)” is an abstract con-
cept enshrined in 5 terms in 5 languages (Spanish, Catalan, Galician, Basque, Aranese)
for which the creative commons licenses have a translation of the Spanish port. One of
these terms is shown in lines 21–25, “obra derivada” in Galician language. The con-
vention of using uppercase for denoting the country code of a jurisdiction has been
used, as well as using lowercase to denote the language code.
418 V. Rodriguez-Doncel et al.

01 @prefix rdfs:<http://www.w3.org/2000/01/rdf-schema#> .
02 @prefix skos:<http://www.w3.org/2004/02/skos/core#> .
03 @prefix tbx:<http://tbx2rdf.lider-project.eu/tbx#> .
04 @prefix ontolex:<http://www.w3.org/ns/lemon/ontolex#> .
05 @prefix dct:<http://purl.org/dc/terms/> .
06 @prefix ctb:< http://tbx2rdf.lider-project.eu/converter/resource/cc/> .

07 ctb:derivative_work
08 a skos:Concept;
09 rdfs:label "derivative work";
10 skos:definition "a new work that translates or transforms one or more
original copyrighted pre-existing works"@en;
11 dct:source "WIPO";
12 owl:sameAs <http://dbpedia.org/resource/Derivative_work> ;
13 skos:closeMatch <http://tbx2rdf.lider-project.eu/data/iate/IATE-74645> .

14 ctb:derivative_work_(ES)
15 a skos:Concept;
16 rdfs:label "derivative work (ES)";
17 cc:jurisdiction <http://dbpedia.org/resource/Spain> ;
18 skos:narrower ctb:derivative_work ;
19 ontolex:isDenotedBy ctb:obra_derivada_gl, ctb:lan_eratorri_eu, ctb:%C3%B2b
ra_derivada_oci, ctb:obras_derivadas_es, ctb:obra_derivada_ca ;
20 skos:definition "e. Consideraranse obras derivadas aquelas obras creadas a
partir da licenciada, como por exemplo: as traducións e adaptacións; as revisi
óns, actualizacións e anotacións; os compendios, resumos e extractos; os arranx
os musicais e, en xeral, calquera transformación dunha obra literaria, artístic
a ou científica. Para evitar a dúbida, se a obra consiste nunha composición mus
ical ou gravación de sons, a sincronización temporal da obra cunha imaxe en mov
emento (synching) será considerada como unha obra derivada para os efectos dest
a licenza."@gl .

21 ctb:obra_derivada_gl
22 a ontolex:LexicalEntry;
23 ontolex:lexicalForm [ontolex:writtenRep "obra derivada"@gl ] ;
24 dct:source <http://creativecommons.org/licenses/by/3.0/es/legalcode.gl>;
25 tbx:reliabilityCode "3" .

Code excerpt 2. RDF Turtle serialization of one general concept (Derivative work), its derived concept for
the Spanish legislation (Derivative_work (ES)) and one of its lexical entries ("obra derivada") in the galician
language. To improve the legibility, the chars '%20' in the namespaced URIs have been replaced by a
blankspace. Equivalently, parentheses have been introduced.

4 Qualified Translations

The above RDF representation based on lemon supports the modeling of copyright and
related rights from a multilingual perspective. In this way, translations among different
lexical representations of terms, expressed in different natural languages, can be
inferred by traversing the RDF graph through their ontolex:LexicalSense. For
instance one can translate “obra derivada” from Galician into Spanish by pivoting on
their common sense26 in the above example. However, the meta-operational relation-
ship between legal reference and coreference has to be worked out.

26
http://tbx2rdf.lider-project.eu/converter/resource/cc/derivative%20work%20%28ES%29
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 419

However, this method does not account for the specific type of linguistic translation
that is taken place (e.g. literal translation, cultural equivalence, etc.). There exist,
however, an extension of the lemon model, the so called lemon translation module27,
that reifies the translation relation and allows associating additional information to it,
such as type of translation, confidence degree, provenance, and even the directionality
of the relation [17]. This module has been integrated in the new Ontolex-lemon model
as part of the new vartrans module.
In the case of the copyright term bank, using such mechanism to represent trans-
lations allows distinguishing between term descriptions that are a literal translation one
from the other (for instance “obra derivada” in the previous example) from other
situations in which the translated description has been adapted to the cultural or
jurisdictional specificities of the target language or legal system. This might be ben-
eficial for future semantic-aware applications in the legal domain. For instance, when
legal terminology has to be compared across language, it can be done within the same
jurisdictional domain, thus being a literal translations acceptable, or across jurisdic-
tions, in which legal equivalents (rather than literal translations) have to be found.
The use of the vartrans module is exemplified in the following code excerpt. The
lexical entries have two senses, which related by means of the reference to a common
concept. The translation is reified and can be qualified as trdcat:directEquiv-
alent or similar.

01 @prefix vartrans: <http://www.w3.org/ns/lemon/vartrans#> .

02 @prefix trcat: <http://purl.org/net/translation-categories#>

03 ctb:derivative_work_(ES)
04 a skos:Concept;
05 rdfs:label "derivative work"@en ;
06 cc:jurisdiction <http://dbpedia.org/resource/Spain> .

07 ctb:lan_eratorri_eu
08 a ontolex:LexicalEntry;
09 ontolex:lexicalForm [ontolex:writtenRep “lan eratorri”@eu ] ;
10 ontolex:sense <http://example.org/sense_1> .

11 ctb:obra%20derivada_gl
12 a ontolex:LexicalEntry;
13 ontolex:lexicalForm [ontolex:writtenRep “obra derivada”@gl ];
14 ontolex:sense <http://example.org/sense_2> .

15 <http://example.org/sense_1> ontolex:reference ctb:derivative_work_(ES) .

16 <http://example.org/sense_2> ontolex:reference ctb:derivate_work_(ES) .
17 <http://example.org/sense_1-sense_2-trans> a vartrans:Translation ;
18 vartrans:relates <http://example.org/sense_1> ;
19 vartrans:relates <http://example.org/sense_2> ;
20 vartrans:category trcat:directEquivalent .

Code excerpt 3. Example of use of the ontolex-lemon vartrans module. Prefixes from Code excerpt 1 also
apply.

27
http://purl.org/net/translation
420 V. Rodriguez-Doncel et al.

5 Related Work

In the literature, different methods exist for approaching the multilingual complexity of
European law, for example controlled vocabularies, implemented in terminology
database (such as IATE run by all the main EU Institutions that we have resort to in our
work), thesauri (as EUROVOC), semantic lexicons or lightweight ontologies(as
WordNet, EuroWordNet and, in the legal domain, JurWordNet) that we evoke here.
EuroVoc Thesaurus28 is the most important multilingual, multidisciplinary standard-
ized thesaurus created by the EU, covering the activities of the EU. EuroVoc is
managed by the Publications Office, which moved forward to ontology-based the-
saurus management and semantic web technologies conformant to W3C recommen-
dations as well as latest trends in thesaurus standards. However, EuroVoc represents a
wide-coverage and faceted thesaurus built specifically for processing the documentary
information of the EU institutions: the legal terminology is quite poor and limited to the
legal fields belonging to the competence of EU.
The CELLAR repository provides semantic indexing, advanced search and data
retrieval for multilingual resources to the information system of the Publications Office
of the European Union information system. Resources and their Functional Require-
ments for Bibliographic Records (FRBR) embrace both the web of data perspective and
the library or “bibliographic” data perspective [16]. Its new ontology development
assumes that “the FRBR classes are collectors of resource metadata at their specific
taxonomy level”, thus, allowing a direct constant access to the FRBR levels [16, p. 35].
This represents certainly an improvement over the existing model, as it enhances the
accessibility of the OP multilingual documents. However, its scope is also limited to
the vocabulary of EU documents.
The Legal Taxonomy Syllabus [6] is a multilevel, multilingual ontology that takes a
comparative law perspective to the modeling of legal terms and concepts from EU
Directives, helping to increase European terminological consistency. Syllabus is an
open-access database linking European terms with national transposition law and also
linking terms horizontally (i.e., between national legal orders).
LexALP [14] uses a technique defined for general lexical databases to achieve cross
language interoperability between languages of the Alpine Convention. This multi-
lingual legal information system combines three main components, (i) a terminology
data base, (ii) a multilingual corpus, and (iii) the relative bibliographic database. In this
way the manually revised, elaborated and validated (harmonised) quadrilingual infor-
mation on the legal terminology (i.e. complete terminological entries) will be closely
interacting with a facility to dynamically search for additional contexts in a relevant set
of legal texts in all languages and for all main legal systems involved.
The multilingual lexical database version of WordNet, EuroWordNet [13], com-
pounds wordnets expressing lexica of 8 European languages. The wordnets are
structured in terms of synsets (sets of synonymous words). Each synset in the

28
http://eurovoc.europa.eu/drupal/
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 421

monolingual WordNets is linked to the others by cross-lingual equivalence relations to

the English synsets recorded by the Inter-Lingual-Index (ILI). The database can be used
for monolingual and cross-lingual information retrieval. The LOIS [13] database is
compatible with the EuroWordNet architecture, and forms an extension of the EWN
semantic coverage into the legal domain. Within this framework, LOIS contributes to
the creation of a European Legal WordNet.

6 Conclusions and Future Work

We have perceived a particular European policy deference towards rendering the

copyright and related rights domain more accurate. We have framed some of the
advantages for yielding a multi-lingual, multi-jurisdictional legal term bank published
as linked data in this domain. Therefore our work presents an effort to achieve a
technical and semantic interoperability among linguistic domain concepts.
However, creating a term bank of legal terms is a time-consuming task where
expertise in the law of different countries is needed and even domain-specific termi-
nologies require a considerable effort. Legal terminologies, legal concepts and legal
knowledge are not synonymous.
Several problems might be raised: (i) ISO standards secure the exchange of ter-
minologies but do not manage the legal value of such terminologies, (ii) as said,
Version 4.0 of Creative Commons aims at a neutral text, capable of fitting every legal
system, but nothing prevents legal operators (e.g. judges) to offer different interpreta-
tions of general concepts at the jurisdiction-specific level; (iii) the term “bank” related
to other linked resources such as DBpedia or Lexvo.org entails a more careful exam-
ination of this kind of relationship, as the valence (i.e. the number of edges incident to
the vertex) of resources might not be equivalent, and in fact they are not.
Defining owl:sameAs relationships requires a more careful examination of the
functional entrenchment of legal sources [15], a more extended comparative work of
both legal and cultural systems [16] and and a closer attention to limited multi-
lingualism in EU and national Courts [18]. This work will be completed in the next
future with the annotations of a complete dataset of existing licenses.
By taking advantage of the fact that a lemon-based linguistic representation is
available, other possibilities beyond direct matching (owl:sameAs) can be explored
when mapping to IATE in order to make correspondences effective at the linguistic
level [19], and more sophisticated automatic mapping processes implemented [20].

Acknowledgements. This work is supported by the EU FP7 LIDER project (FP7 – 610782), by
DER2012-39492-C02-01 CROWDSOURCING, by Ministerio de Economía y Competitividad
(Juan de la Cierva Incorpora) and by the fellowship 520250-1-2011-1-IT-ERASMUNDUS
EMJD.
422 V. Rodriguez-Doncel et al.

References
1. ISO 30042:2008. Systems to manage terminology, knowledge and content – TermBase
eXchange (2008)
2. McCrae, J., et al.: Interchanging lexical resources on the semantic web. Lang. Resour. Eval.
46, 701–719 (2012)
3. Rodriguez-Doncel, V., Villata, S., Gomez-Perez, A.: A dataset of RDF licenses. In:
Hoekstra, R. (ed.) Proceedings of the 27th International Conference on Legal Knowledge
and Information System, p. 189 (2014)
4. Cimiano, P., McCrae, J., Rodriguez-Doncel, V., Gornostay, A., Gomez-Perez, A., Simoneit,
B.: Linked terminology: applying linked data principles to terminological resources. In:
Kozem, S. et al. (eds.) Proceedings of the 4th Biennial Conference on Electronic
Lexicography, pp. 504–517 (2014)
5. Cabrio, E., Palmero Aprosio, A., Villata, S.: These are your rights. In: Presutti, V., d’Amato,
C., Gandon, F., d’Aquin, M., Staab, S., Tordai, A. (eds.) ESWC 2014. LNCS, vol. 8465,
pp. 255–269. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-07443-6_18
6. Ajani, G., Boella, G., Lesmo, L., Mazzei, A., Rossi, P.: Terminological and ontological
analysis of European directives: multilinguism in law. In: 11th International Conference on
Artificial Intelligence and Law (ICAIL), pp. 43–48 (2007)
7. Ajani, G., Boella, G., Martin, M., Mazzei, A., Radicioni D., Rossi, P.: Legal taxonomy
syllabus 2.0. In: 3rd Workshop on Legal Ontologies and Artificial Intelligence Techniques
Joint with 2nd Workshop on Semantic Processing of Legal Texts (2009)
8. Nakamura, M., Ogawa, Y., Toyama, K.: Extraction of legal definitions and their
explanations with accessible citations. In: Casanovas, P., Pagallo, U., Palmirani, M., Sartor,
G. (eds.) AICOL -2013. LNCS (LNAI), vol. 8929, pp. 157–171. Springer, Heidelberg
(2014). https://doi.org/10.1007/978-3-662-45960-7_12
9. Lesmo, L., Boella, G., Mazzei, A.: Multilingual conceptual dictionaries based on ontologies:
analytical tools and case studies. In: Proceedings of V Legislative XML Workshop, pp. 1–14
(2006)
10. Maduro, P.P.: Interpreting European law: judicial adjudication in a context of constitutional
pluralism. Eur. J. Legal Stud. 1, 137 (2007)
11. Hanf, D., Muir, E.: Le droit de l’Union européenne et le multilinguisme. In: Hanf, D.E.,
Malacek, M.K. (eds.) Langue et Construction Européenne (2010)
12. Casellas, N.: Linked legal data: a SKOS vocabulary for the code of federal regulations. http://
www.semantic-web-journal.net/system/files/swj311_2.pdf (2012)
13. Peters, W., Sagri, M.T., Tiscornia, D.: The structuring of legal knowledge in LOIS. Artif.
Intell. Law 15, 117–135 (2007)
14. Lyding, V. et al.: The LexALP information system: term bank and corpus for multilingual
legal terminology consolidated. In: Proceedings of the Workshop on Multilingual Language
Resources and Interoperability, pp. 25–31 (2006)
15. Casanovas, P., Casellas, N., Vallbé, J.J.: Empirically grounded developments of legal
ontologies: a socio-legal perspective. In: Casanovas, P., et al. (eds.) Approaches to Legal
Ontologies, pp. 49–67. Springer, Dordrecht (2011). https://doi.org/10.1007/978-94-007-
0120-5_3
16. Francesconi, E., Küster, Marc W., Gratz, P., Thelen, S.: The ontology-based approach of the
publications office of the eu for document accessibility and open data services. In: Kő, A.,
Francesconi, E. (eds.) EGOVIS 2015. LNCS, vol. 9265, pp. 29–39. Springer, Cham (2015).
https://doi.org/10.1007/978-3-319-22389-6_3
 A Linked Data Terminology for Copyright Based on Ontolex-Lemon 423

17. Gracia, J., Montiel-Ponsoda, E., Vila-Suero, D., Aguado-de Cea, G.: Enabling language
resources to expose translations as linked data on the web. In: Proceedings of 9th Language
Resources and Evaluation Conference. European Language Resources Association, pp. 409–
413 (2014)
18. Derlén, M.: Multilingual Interpretation of European Union Law. Kluwer Law International,
Dordrecht (2009)
19. Gracia, J., Montiel-Ponsoda, E., Cimiano, P., Gómez-Pérez, A., Buitelaar, P., McCrae, J.:
Challenges for the multilingual web of data. J. Web Semant. 11, 63–71 (2012)
20. Euzenat, J., Shvaiko, P.: Ontology Matching. Springer, Berlin (2007). https://doi.org/10.
1007/978-3-540-49612-0
Legal Argumentation
 Anything You Say May Be Used Against
You in a Court of Law
Abstract Agent Argumentation (Triple-A)

Ryuta Arisaka1(B) , Ken Satoh2 , and Leendert van der Torre3

1
University of Perugia, Perugia, Italy
ryutaarisaka@gmail.com
2
National Institute of Informatics, Tokyo, Japan
3
University of Luxembourg, Luxembourg City, Luxembourg

Abstract. Triple-A is an abstract argumentation model, distinguish-

ing the global argumentation of judges from the local argumentation of
accused, prosecutors, witnesses, lawyers, and experts. In Triple-A, agents
have partial knowledge of the arguments and attacks of other agents,
and they decide autonomously whether to accept or reject their own
arguments, and whether to bring their arguments forward in court. The
arguments accepted by the judge are based on a game-theoretic equilib-
rium among the argumentation of the other agents. The Triple-A theory
can be used to distinguish various direct and indirect ways in which the
arguments of an agent can be used against his or her other arguments.

1 Introduction

There has been a long and fruitful interaction between formal argumentation
and legal reasoning. Legal reasoning has inspired the development of formal
argumentation techniques, and formal argumentation is frequently used as a
methodology in the legal domain. Whereas legal applications are typically based
on structured argumentation theories, following the work of Dung [16], also more
abstract argumentation theories are considered in the law [7,9,13,18,24]. For
example, formal argumentation has been applied to modelling doctrines in com-
mon law of contract [17], to US Trade Secrets Law [1], and to hypothetical,
evidential and abductive reasoning [8,10,11].
In this paper we introduce Triple-A, which stands for Abstract Agent Argu-
mentation. We are interested in argumentation of agents such as judges, accused,
prosecutors, witnesses, lawyers, and experts in court. In particular, in this paper
we address the following questions using Triple-A:

This work was partially supported by JSPS KAKENHI Grant Number 17H06103.
Moreover, the third author has received funding from the European Union’s Horizon
2020 research and innovation programme under the Marie Skodowska-Curie grant
agreement No 690974 for the project MIREL: MIning and REasoning with Legal
texts.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 427–442, 2018.
https://doi.org/10.1007/978-3-030-00178-0_29
428 R. Arisaka et al.

1. How is the argumentation of judges related to the argumentation of the other

agents?
2. How can the arguments an agent puts forward, be used against him?
To answer the first question, we use abstract Dung style semantics, but
with a twist. Whereas in Dung style argumentation, an agent presents a set
of accepted arguments, called an extension, in our approach an agent presents a
sub-framework of his or her argumentation framework. We can still recover the
extensions from these sub-frameworks, but the sub-frameworks contain more
information. This additional information can be exploited by the global argu-
mentation of the judge. In particular, since the argumentation of each agent may
depend on the argumentation of the other agents, we define the argumentation
of the judge as a game-theoretic equilibrium among the argumentation of the
other agents.
The second question refers to the title, which in turn refers to the so-called
Miranda warning, also referred to as the Miranda rights. This is a right to
silence warning given by police in the United States to criminal suspects in
police custody before they are interrogated to preserve the admissibility of their
statements against them in criminal proceedings. Every U.S. jurisdiction has its
own regulations regarding what, precisely, must be said to a person arrested or
placed in a custodial situation. The typical warning states:
– You have the right to remain silent and refuse to answer questions.
– Anything you say may be used against you in a court of law.
– You have the right to consult an attorney before speaking to the police and
to have an attorney present during questioning now or in the future.
– Etc.
For formal argumentation, this raises the question how an agent’s arguments
can be used against him or her. In this paper, we use the Triple-A theory to
study different ways in which this can happen. An argument of an agent can be
used directly against him, for example if he contradicts himself. However, more
interestingly, the arguments of an agent can be used against him indirectly, via
the argumentation of other agents. We associate with every agent the set of
arguments and attacks of other agents he is aware of. Moreover, we assume that
the agent takes these arguments and attacks into account when deciding which
arguments to accept.
This paper combines former research of the authors, in particular the work
on coalitional argumentation of the first two authors [3] and the work on multi-
sorted and input/output argumentation of the third author [5,28].
The layout of this paper is as follows. In Sect. 2 we give an overview of abstract
argumentation, including the sub-framework variant of Dung’s abstract argu-
mentation semantics we need for agents taking the arguments of other agents
into account. In Sect. 3 we introduce Triple-A and we explain how the argumen-
tation of judges is related to the argumentation of the other agents. In Sect. 4
we discuss how arguments of an agent can be used against his or her other
arguments.
 Anything You Say May Be Used Against You in a Court of Law 429

2 Sub-framework Semantics
A sub-framework semantics called AFRA semantics was introduced by Baroni
et al. [6]. In this section we introduce our version of sub-framework semantics.
We first recall Dung’s extension-based semantics. In this paper we consider
only stable semantics.
Definition 1 (Stable semantics). Let F = (A , R) be a graph called an
argumentation framework, where the elements of A are called arguments and
the binary relation R is called attack. A set of arguments B ⊆ A is a stable
extension if and only if it does not contain two arguments a1 and a2 such that
a1 attacks a2 , and for all arguments a2 not in B, there is an argument a1 in
B such that a1 attacks a2 . We write stb(F ) for the set of all stable extensions
of F .
We now consider sub-framework semantics. Close-minded stable semantics
returns exactly the stable extensions, and no attacks. Attack-minded stable
semantics returns all arguments, and all attacks except the ones on arguments
in the stable extension. This semantics was introduced as attack semantics by
Villata et al. [31]. Open-minded semantics returns the attack-minded seman-
tics, as well as attacks from arguments on arguments in the grounded extension.
The grounded extension is defined as usual as the fixed point of the charac-
teristic function, that returns all arguments defended by a set of arguments.
Max-minded semantics returns all sub-frameworks that have exactly one of the
stable extensions of the framework, and that are maximal in the sense that its
super-frameworks do not return this stable extension. Also other sub-framework
semantics can be defined, but we do not consider them in this paper.
Definition 2 (X-minded stable semantics). Let f be the defence (or char-
acteristic) function f (S) = {a ∈ A | ∀b ∈ A : R(b, a)∃c ∈ S : R(c, b)}, and f ∗
be the fixed point f (. . . (f (∅))). Close-minded stable semantics C, attack-minded
stable semantics A, open-minded stable semantics O, and max-minded stable
semantics M are given by
– C(A , R) = {(E, ∅) | E ∈ stb((A , R))},
– A(A , R) = {(A , R \ (A × E)) | E ∈ stb((A , R))},
– O(A , R) = {(A , R \ (A × (E \ f ∗ ))) | E ∈ stb((A , R))},
– M (A , R) = {(A , R  ) | stb((A , R  )) = {E} ⊆ stb((A , R)), if R ⊇ R  ⊃ R 
then stb((A , R  )) = {E}}.
Sub-framework semantics can be used as a model of dynamic argumenta-
tion, because open-minded agents may on the one hand accept an argument a
when also the arguments of other agents are considered, while on the other hand
rejecting this argument a when only their own arguments are considered. For
example, consider an argument framework F where argument b attacks argu-
ment a. The open-minded sub-framework is F and expresses the extension b.
However, if later argument b is removed, then argument a is accepted. The close-
minded single argument framework contains only argument b. Once b is removed,
no argument is accepted at all.
430 R. Arisaka et al.

Attack-minded and open-minded stable semantics satisfy the property that

the stable extensions coincide with the stable extensions of the sub-frameworks.
Close-minded and max-minded stable semantics satisfy the same property
trivially.

Proposition 1 (Relation). ∪{stb(F  ) | F  ∈ X(F )} = stb(F ) for X ∈

{C, A, O, M }.

Proof. Vacuous for X ∈ {C, M }. Consider X = A, we note that every member

(A , R\(A × E)) of A(A , R) removes from (A , R) all attacks on members of
E. Hence it is not possible that stb((A , R\(A × E))) does not contain members
of E. For X = O, note that every member (A , R\(A × (E\f ∗ ))) of A(A , R)
removes from (A , R) all attacks on members of E except when the members of
E belong to the grounded extension. Let us denote the set of all such members of
E by Ag . However, as the arguments attacking any member of Ag cannot be in a
member of stb((A , R)) by the definition of the grounded extension, and, further,
that all attacks on those attacking arguments will be preserved in (A , R\(A ×
(E\f ∗ ))), we see that Ag continue to be in stb((A , R\(A × (E\f ∗ )))), which,
moreover, is {E}.

Sub-framework stable semantics is illustrated in Example 1 below.

Example 1 (X-minded stable semantics). In Fig. 1, we find an argumentation
framework F1 , as shown in (A). It consists of three arguments a1 , a2 , a3 , and
the following attacks: (a1 , a2 ), (a2 , a3 ), (a3 , a2 ), (a3 , a1 ). We have stb(F1 ) = {a3 }.
Hence C(F1 ) = {(a3 , ∅)}. To obtain A(F1 ), we compute R\(A × E). We have
that A × E = {(a1 , a3 ), (a2 , a3 ), (a3 , a3 )}. Therefore the result of computation is
{(a1 , a2 ), (a3 , a2 ), (a3 , a1 )}. Consequently, A(F1 ) has the argumentation frame-
work shown in Fig. 1 as its sole member. In this example, E\f ∗ = E\{} = E,
and O(F1 ) = A(F1 ). Finally, there being only one stable extension {a3 } for
F1 , we have M (F1 ) = F1 . Any sub-framework of F1 having {a3 } as its stable
extension cannot belong to M because F1 as a larger sub-framework has the
same extension.
To show that A = O, consider F2 in (B) of Fig. 1. Here f ∗ (F2 ), i.e. the
grounded extension of F2 , is not empty and a2 (not in stb(F2 )) attacks a3 (in
f ∗ (F2 )), thus (a2 , a3 ) remains in O(F2 ). A(F2 ) can be compared to it.

Proposition 2. We write (A , R) ⊆ (A  , R  ) for (A ⊆ A  ) and (R ⊆ R  ).

Let : X × X, X ∈ {C, A, O, M }, be a predicate such that X1 X2 iff, for each
(finite) argumentation framework F and for each F2 ∈ X2 (F ), there exists
some F1 ∈ X1 (F ) such that F1 ⊆ F2 and that stb(F2 ) = stb(F1 ).1 Then
C A O M.

1
The other direction (i.e. for each F1 ∈ X1 (F )) is redundant by Proposition 1.
 Anything You Say May Be Used Against You in a Court of Law 431

Fig. 1. (A) An argumentation framework F1 with the only one member of: C(F1 )
(close-minded semantics); A(F1 ) (attack-minded semantics); O(F1 ); and M (F1 )
(max-minded semantics). (B) An argumentation framework F2 with the only member
of: C(F2 ); A(F2 ); O(F2 ); and M (F2 ).

Proof.
C A: Immediate from E ⊆ A .
A O: Follows from R\(A × E) ⊆ R\(A × (E\f ∗ )) (also see in Example 1).
O M : It is vacuous from the definition of M that for each member (A , R  ) of
M (A , R) there does not exist any larger R  than R  such that (A , R  ) has
the same stable extension E as (A , R  ) for each such E. Hence it suffices to
show that O(A , R) contains an element (A , R\(A × (E\f ∗ ))) for each such
E, which, however, follows vacuously from the definition of O.
An argumentation framework with input consists of an argumentation framework
F = (A , R), a set of external input arguments I , an input extension EI of I
and an attack relation RI from I to A . Given an argumentation framework
with input, a local function returns a corresponding set of sub-frameworks of
F . This is a sub-framework version of the local functions defined by Baroni et
al. [5], returning a corresponding set of extensions of F . Since they define their
local acceptance functions for all Dung semantics, not only for stable semantics,
their definitions are more general than ours. Similar notions are defined also by
Liao [21]. We refer to these papers for further explanations and examples about
local functions.
Definition 3 (Local stable semantics). We write 2F for {F  | F  ⊆ F }.
For a set of arguments B ⊆ A , we write (A , R)B for (B, R ∩ (B × B)), and
for a set of frameworks S, we write SB for {FB | F ∈ S}.
An argumentation framework with input is a tuple (F , I , EI , RI ), includ-
ing an argumentation framework F = (A , R), a set of arguments I such that
I ∩ A = ∅, an input extension EI ⊆ I and a relation RI ⊆ I × A . A local
function assigns to any argumentation framework with input a (possibly empty)
set of sub-frameworks of F , i.e. f (F , I , EI , RI ) ⊆ 2F .
Local X-minded stable semantics is defined by fX ((A , R), I , EI , RI )) =
X(A ∪ EI , R ∪ (RI ∩ (EI × A )))A for X ∈ {C, A, O, M }.
432 R. Arisaka et al.

Example 2 (Continued). From a point of view, X-mindedness in local stable

semantics of F can be regarded as controlling the amount of information about
stb(F ) to be revealed. Let us consider the simple argumentation framework F2
in Fig. 1 for illustration. For the context, suppose there was a murder, and that
the scene of murder was Laboratory A. A man is being suspected. Another man
reasons that the suspect could not have been able to enter Laboratory A, as the
facility enforces a very strict access control (argument a3 ). He reflects that with
an ID card, the suspect could have been there (argument a2 ). At this point there
occurs a mutual conflict between a3 and a2 . However, he remembers that the
suspect had lost his ID card (argument a1 ), dismissing a2 thereupon.
As can be easily seen, all the X-minded local stable semantics return the same
Dung stable semantics: that a1 (the suspect had lost his ID card) and a3 (the
suspect could not enter Laboratory A) but not a2 are in the extension. However,
the man could be as liberal as giving away all the reasoning as conducted in his
mind or, on the other hand, could choose just to state concisely what he thinks
are acceptable.
X = C represents the latter, under which he merely reveals a1 and a3 , which
does not necessarily make clear the link between the suspect’s possession of an
ID card and access to Laboratory A. In this example, X ∈ {O, M } represents
the former, revealing his reasoning precisely. Sitting between the two attitudes is
X = A under which he nuances his reasoning, that, although the suspect could
have been in Laboratory A with an ID card, such a thought should be dismissed
because the suspect had lost his ID card and could not get in there.

In the rest, we will focus on C and O.

3 Argumentation of Judge vs Argumentation of Others

An AAA framework is an argumentation framework together with a set of agents

and an assignment of the arguments to the agents. We call the agents also the
sources of the arguments. Rienstra et al. [28] call it a multi-sorted argumentation
framework.

Definition 4 (Agents). An abstract agent argumentation (or AAA, or Triple-

A) framework is a tuple F = (A , R, Ag, Src) where A is a non-empty set (of
arguments), R ⊆ A ×A is a binary relation over arguments (expressing attack),
Ag is a set (of agents) and Src : A → Ag is a function mapping each argument
to the agent that put it forward (also known as its source). For agent A ∈ Ag,
we write AA for {a ∈ A | Src(a) = A} and (A , R)A for (AA , R ∩ (AA × AA )),
and for a set of frameworks S, we write SA for {FA | F ∈ S}.

To illustrate Triple-A, we introduce a legal running example, based on two wit-

nesses.

Example 3. The following scenario is inspired by an example from Okuno and

Takahashi [22]. Suppose there occurred a murder, and that there are two persons
 Anything You Say May Be Used Against You in a Court of Law 433

individually reasoning over the case. One of them, say Wit 1, thinks firmly
that the suspect killed the victim at Laboratory C at around 7 pm (argument
a2 ). All material evidence imply that only the suspect could have killed the
victim. However, there is one remaining counter-evidence: the suspect had lost
his ID card which is required to enter the laboratory (argument a1 ). Thus Wit
1 is unable to prove a2 beyond doubt. Meanwhile, the other witness Wit 2 has
heard from the suspect that he was at restaurant until 7:10 pm (argument a4 ).
Wit 2 wants to trust the suspect’s words. Just at a corner of Wit 2’s mind, a
doubt lingers: that the suspect could have been at the laboratory, because the
security system of the building was faulty and no ID card was needed on the
day (argument a3 ).
Let us suppose that the overall argumentation Fx is as shown in Fig. 2. There
Src(a1 ) = Src(a2 ) = Wit 1 and Src(a3 ) = Src(a4 ) = Wit 2. However, for both
Wit 1 and Wit 2, the conclusion that they would draw could be very different if
they communicated or otherwise. If Wit 1 learns of a3 , for Wit 1, there will be
no longer any doubt that a2 holds, and if Wit 2 learns of a2 , Wit 2 would reason
that a4 was a lie by the suspect after all. On the other hand, if Wit 1 (Wit 2) is
unaware of a3 (a2 ), Wit 1 (Wit 2) would remain inconclusive of his/her decision.

Fig. 2. An argumentation framework Fx with two agents: Wit 1 and Wit 2, who each
has two arguments.

Epistemic Triple-A extends Triple-A with the knowledge of the agents, which
reflects the arguments and attacks the agents are aware of. As two extreme cases,
we consider so-called aware agents that know all the arguments and attacks of
other agents, and so-called unaware agents that only know their own arguments
and the attacks among them.
Definition 5 (Aware and unaware agents). Epistemic Abstract Agent
Argumentation (or EAAA, or Epistemic Triple-A) framework F =
A , R, Ag, Src, K extends a Triple-A framework with the agent knowledge K, a
function KA (F ) ⊆ F that maps agent A ∈ Ag to a sub-framework of F , repre-
senting the arguments and attacks the agent knows. We assume FA ⊆ KA (F ),
representing that agents know their own arguments and the attacks among them.
An agent A is called aware if KA (F ) = F and unaware if KA (F ) = FA .
434 R. Arisaka et al.

Agent semantics is a local function with respect to the agent input, where
the knowledge of the agents is used to define the input of the agents.
Definition 6 (Agent stable semantics). Let F = A , R, Ag, Src, K be an
Epistemic Triple-A framework, A ∈ Ag be an agent and KA (F ) = (A  , R  ) be
agent A’s knowledge. The input of A is IA = {a1 ∈ A  | (a1 , a2 ) ∈ R  ∩ ((A \
AA ) × AA )}, and the conditioning relation of A is RIA = R  ∩ (IA × AA ),
Given a sub-framework stable semantics S ∈ {C, O}, agent stable semantics
is local function fS (F , IA , EIA , RIA ) = S(A ∪ EI , R ∪ (RI ∩ (EI × A )))A .
The running witnesses example illustrates the local agent argumentation
semantics of the witnesses.
Example 4 (Continued). We look at C and O for both aware and unaware agents.
Let us consider open-minded semantics to begin with ((A)-(B2) in Fig. 3). Sup-
pose Wit 1 is aware, then FA (Cf. (A) of Fig. 3) is considered for local agent
semantics. When argument a3 , the input argument, is in the input extension,
Wit 1’s local agent semantics, fO (FWit1 , {a3 }, {a3 }, {(a3 , a1 )}), must be the set
of sub-argumentation frameworks of FWit1 whose Dung stable extension is {a2 },
and which have all arguments that appear in FWit1 . F (1a) as shown in (A1)
of Fig. 3 is the only one possibility. Intuitive meaning of this semantics is that,
when Wit 2 presumes a3 acceptable, Wit 1 will presume a2 acceptable and a1
unacceptable as his/her response.
When a3 is not in the input extension, there are two Dung stable extensions
for FWit1 , i.e. stb(FWit1 ) = {{a1 }, {a2 }}. fO (FWit1 , {a3 }, ∅, {(a3 , a1 )}), Wit 1’s
local agent semantics, must therefore be the set of sub-argumentation frame-
works of FWit1 whose Dung extension is either {a1 } or {a2 }, and which have all
arguments that appear in FWit1 . We consequently obtain {F (1a), F (1b)} (Cf.
(B) of Fig. 3). In a similar manner, we obtain open-minded local agent semantics
for aware Wit 2 (Cf. (B)-(B2) of Fig. 3).
Now, suppose that Wit 1 is unaware. Then we consider FWit1
(Cf. (C) of Fig. 3). Unaware Wit 1’s open-minded local agent semantics
fO (FWit1 , {a3 }, ∅, {(a3 , a1 )}) is {F (1a), F (1b)}, as shown in (b) of Fig. 3. As
expected, this case is the same as for aware Wit 1 when a3 is not in the input
extension. In a similar manner, we obtain open-minded local agent semantics for
unaware Wit 2 ((D1) of Fig. 3).
For close-minded semantics ((A2), (B2), (C2), (D2) in Fig. 3), we obtain only
arguments for agents local semantics without attack arrows.
We now give the central definition of our multi-agent interaction. Since the
local semantics is expressed as sub-frameworks, but the input is expressed as an
extension, we represent the equilibria by pairs of an argumentation framework
and an extension.2
2
We can define multiagent semantics also as a set of arguments only, as follows: A
stable MAS extension E is a subset of A such that there exists a sub-framework
F  of F such that . . . . As suggested by Massiliano Giacomin, there seems to be no
loss of information using just extensions. We did not do so, because the resulting
definition seems to become more difficult to read.
 Anything You Say May Be Used Against You in a Court of Law 435

Fig. 3. Illustration of agent stable semantics for aware/unaware agents.

Definition 7 (Multi-agent semantics). Let F = A , R, Ag, Src, K be an

Epistemic Triple-A framework, and for each agent A ∈ Ag, let SA ∈ {C, O} be
an agent stable semantics for agent A. The semantics of a triple-A framework F
436 R. Arisaka et al.

Fig. 4. Multi-agent semantics of Wit 1’s and Wit 2’s local agent stable semantics.

is a pair of an argumentation framework F  = (A  , R  ) and a stable extension

E of F  such that we have

1. FA ∈ S(FA , IA , E ∩ IA , RIA )A , and

2. for all (a1 , a2 ) such that Src(a1 ) = Src(a2 ), we have (a1 , a2 ) ∈ R  iff
(a1 , a2 ) ∈ R ∩ (A  × A  ).

The running witnesses example illustrates the global argumentation seman-

tics of the judge.
Example 5 (Continued). Consider Fx (Cf. Fig. 2) once again. We saw open-
minded and close-minded local agent stable semantics for Wit 1 and Wit
2 in Example 4. Figure 4 lists all combinations of open/close-minded and
aware/unaware agents and their multi-agent semantics, except for (Wit 1:
 Anything You Say May Be Used Against You in a Court of Law 437

unaware, Wit 2: aware) case, as it is symmetric to (Wit 1: aware, Wit 2: unaware)

case.
Notice that, even when local agent Wit 1 in its local agent stable semantics
accepts a1 (F (1c)), a1 may not be accepted in multi-agent stable semantics (see
the combination F (1b)F (2a) when both Wit 1 and Wit 2 are unaware). As
such, the multi-agent semantics reflects judge’s viewpoint who must evaluate all
arguments provided by involved agents to deliver fair judgement.
For comparisons of open-minded and close-minded semantics, notice that we
could gain a set-theoretically smaller extension with close-minded agents than
with open-minded agents. As Fig. 4 shows, if, for unaware agents, open-minded
Wit 1’s F (1b) is combined with close-minded Wit 2’s F (2c), we obtain {a2 , a3 }
as the extension for the multi-agent semantics, whereas with close-minded Wit
1’s F (1d) and close-minded Wit 2’s F (2c), we obtain {a3 } as the extension,
even though Dung stable extensions for F (1b) and F (1d) are the same. The
rationale is that open-minded witnesses may provide more information for the
judge to evaluate.

4 Argument Against the Agent

In this section we illustrate ways in which agent’s arguments may be used against
him/her. We use a variation of the running example with an accused, a witness,
a prosecutor and finally a judge who globally evaluates these arguments.

Example 6. There occurred a murder at Facility C. Acc is being accused. There

is a witness Wit and a prosecutor Prc. Acc has in mind two arguments:
a1 that he was at Facility A on the day of the murder. (This is a fact known to
Acc)
a2 that he is innocent. (This is Acc’s claim)
Prc entertains:
a6 that only Acc could have killed the victim. (This is Prc’s claim)
Meanwhile, Wit believes certain information. He has three arguments:
a3 Acc stayed at home on the day of the murder, having previously lost his ID
card. (This Wit originally believes to be a fact)
a4 Acc could enter any facility provided he is with his ID card. (This is a fact
known to Wit)
a5 Acc could not have been at Facility C at the time of the murder. (This is
Wit’s claim)
Further, the relation between the three arguments is such that a3 attacks a4
which attacks a5 . Altogether, these arguments by the three agents form the
argumentation framework in Fig. 5.
Each of Acc, Prc and Wit reveals his/her internal argumentation, partially
or elaborately, for the judge to evaluate. But since an agent may come to learn
438 R. Arisaka et al.

Fig. 5. Left: Argumentation by an accused (marked Acc), by a witness (marked Wit),

and by a prosecutor (marked Prc). Right: multi-agent semantics with open-minded
unaware Wit (Fi ), with close-minded unaware Wit (Fj ), with open-minded aware
Wit (Fk ), and with close-minded aware Wit (Fl ).

arguments by other agents when, say, they are expressed before he/she pre-
sented his/her own, it is possible that he/she take the additional information
into account in delivering his/her decision.
In this example, both Prc and Acc have the trait of our unaware agents,
for Prc has no reason to drop his/her argument a6 ; neither does Acc, seeing no
benefit in admitting to a6 . However, how Wit responds to the fact known to Acc
(a1 ) can prove crucial for him to be judged innocent or guilty.

Case A. Suppose Wit is unaware and open-minded. Wit presents what he

believes, namely his local argumentation framework (see Fi in Fig. 5). She locally
accepts a3 and a5 . The judge evaluates all arguments, concluding that a2 is not
acceptable, i.e. Acc is guilty. The judge starts his inference with Acc’s acceptable
a1 , to reject a2 . The two arguments a3 and a5 accepted by Wit are not accepted
by the judge. This illustrates indirect use of an argument against Acc.
 Anything You Say May Be Used Against You in a Court of Law 439

Case B. Suppose Wit is unaware and close-minded. Wit, instead of presenting

all reasoning as he did to develop his local argumentation, states the key points
concisely: that Acc stayed at home on the day of the murder, and that Acc
could not have been at Facility C (see Fj ). The omission of the fact known to
Wit (a4 ), which Wit perhaps considers irrelevant to the criminal case, changes
judge’s decision completely. In Fj , a5 is an argument that is globally acceptable,
which rejects a6 in favour of a2 .

Case C. Suppose Wit learns a1 beforehand. Wit realises that a3 which she
thought was a fact is not factual. She no longer claims a5 in her local argumen-
tation, although she discloses her argumentation entirely (see Fk ). Her decision
that a4 is acceptable agrees with the judge’s, and the latter concludes that a2
shall be rejected.

Case D. Suppose again Wit learns a1 beforehand, but that she mentions the
key arguments concisely. She states that entry to any facility requires an ID card
(see Fl ). Here again, the judge has no objection to the evidence that might have
been provided by Proc. a6 is accepted, to prove Acc guilty. This illustrates a
direct use of an argument by Wit against Acc.
As we highlighted, agents attributes influence global judgement. Further, an
argument may be used against an agent directly or indirectly.

5 Related Work
Argumentation by autonomous agents have been studied mostly in the context
of strategic argumentation games [2,20,23,25,26,29]. Negotiation dialogues [2]
characterise changes in the set of arguments acceptable to an agent in accordance
to new arguments another agent introduces into his/her local scope, which, as
ours, respects agents locality. Compared to the existing studies, however, the
focus of our work is more on analysis of how derivation of local semantics by
local agents influences global arguments’ acceptance. As we observed, multi-
agent semantics for open-minded agents may not be the same as for close-minded
agents, even when they accept the same sets of arguments. We observed also
that locally accepted arguments may not belong to globally acceptable argu-
ments. Agents attributes: open-minded vs close-minded and aware vs unaware,
play a role in the interaction of local agent semantics. There are other agents
attributes [23]. While many studies on game-theoretic strategic argumentation
games have presupposed complete information (see [19]), realistic legal exam-
ples often involve uncertainty of the belief state of other agents’, and a theory
that adapts to incomplete information is highly relevant. In our framework, local
agents may be uncertain of acceptability statuses of the arguments that could be
attacking their arguments (the input arguments). As such, they consider their
best response to each possible scenario for dealing with the element of uncer-
tainty based on their agents attributes.
440 R. Arisaka et al.

Rahwan and Larson [27] contemplate (re)construction of an argumentation

framework from the arguments in a given argumentation framework that are
distributed across agents. In the construction process, the agents may or may
not reveal their arguments, and the global outcome to be obtained varies with
their decisions. Concerning this, we saw that even if all involved agents are open-
minded and revealing all arguments in their local space, the global outcome may
still differ.
Judgement aggregation [4,12,14,30] to determine acceptable arguments
based on social choice theory or aggregation of argumentation frameworks [15]
are being studied. While they are not the main focus of this paper, such studies
become important when we deal with agents perception of other agents’ local
argumentation. We aim to extend our theory for that kind of a situation in future
work.

6 Conclusion

Triple-A is an agent abstract argumentation model, distinguishing the global

argumentation of judges from the local argumentation of accused, prosecu-
tors, witnesses, lawyers, experts and so on. In Triple-A, the agents decide
autonomously whether to trust the other agents in the sense of taking some
of their arguments into account. Moreover, they decide autonomously whether
to accept or reject their own arguments, and whether to bring their arguments
forward in court. The globally accepted arguments by the judge are defined using
a game theoretic equilibrium definition.
Triple-A framework distinguishes various direct and indirect ways in which
the arguments of an agent can be used against his other arguments.

Acknowledgement. We thank Massimiliano Giacomin for insightful feedback on an

earlier version of this paper.

References
1. Al-abdulkarim, L., Atkinson, K., Bench-Capon, T.: ANGELIC secrets: bridging
from factors to facts in US trade secrets. In: JURIX, pp. 113–118 (2016)
2. Amgoud, L., Vesic, S.: A formal analyis of the role of argumentation in negotiation
dialogues. J. Log. Comput. 5, 957–978 (2012)
3. Arisaka, R., Satoh, K.: Coalition formability semantics with conflict-eliminable sets
of arguments (Extended Abstracts). In: AAMAS, pp. 1469–1471 (2017)
4. Awad, E., Booth, R., Tohmé, F., Rahwan, I.: Judgement aggregation in multi-agent
argumentation. J. Log. Comput. 27(1), 227–259 (2017)
5. Baroni, P., Boella, G., Cerutti, F., Giacomin, M., van der Torre, L.W.N., Villata,
S.: On the input/output behavior of argumentation frameworks. Artif. Intell. 217,
144–197 (2014)
6. Baroni, P., Cerutti, F., Giacomin, M., Guida, G.: AFRA: argumentation framework
with recursive attacks. Int. J. Approx. Reason. 52(1), 19–37 (2011)
 Anything You Say May Be Used Against You in a Court of Law 441

7. Bench-Capon, T.: Representation of case law as an argumentation framework. In:

JURIX, pp. 53–62 (2002)
8. Bench-Capon, T., Prakken, H.: Using argument schemes for hypothetical reasoning
in law. Artif. Intell. Law 18(2), 153–174 (2010)
9. Bench-Capon, T., Sartor, G.: A model of legal reasoning with cases incorporating
theories and values. Artif. Intell. 150, 97–143 (2003)
10. Bex, F., Bench-Capon, T., Atkinson, K.: Did he jump or was he pushed? Abductive
practical reasoning. Artif. Intell. Law 17(2), 79–99 (2009)
11. Bex, F., Prakken, H., Reed, C., Walton, D.: Towards a formal account of reason-
ing about evidence: argumentation schemes and generalisations. Artif. Intell. Law
11(2), 125–165 (2003)
12. Bodanza, G.A., Auday, M.R.: Social argument justification: some mechanisms and
conditions for their coincidence. In: Sossai, C., Chemello, G. (eds.) ECSQARU
2009. LNCS (LNAI), vol. 5590, pp. 95–106. Springer, Heidelberg (2009). https://
doi.org/10.1007/978-3-642-02906-6 10
13. Brewka, G., Strass, H., Ellmauthaer, S., Wallner, J., Woltran, S.: Abstract dialec-
tical frameworks revisited. In: IJCAI (2010)
14. Caminada, M., Pigozzi, G.: On judgement aggregation in abstract argumentation.
Auton. Agents Multi-Agent Syst. 22(1), 64–102 (2011)
15. Coste-Marquis, S., Devred, C., Konieczny, S.: On the merging of dung’s argumen-
tation systems. Artif. Intell. 171(10–15), 730–753 (2007)
16. Dung, P.M.: On the acceptability of arguments and its fundamental role in non-
monotonic reasoning, logic programming, and n-person games. Artif. Intell. 77(2),
321–357 (1995)
17. Dung, P.M., Thang, P.M.: Modular argumentation for modelling legal doctrines in
common law of contract. Artif. Intell. Law 17, 167–182 (2008)
18. Gordon, T., Prakken, H., Walton, D.: The carneades model of argument and burden
of proof. Artif. Intell. 171, 875–896 (2007)
19. Governatori, G., Olivieri, F., Rotolo, A., Scannapieco, S., Sartor, G.: Two faces of
strategic argumentation in the law. In: JURIX, pp. 81–90 (2014)
20. Grossi, D., van der Hoek, W.: Audience-based uncertainty in abstract argument
games. In: IJCAI, pp. 143–149 (2013)
21. Liao, B.: Toward incremental computation of argumentation semantics: a
decomposition-based approach. Ann. Math. Artif. Intell. 67(3–4), 319–358 (2013)
22. Okuno, K., Takahashi, K.: Argumentation system with changes of an agent’s knowl-
edge base. In: IJCAI vol. 09, pp. 226–232 (2009)
23. Parsons, S., Sklar, E.: How agents alter their beliefs after an argumentation-based
dialogue. In: Parsons, S., Maudet, N., Moraitis, P., Rahwan, I. (eds.) ArgMAS
2005. LNCS (LNAI), vol. 4049, pp. 297–312. Springer, Heidelberg (2006). https://
doi.org/10.1007/11794578 19
24. Prakken, H., Sartor, G.: A dialectical model of assessing conflicting arguments in
legal reasoning. Artif. Intell. Law 4, 331–368 (1996)
25. Procaccia, A., Rosenschein, J.: Extensive-form argumentation games. In: EUMAS,
pp. 312–322 (2005)
26. Rahwan, I., Larson, K.: Argumentation and game theory. In: Simari, G., Rahwan,
I. (eds.) Argumentation in Artificial Intelligence, pp. 321–339. Springer, Heidelberg
(2009). https://doi.org/10.1007/978-0-387-98197-0 16
27. Rahwan, I., Larson, K.: Mechanism design for abstract argumentation. In:
AAMAS, pp. 1031–1038 (2008)
442 R. Arisaka et al.

28. Rienstra, T., Perotti, A., Villata, S., Gabbay, D.M., van der Torre, L.: Multi-sorted
argumentation. In: Modgil, S., Oren, N., Toni, F. (eds.) TAFA 2011. LNCS (LNAI),
vol. 7132, pp. 215–231. Springer, Heidelberg (2012). https://doi.org/10.1007/978-
3-642-29184-5 14
29. Riveret, R., Prakken, H.: Heuristics in argumentation: a game theory investigation.
In: COMMA, pp. 324–335 (2008)
30. Tohmé, F.A., Bodanza, G.A., Simari, G.R.: Aggregation of attack relations: a
social-choice theoretical analysis of defeasibility criteria. In: Hartmann, S., Kern-
Isberner, G. (eds.) FoIKS 2008. LNCS, vol. 4932, pp. 8–23. Springer, Heidelberg
(2008). https://doi.org/10.1007/978-3-540-77684-0 4
31. Villata, S., Boella, G., van der Torre, L.: Attack semantics for abstract argumen-
tation. In: IJCAI 2011, Proceedings of the 22nd International Joint Conference on
Artificial Intelligence, Barcelona, Catalonia, Spain, 16–22 July 2011, pp. 406–413
(2011)
 A Machine Learning Approach
to Argument Mining in Legal Documents

Prakash Poudyal(B)

Department of Informatics, University of Évora, Évora, Portugal

prakashpoudyal@gmail.com

Abstract. This study aims to analyze and evaluate the natural lan-
guage arguments present in legal documents. The research is divided
into three modules or stages: an Argument Element Identifier Mod-
ule identifying argumentative and non-argumentative sentences in legal
texts; an Argument Builder Module handling clustering of argument’s
components; and an Argument Structurer Module distinguishing argu-
ment’s components (premises and conclusion). The corpus selected for
this research was the set of Case-Laws issued by the European Court
of Human Rights (ECHR) annotated by Mochales-Palau and Moens [8].
The preliminary results of the Argument Element Identifier Module are
presented, including its main features. The performance of two machine
learning algorithms (Support Vector Machine Algorithm and Random
Forest Algorithm) is also measured.

Keywords: Legal argument · Natural language analysis

Machine learning

1 Introduction

An argument combines a premise or a set of premises and a conclusion. Histori-

cally, Dialectics and Philosophy are the ancient roots of the discipline of argumen-
tation. Arguments have always been considered an important branch of Philoso-
phy and, with the passage of time and advancement in technology, its relevance
has grown exponentially in other fields such as Literature, Logic, Law, and also in
Mass Communication and Artificial Intelligence. Arguments are the fundamental
tools for human beings to argue and reach their objectives. During debates, the
conclusion of an argument is the focal point of the discussion. Premises are the
vehicle that supports the conclusion’s reasoning and approval. There are premises
that reinforce other premises and as such add strength to the conclusion. During
a discussion, facts, figures and further evidence as well as logic are provided to
support, attack and/or refute the opponent’s arguments. At a time when social
media is one of the most important discussion platforms available, the number
of users expressing their opinion has grown exponentially. Usually, such opinions
are expressed through an array of premises that generate ideas and claims. Con-
sidering the relevance of argumentation in everyday life and its ubiquity in the
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 443–450, 2018.
https://doi.org/10.1007/978-3-030-00178-0_30
444 P. Poudyal

judiciary, this study was made to analyse and evaluate the natural language used
in argumentative legal documents. To automatically identify the argument in an
unstructured text, a system was developed in three stages or modules. The first
stage or module is the Argument Element Identifier, henceforth referred to by its
acronym AEI. In this module, the main aim was to identify the argumentative
and non-argumentative sentences in a corpus of legal documents. The structur-
ing of arguments is addressed in the second stage or the Argument Builder Mod-
ule, henceforth referred to as AB. In the third stage, the Argument Structurer
Module (henceforth referred to as AS), the system will distinguish the arguments’
components (premise and conclusion). The corpus selected for this study was the
Case-Law issued by the European Court of Human Rights (ECHR) annotated by
Mochales-Palau and Moens [8]. Details of the corpus are described in [11].
Mochales-Palau and her colleagues [6–10,13] have published several papers
identifying and extracting arguments from both the ECHR Corpus and the Arau-
caria Corpus1 . Moens et al. [9] used features such as n-gram, verb nodes, word
couples, and punctuation and their average accuracy results was close to 74% in
various types of text but dropped slightly to 68% in the legal corpus. Mochales-
Palau and Moens [8] added more features such as modal auxiliary, keywords,
negative/positive words, text statistics, punctuation keywords, same words in
both the previous as well as the following sentence, and first and last words in
the next sentence and reported accuracy results of 90%. Mochales-Palau and
Moens [10] also defined the argument boundaries i.e. the beginning as well as
the end of an argument. Since components of the argument can be found scat-
tered throughout the text, the authors suggest using semantic distance to solve
this issue and argue for the use of context-free grammars (CFG) to detect the
argument structure and claim to have reached and accuracy of 60%. The tech-
nique presented by these authors is applied only to a very limited number of
Case-Laws.
Stab et al. [15,16] analysed argumentative writings from a discourse structure
perspective. They used structural, lexical, syntactic and contextual features to
determine argumentative discourse structures in persuasive essays. Their exper-
iment succeeded in establishing the f-measure for identifying argument com-
ponents at 0.726. They focused on word indicators and lexical features that
highlight an argumentative sentence. Doddington et al. [4] described four chal-
lenges and identified five types and 24 subtypes of relations. The “Role” type
of relation, which refers to the part a person plays in an organization, can be
subtyped as Manager, General Staff, Member, Owner, Founder, Client, Affiliate-
Partner, Citizen-of or Other. The “Part” type of the relation can be subtyped
as Subsidiary, Part-of or Other. The “Near” type identifies relative locations.
The “Social” type can be subtyped as Parent, Sibling, Spouse, Grandparent,
Other-Relative, Other-Personal, Associate, or Other-Professional.
Bunescu and Mooney [2] presented a novel approach to extract the relation
between entities by presenting a new kernel for the relation extraction, based on
the shortest path between the two relation entities in a dependency graph. They
1
http://araucaria.computing.dundee.ac.uk/doku.php.
 A Machine Learning Approach to Argument Mining in Legal Documents 445

Fig. 1. Proposed Architecture of the System

deployed an “Automatic Content Extraction” on a corpus of newspaper articles

and were able to show significant improvements over a recent dependency tree
kernel. Biran and Rambow [1] also aimed to identify argumentative relations
while Cabrio and Villata et al. [3] used a combination of textual entailment
framework and bipolar abstract argumentation approach to evaluate argument
texts and find the relation between the arguments. Florou et al. [5] used a gram-
matical approach of future and conditional tenses and moods. They highlight
the impact of illustration, justification, and rebuttal wording in the argument.
Poudyal and Quaresma [12] have found that the Support Vector Machine is the
best machine learning algorithm in identifying name entity relation.

2 Proposed Approach
The system we propose consists of three sequential modules or phases as illus-
trated by Fig. 1.
1. Argument Element Identifier (AEI): identifies argumentative and non - argu-
mentative sentences in legal texts;
2. Argument Builder (AB): handles arguments’ components’ clustering;
3. Argument Structurer (AS): distinguishes arguments components (premise
and conclusion).

During the Argument Element Identifier (AEI) phase, our main task was to
find an optimal machine learning algorithm with appropriate features to distin-
guish an argumentative from a non-argumentative sentence in legal documents.
We conducted several experiments with various machine learning algorithms
and classified them according to the type of features used. Figure 2 presents an
overview of the AEI phase. After identifying the argumentative sentences in a
legal text, it is necessary to organize these sentences into argumentative clus-
ters composed by a set of argumentative sentences interconnected or related to
each other. Detecting the boundaries of an argument is a very challenging task
mainly due to the fact that its components (premise and conclusion) may be
connected or related to other arguments. To cluster such sentences, we deployed
a fuzzy clustering algorithm (FCA) that provides a membership value ranging
from 0 to 1 for each sentence cluster. The membership values are the key assets
of the FCA, which allows us to associate each sentence to more than one argu-
ment cluster. The performance of the algorithm depends on the type of features
446 P. Poudyal

Fig. 2. Overview of the Argument Element Identifier Module (AEI)

selected. In this study, we focus on the following features: ‘N-gram’, ‘Word2vec’,

and ‘Sentence Position’. Figure 3 offers an overview of this phase. On the AS
phase, argument components (premise and conclusion) are identified as having a
premise or a conclusion basis. The sentences identified as having a premise basis
are outright premises or consist of a premise clause. The sentences identified as
having a conclusion basis are obvious conclusions or point towards one. Many
sentences that have a premise basis and are tagged as such may also include a
conclusion clause and the same happens to the sentences labeled as displaying a
conclusion basis. To accomplish this task, we deployed indicator features. Indica-
tor features play an important role in identifying argument’s premises and con-
clusions. Words such as “finally,” “therefore,” “concluding,” and “thus” clearly
introduce a conclusion and play an important role in the process of identify-
ing argument’s conclusions. It is also highly probable that sentences containing
words like “should,” “could,” “almost,” “must be,” “because,” “seems,” and
“would like,” are premises. A major limitation in the AS phase is that each
sentence may have one or several premises but only one conclusion, and also
the system’s accuracy rate will diminish whenever the classifier is not able to
identify the sentence’s conclusion, or identifies more than one conclusion in a
single argument.

3 AEI Preliminary Results

The main goal of the AEI phase was to select the algorithm with the most
appropriate parameters. We aimed to develop a system that will automatically
 A Machine Learning Approach to Argument Mining in Legal Documents 447

Fig. 3. Overview of the Argument Builder (AB) Module

identify the argumentative sentences on an unstructured textual document. As

Fig. 4 illustrates, the AEI system’s architecture follows several steps. Initially,
the corpus needs to be refined. Once the features are extracted, the classifier can
then be built and its performance evaluated.
The words that form a document must be mapped in accordance to a prede-
termined token and TF-IDF in order to normalise the length of each unit. In our
experiment, this procedure created 11374 features. The TF-IDF [11] function
was calculated as:
N
tf − idf (wi , d) = tf (wi , d)ln (1)
df (wi )

where tf (wi d) is the frequency word wi in document d and df (wi ) is the num-
ber of documents where wi appears and N is the number of documents in the
corpus. To measure performance we used precision, recall and f-measure [14]
methods. We ran several experiments with the machine learning algorithms
Support Vector Machine (SVM) and Random Forest (RF) to determine their
performance in identifying argumentative sentences in accordance with the fea-
tures provided. We selected the top-n informative features (using the gain ratio
measures) with n ∈ {100, 200, 500, 1000, 2000, 5000, 11374} and tested the poly-
nomial kernel SVM with various values for the complexity parameter (C ∈
{0.001, 0.01, 0.1, 1, 10, and 100}). Similar experiments were conducted deploy-
ing the Random Forest algorithm using several trees (nt ∈ {7, 11, 17, 50, 100}).
Figures 5 and 6 show the graph of f-measure vs. Support Vector Machine
(SVM) algorithm and f-measure vs. Random Forest Algorithm respectively. In
the SVM chart (Fig. 5), as the number of features increases, the performance of
f-measure increases, up to 2000 features. The highest f-measure value of 0.595
was achieved with c = 0.1 and 2000 features in the SVM algorithm experiment.
448 P. Poudyal

Fig. 4. Architecture of argument element identifier (AEI) Module

Fig. 5. F-Measure of SVM algorithm

In case of the graph of f-measure obtained from the Random Forest Algorithm
chart, (Fig. 6) as the number of features increases, a peak f-measure of 0.52
was reached with 1000 features and 100 trees. Then, the f-measure value
decreases up to 2000 and remains constant till 11681 features. We can there-
fore conclude that the SVM algorithm produced better results than the RF
algorithm. Overall, the results achieved are quite promising and support our
proposal for the creation of a new argument mining framework.
 A Machine Learning Approach to Argument Mining in Legal Documents 449

Fig. 6. F-Measure of RF algorithm

4 Conclusion and Future Works

We are proposing a new approach to automatically identify arguments in legal

documents which is phased in three modules: Argument Element Identifier
(AEI), Argument Builder (AB) and Argument Structurer (AS). The preliminary
results of the AEI are extremely promising and support to the development of
a new argument mining framework. Further research must be done on the use
of string kernel as well as other alternative representation models, including
linguistic features such as POS tags, Parse trees and Tree Kernel.

Acknowledgment. The current work is funded by EMMA-WEST in the framework

of the EU Erasmus Mundus Action 2.

References
1. Biran, O., Rambow, O.: Identifying justifications in written dialogs by classifying
text as argumentative. Int. J. Semant. Comput. 5(04), 363–381 (2011). https://
doi.org/10.1142/S1793351X11001328
2. Bunescu, R.C., Mooney, R.J.: A shortest path dependency kernel for relation
extraction. In: Proceedings of the Human Language Technology Conference and
Conference Empirical methods in Natural Language Processing (HLT/EMNLP-
05), pp. 724–731. Association for Computational Linguistics, Stroudsburg (2005).
https://doi.org/10.3115/1220575.1220666
3. Cabrio, E., Villata, S.: Towards a benchmark of natural language arguments. In:
Proceedings of the 15th International Workshop on Non-Monotonic Reasoning
(NMR 2014), Vienna (2014)
4. Doddington, G., Mitchell, A., Przybocki, M., Ramshaw, L., Strassel, S.,
Weischedel, R.: The automatic content extraction (ace) program-tasks, data, and
450 P. Poudyal

evaluation. In: Proceedings of the Fourth International Conference on Language

Resources and Evaluation, vol. 2, pp. 837–840 (2004)
5. Florou, E., Konstantopoulos, S., Koukourikos, A., Karampiperis, P.: Argument
extraction for supporting public policy formulation. In: Proceedings of the 7th
Workshop on Language Technology for Cultural Heritage, Social Sciences, and
Humanities, pp. 49–54 (2013)
6. Mochales, R., Ieven, A.: Creating an argumentation corpus: do theories apply to
real arguments?: a case study on the legal argumentation of the ECHR. In: Pro-
ceedings of the 12th International Conference on Artificial Intelligence and Law,
pp. 21–30. ACM, New York (2009). https://doi.org/10.1145/1568234.1568238
7. Mochales, R., Moens, M.F.: Study on the structure of argumentation in case law.
In: Proceedings of the 2008 Conference on Legal Knowledge and Information Sys-
tems, pp. 11–20. IOS Press, Amsterdam (2008)
8. Mochales-Palau, R., Moens, M.F.: Study on sentence relations in the automatic
detection of argumentation in legal cases. Front. Artif. Intell. Appl. 165, 89–98
(2007)
9. Moens, M.F., Boiy, E., Palau, R.M., Reed, C.: Automatic detection of arguments
in legal texts. In: Proceedings of the 11th International Conference on Artificial
Intelligence and Law, pp. 225–230. ACM (2007)
10. Palau, R.M., Moens, M.F.: Argumentation mining: the detection, classification and
structure of arguments in text. In: Proceedings of the 12th International Conference
on Artificial Intelligence and Law, pp. 98–107. ACM (2009). https://doi.org/10.
1145/1568234.1568246
11. Poudyal, P., Goncalves, T., Quaresma, P.: Experiments on identification of argu-
mentative sentences. In: Proceeding of 10th International Conference on Soft-
ware, Knowledge, Information Management & Applications (SKIMA), pp. 398–
403. IEEE (2016). https://doi.org/10.1109/SKIMA.2016.7916254
12. Poudyal, P., Quaresma, P.: An hybrid approach for legal information extraction.
Front. Artif. Intell. Appl. (JURIX) 250, 115–118 (2012). https://doi.org/10.3233/
978-1-61499-167-0-115
13. Reed, C., Palau, R.M., Rowe, G., Moens, M.F.: Language resources for study-
ing argument. In: Proceedings of the Sixth International Conference on Language
Resources and Evaluation (LREC 2008), Marrakech, Morocco, pp. 91–100 (2008)
14. Salton, G., Wong, A., Yang, C.S.: A vector space model for automatic indexing.
Commun. ACM 18(11), 613–620 (1975). https://doi.org/10.1145/361219.361220
15. Stab, C., Gurevych, I.: Identifying argumentative discourse structures in persuasive
essays. In: Proceedings of the Conference on Empirical Methods in Natural Lan-
guage Processing (EMNLP), pp. 46–56 (2014). https://doi.org/10.3115/v1/D14-
1006
16. Stab, C., Kirschner, C., Eckle-Kohler, J., Gurevych, I.: Argumentation mining in
persuasive essays and scientific articles from the discourse structure perspective.
In: Proceedings with the Workshop on Frontiers and Connections between Argu-
mentation Theory and Natural Language Processing, Bertinoro, Italy, pp. 40–49
(2014)
 Answering Complex Queries on Legal
Networks: A Direct and a Structured IR
Approaches

Nada Mimouni1(B) , Adeline Nazarenko2 , and Sylvie Salotti2

1
Chair Governance Analytics, Université Paris Dauphine – PSL,
75016 Paris, France
nada.mimouni@dauphine.fr
2
LIPN (UMR 7030), Université Paris 13 – Sorbonne Paris Cité & CNRS,
93430 Villetaneuse, France
{adeline.nazarenko,sylvie.salotti}@lipn.univ-paris13.fr

Abstract. This paper highlights the benefit of semantic information

retrieval in legal networks. User queries get more complex when they
combine constraints on semantic content and intertextual links between
documents. Comparing two methods of search in legal collection net-
works, we present new functionalities of search and browsing. Relying
on a structured representation of the collection graph, the first approach
allows for approximate answers and knowledge discovery. The second one
supports richer semantics and scalability but offers fewer search function-
alities. We indicate how those approaches could be combined to get the
best of both.

Keywords: Semantic information retrieval · Legal collection graphs

Exact answers · Approximate search · Knowledge discovery

1 Introduction
Legal practitioners are permanently in the need to search legal collections in
order to assess regulatory texts applicable to a specific case. Needs analysis with
legal experts and partners of the Légilocal project [1] showed that users express
their needs in the form of complex queries that address both the semantic content
and intertextual links between documents. For instance, when drafting an order,
municipal clerks typically have to identify the legislation to refer to as well
as former orders published on the same topic, especially those that have been
appealed. They would enter queries like “Which local acts concerning rural roads
have been appealed and were canceled by court decision? ”.
Legal information retrieval (IR) systems widely used by both citizens and
practitioners are not able to handle such complex queries. They return docu-
ments based on the keyword they contain or are associated to as metadata and
not on intertextual grounds. Answers are returned as a list of documents without
taking into account the graphs to which those documents belong (esp. references
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 451–464, 2018.
https://doi.org/10.1007/978-3-030-00178-0_31
452 N. Mimouni et al.

to other texts). To build the context of a given document, a step of exploratory

search – starting from one returned result and going through its hypertextual
links – is required. In legal collections, where numerous and various types of
links exist between legal sources, users get easily lost.
Moreover, legal IR systems operate as logical search systems since a user
needs to get all texts related to his query and not only the most relevant ones.
Answers are returned when documents match exactly the sent queries. When no
exact answer is found or when too few or too many answers are returned, users
are not satisfied with the results and must continue to query the system until
they get a relevant response.
In an earlier research, we have proposed an approach towards semantic and
graph-based search in networks of legal documents [5] and demonstrated the
potentialities that a graph-based semantic IR approach could offer to legal con-
tent management systems. In this approach, the intertextual information is taken
into account at the querying level to that it is possible to answer complex queries
like “What are the legal decisions that cite Article 1382 of the Civil Code? ”, “I
am looking for decisions that were canceled by the Court of Cassation.” or “What
are the code articles that have been confirmed and which are cited by the munic-
ipal orders dealing with rural roads? ”.
In this work, we propose to extend that logical approach with advanced
search and discovery functionalities, by returning approximate answers when no
exact one is found, by allowing to restrict or enlarge the answers set when too
many or too few are found and by discovering new knowledge not easily detected
by hypertextual navigation from the list of results. Needs analysis showed that
such functionalities are increasingly expected. For instance, legal practitioners
may need to understand what laws cited by a given text are talking about or
what articles citing a given law are dealing with.
We implemented graph-based search using two different approaches: a struc-
tured and a direct one. The first approach creates a semantic hierarchical and
relational structure on the top of the collection thus allowing for advanced search
functionalities but it is limited to sub-collections1 . The second approach can han-
dle larger collections but supports only exact querying.
We present our experimental data on Sect. 2 and our methods of graph-based
IR in Sect. 3. Exact search and browsing functionalities are discussed respectively
in Sects. 4 and 5. Section 6 introduces related works addressing the problem of
modeling and querying graphs of legal collections.

2 Experimental Data

We test and evaluate our approaches on three collections of documents:

– The Noise corpus was used as an illustrative example in Légilocal. It is com-

posed of 10 documents annotated with 10 attributes related to noise and a
1
The target scale is a few hundred documents, but we must further optimize the code
for it (solutions exist, it is an implementation issue).
 Answering Complex Queries on Legal Networks 453

single reference relation. Documents are of different types: local acts (orders)
making reference to legislative texts (decree, law, code, ordinance).
– The Légilocal corpus is a richer collection extracted from the Légilocal base
for demonstration purposes: it is composed of 25 documents and 30 legal arti-
cles, various types of documents are represented (local decisions, legislative
texts, judgements and editorial documents) – sometimes with different ver-
sions of the same document –, as well as various types of links (application,
composition, decision either confirmation or cancelation, modification) [5].
– The ILO corpus is the largest collection with almost 400 documents collected
from the International Labour Office (ILO)2 . Documents are conventions and
recommendations, linked by an implementation relation.

These collections have served us as a primary base to test the proposed

approaches. The documents are analyzed to identify their types and extract their
structure, they are annotated with semantic descriptors and the intertextual
citations are themselves identified and semantically typed. Semantic descriptors
annotating documents correspond to concepts extracted from semantic resources
dealing with the texts domains. Real scale collections are to be created and
documents analyzed by partners of the project.
Needs analysis realized with collaboration of Légilocal legal experts shows
that practitioners’ spontaneous queries naturally combine content and intertex-
tual criteria. We built a list of queries which are formally different from each
other but which covers a wide range of practitioners’ queries. We created a set
of queries on both ILO and Légilocal (LEGI) corpus along with their relevant
answers to test and evaluate our search approaches, the quality of a method
being related to the variety of query types that it supports.

– ILO1: “Which convention implements Recommendation 113 on collective bar-

gaining?”
– ILO2: “Which convention implements the recommendation talking about sea-
farers occupational accidents?”
– ILO3: “What are the recommendations implemented by conventions dealing
with air pollution?”
– LEGI1: “What are the legal decisions that cite Article 1382 of the Civil Code?”
– LEGI2: “What are the articles of the Environment Code that deal with motor
vehicles and rural roads?”
– LEGI3: “I am looking for the decision that is the subject of the judgment A
of the Court of Cassation.”
– LEGI4: “I am looking for decisions that were canceled by the Court of
Cassation.”
– LEGI5: “I wonder if this judgement of the Court of Appeal was itself the
subject of an appeal.”
– LEGI6: “I would like to see municipal orders regarding rural roads that have
been appealed and were canceled by decision of jurisprudence.”

2
www.ilo.org.
454 N. Mimouni et al.

– LEGI7: “What are the code articles which are cited by the municipal orders
dealing with rural roads that have been confirmed?”
– LEGI8: “What are the decisions prior to decision D?”
– LEGI9: “I wonder if the texts referred by municipal orders dealing with rural
roads are also cited by those concerning the motor vehicles.”
– LEGI10: “What are municipal orders that have been the subject of two
appeals?”
– LEGI11: “I wonder if the texts referred by the municipal order 97-17 of
Champigné have been modified, and if so, what are the new versions of these
texts and the source texts of the amendment.”

3 Modeling Legal Collections as Graphs

The document collection is modeled as a graph of documents, which can be
exploited directly using semantic technologies (direct approach) or which can be
further structured into a hierarchical and relational conceptual structure, that
itself supports search and browsing (structured approach).

3.1 Direct Approach

The direct approach models the collection as an RDF graph where the objects
(nodes) are document and descriptor identifiers and where properties (edges) are
intertextual links (between documents) or annotations (between documents and
descriptors). Document descriptors are organized into hierarchies allowing to
exploit the links of specialization/generalization in order to adjust the precision
of queries. Document types (law, decree, etc.) are also organized into a highly
structured hierarchy, which is a feature of the legal domain. The document struc-
ture (sections, paragraphs, etc.) is represented rather the whole document. The
article is considered as the basic documentary unit, it can be directly cited and
has its own life cycle (may be amended, codified, etc.). Intertextual links between
documents are represented as direct relations between two documentary units
(as for citation, visas, application) or as document operations modeling relations
involving more documentary units: source, target and result (as for modification,
codification).
Taking into account the history of the different versions of a legal document
led to distinguish several levels of representation of documentary objects: the
master document, the work, and the various versions that are given, expres-
sions. This involves specifying at what level the intertextual relations are intro-
duced and allows to factorize some document properties on the work avoiding
to duplicate them on each of its expressions.
An OWL document ontology controls the semantics of the RDF graph. It
provides for fine-grained descriptions of legal collections, taking any type of
features into account: semantic metadata, document structure, document types
and intertextual relationships. User queries are expressed in SPARQL and are
directly matched against the RDF graph.
An example of an instance graph is given in Fig. 1 showing the codification
of article L362 − 1 of the Environment Code by the Ordinance n2000 − 914.
 Answering Complex Queries on Legal Networks 455

Fig. 1. Codification of article L362 − 1 of the environment code by the ordinance

n◦ 2000 − 914

3.2 Structured Approach

Relying on FCA/RCA, the structured approach [6] models the collection as a

set of interlinked hierarchical structures which organizes documents into classes
on the basis of their semantic content and their interrelationships (Fig. 2).

Fig. 2. Example of relational lattice family with relations among the formal concepts
of the set of lattices.

The approach is composed of four main steps. The document collection is

first structured into formal concept lattices based on the semantic descriptors
associated with the documents (Semantic content modeling). The formal con-
cepts are clusters of documents (Extent) sharing a set of semantic and rela-
tional attributes (Intent). Those lattices are then enriched with intertextuality
(relational information: relational attributes are added to the extents of formal
concepts) which produces a relational lattice family (Intertextuality modeling).
User queries, possibly combining semantic descriptors and cross-references con-
straints, are matched against the relational lattice family, which allow for return-
ing graphs of documents as answers (Direct querying). The lattice structure can
be further exploited to retrieve approximate results as an alternative or in addi-
tion to direct ones (Browsing).
456 N. Mimouni et al.

Semantic Content. The semantic content of documents is first modeled as a

set of formal contexts using FCA. Each context corresponds to a specific type
of documents and describes a binary relation between a set of objects and a
set of attributes (object × attributes). The objects correspond to documents.
The attributes correspond to the semantic descriptors characterizing the content
of these documents. In an information retrieval (IR) perspective, the lattice
built by FCA on formal contexts gathers all possible combinations of documents
attributes. These combinations are represented by the intensions of concepts
having as extensions all the documents sharing these properties and correspond
to the set of all satisfiable queries on the collection.

Relationships. We use RCA, the relational extension of FCA, to take into

account the intertextual information while modeling the collection. The approach
creates a family of relational contexts from binary contexts (documents × seman-
tic descriptors) and relations represented separately in new contexts, each of
which defines one type of relation between documents (documents×documents).
This family of contexts forms the starting point for the creation of corresponding
conceptual structures called Relational Lattice Family (RLF).

3.3 Comparaison

Figure 3 compares these approaches. We can notice that modeling and encoding
the collection (first and second steps on Fig. 3, bottom side) are more sophisti-
cated with the direct approach: ontological model and RDF triples (the figure
shows the graph of instances). This allows the direct approach to express more
complex query graphs (fourth step of Fig. 3). In the structured approach (top
side on Fig. 3) the initial modeling step is simpler: the collection is modeled as
a set of objects/attributes tables, but there is an additional structuring phase
(third step) creating a structure that can be further exploited for browsing.
In information retrieval, structuring a document collection as a lattice pre-
computes the answers to all queries that are satisfiable on that collection. In
addition, browsing the structure provides approximate answers as it allows for
generalizing or restricting queries. In the direct approach, answers are calculated
on the fly when the SPARQL query is sent to the system. This approach is more
flexible (when the model of the collection or the collection itself evolves), is
not limited by the collection size and allows to express richer queries. On the
contrary, the structured approach performs well on small collections or local
perspectives over large collections and allows browsing and visualizing, but it
cannot answer complex queries like LEGI11. We have developed a prototype to
visualize the answer graphs from the structured approach displaying the objects
of the answer along with all their attributes even if they are not mentioned in
the initial query.
The choice to use one or the other approach depends tightly on the appli-
cation requirements (user interfaces, number of documents, granularity of docu-
ment description or the evolution of the collection).
 Answering Complex Queries on Legal Networks 457

Fig. 3. Structured vs. direct approaches. The first approach (top) relies on FCA/RCA
to build a semantic structure over the graph of documents. Queries are answered
based on that structure. The second approach (bottom) exploits semantic technologies
(OWL/RDF/SPARQL) to model the collection as a graph. Query answers are directly
extracted from that graph. Compared with the structured approach, the direct one
supports richer semantic graphs but offers no exploration facility.

4 Exact Search in Collection Graphs

As presented in [5], document collections and queries are formalized as graphs.

The proposed approaches enable answering different types of elementary and
complex queries. We tested 14 types having or not a focus and with or without
constraints. For instance, Fig. 4 shows two examples of query graphs: qa corre-
sponds to an elementary query and qb to a relational one. The query qa looks for
documents of type t2 annotated with semantic attributes s4 and s5 . The query
qb looks for documents of type t1 annotated with s1 having the link r1 with
documents of type t2 annotated with s4 .

Fig. 4. Examples of query graphs: qa represents an elementary query and qb a relational

query, which is represented with a focus in the formula.

On exact search and for elementary as well as for relational queries, both
approaches perform well. All the selected queries are properly answered, with a
list of documents or document graphs that exactly match the query or no answer
if the query is not satisfiable. We show below some examples of queries, testing
the direct and the structured approaches on our three test corpus: Noise, ILO
and Légilocal corpus.
458 N. Mimouni et al.

4.1 Structured Approach

Once structured as a family of relational lattices, the collection can be exploited
to answer elementary and relational queries. We tested this approach on both
Noise and ILO corpus.
Let’s consider an example of a relational query on the Noise corpus: “Which
orders talk about sound level (nvs) and make reference (rf ) to decrees about noisy
activities (ab)?”. Figure 5 shows the structure built for the collection. The query
is modeled as an attributed graph, which nodes are virtual objects (variables)
created for each unidentified document (Qo for orders and Qd for decrees), which
attributes are semantic descriptors (nvs for node Qo and ab for Qd ) and the link
is the intertextual relation (rf ). An iterative search process calculates relevant
answers [6], it instantiates Qo and Qd respectively by documents AS (Strasbourg
order) and O45 (ordinance 45). It returns the following answer graph (the query
has no focus, so answers are presented as graphs): Ga = T ype(AS, order) ∧
Att(AS, nvs) ∧ T ype(O45, decree) ∧ Att(O45, ab) ∧ Rel(AS, rf, 045).

Fig. 5. Exact and approximate search in a relational lattice family.

We tested the approach on the ILO corpus. Two lattices are build: a conven-
tion lattice and a recommendation lattice. Created concepts are classes of docu-
ments (either conventions or recommendations) sharing a given set of attributes
(semantic descriptors). All the collected queries returned the expected relevant
answers. We detail hereafter two examples: ILO1 and ILO2.
ILO1: this query contains an identified document R113 (recommendation
113), we create the variable QueryConv corresponding to the searched one
that we insert in the recommendation lattice. We locate R113 on the rec-
ommendation lattice, check that it is actually associated with the formal
attribute collective bargaining and search for the related convention (we check
relational attributes). The query graph is:
Gq = T ype(QueryConv, convention) ∧ T ype(R113, recommendation) ∧
Att(R113, collective bargaining) ∧ Rel(QueryConv, impl, R113)
 Answering Complex Queries on Legal Networks 459

The search returns convention C144. The answer graph is depicted on Fig. 6,
with the dashed line surrounding the exact answer.

Fig. 6. Answer graph for query ILO1.

ILO2: the query graph contains two variables, two attributes and one rela-
tion. The query graph is:

Gq = T ype(QueryConv, convention) ∧ T ype(QueryRec, recommendation) ∧

Att(QueryRec, occupational accidents) ∧ Att(QueryRec, seaf arers) ∧
Rel(QueryConv, impl, QueryRec)

The search process instantiates variables QueryConv and QueryRec respec-

tively with documents C164 and R142. The visualisation of the answer graph
is given in Fig. 7. The central part of the graph (surrounded by the dashed
line) corresponds to the exact answer having the attributes of the query.

Fig. 7. Exact and approximate answer graphs for query ILO2.

4.2 Direct Approach

As the direct approach allows to model the collection in a more fine-grained
manner and is more naturally prepared to process large volume of data, it was
460 N. Mimouni et al.

more interesting to test it on the Légilocal collection. Nevertheless, we have

tested it on the ILO corpus in the aim of comparing the approaches.
In the ILO2 query, we search for objects of the class Convention having
the relation implement with Recommandation objects described by semantic
concepts occupational accidents and seafarer. The query is translated in
the following SPARQL query (note that the relation implement is translated as
an object property ilo:implement which domain is class Convention and range
is class Recommandation, attributes are attached to documents via the object
property ilo:hasConcept). It returns documents C164 and R142.
1 SELECT ?conv ?recom
2 WHERE {
3 ?recom ilo:hasConcept ilo:occupationalaccidents , ilo:seafarer .
4 ?conv ilo:implement ?recom .
5 }
The expressivity of the query langage enables composing more complex
queries dealing with versions and fragments of documents, or with relations
with more than two arguments. For instance, the query LEGI11 is translated as
follows:
1 SELECT ?text ?newversion ?oldversion ?source
2 WHERE {
3 ?visa rdf:type lido:VisaCitation .
4 ?visa lido:citationSource :ArreteChampigne97-17 .
5 ?visa lido:citationTarget ?text .
6 ?modification rdf:type lido:Modification .
7 ?modification lido:referenceTarget ?oldversion .
8 ?modification lido:referenceSource ?source .
9 ?modification metalex:result ?newversion .
10 ?text metalex:realizedBy ?oldversion .
11 }

5 Browsing for Approximate Search and Knowledge

Discovery
5.1 Exploration Facilities
When a query doesn’t have any exact answer, it is interesting to return an
approximate answer to the user instead of an empty set of results. Also, when
the user gets too few or too many answers to his query, he/she should be able
to broaden or narrow his/her search by removing or adding constraints to the
initial query.
This could be done by exploring the collection space and selecting documents
that are more or less close to the position of the (possibly empty) exact answer
variable, i.e. documents with less or more semantic or relational properties.
The exploration process could be automated and remain transparent to the user
(approximate results are calculated by the system and returned), or manual with
suitable user interfaces to visualize the data space and navigate within it.
For the time being, this functionality is provided by the structured app-
roach thanks to the structure it builds on the top of the collection, and not
 Answering Complex Queries on Legal Networks 461

by the semantic approach [6]. Structuring the collection has the advantage of
pre-computing the answers to all the satisfiable queries, either elementary or
relational, which allows to find approximate answers to users’ queries and dis-
cover new knowledge, relying on the strategy of [11] without extra calculation.
Some exploration facilities are given in Fig. 8.
Such lattice navigation functionalities cannot be implemented on large col-
lections, but is a track that we explore in the idea of simulating the relevant
browsing strategies using SPARQL. We can actually propose some navigation
scenarios, in the form of automatically generated sequences of SPARQL queries
that would simulate the desired navigation strategies. For instance, the following
method corresponds to a query relaxation procedure:
1. Starting from the initial query, relax one or more constraints (RDF triples)
in the query graph.
2. Instantiate the relaxed SPARQL queries on the graphs of the collection.
3. Propose the obtained results to the user as an alternative or complement to
the exact answer.

5.2 Approximate Search: Examples

Noise Corpus. Let’s consider the example presented above (Sect. 4.1) on the
noise corpus. It is possible to explore the order lattice by relaxing a semantic
constraint or by dropping a relational constraint ending up with an elementary
query (as described in Fig. 8). For instance, on the structure of Fig. 5, if we relax a
semantic constraint by removing the descriptor nvs, we search for orders making
reference to laws and decrees about noisy activities (ab), which are grouped in
concept c13. A relaxed answer-graph Gr is given by the document AP linked to
the document D95:
Gr = T ype(AP, order) ∧ T ype(D95, decree) ∧ Att(D95, ab) ∧ Rel(AP, rf, D95)

ILO Corpus. On Fig. 7, the right side of the graph contains, in addition to
the exact answer, several approximate answers that have only one of the query
attributes: seafarer for R107, R138 and occupational accidents for R164 and
R171. They are obtained by upward navigating the conceptual structure which
accounts for query generalization.
This can be simulated by the following sequence of relaxed SPARQL queries:

1 SELECT ?conv ?recom

2 WHERE {
3 ?recom ilo:hasConcept ilo:seafarer .
4 ?conv ilo:implement ?recom .
5 }
1 SELECT ?conv ?recom
2 WHERE {
3 ?recom ilo:hasConcept ilo:occupationalaccidents .
4 ?conv ilo:implement ?recom .
5 }
462 N. Mimouni et al.

Fig. 8. Exploring facilities offered by the conceptual structure: query broadening and
restricting for an elementary query (diagram 1) or a relational query (diagram 2), sup-
pressing a relational attribute (diagrams 3), adding a relational attribute (diagram 4)

5.3 Knowledge Discovery: Examples

On Fig. 7, the left side of the graph represents the convention which implements
the recommendation R142. The visualisation module presents all the attributes
(semantic descriptors) of the convention C164 on the result graph. This enables
the user to extend the search, looking for conventions that talk about similar
topics, i.e. which are annotated with a subset of its attributes. This corresponds
to the function of search by example [11] where the user has initially a sample,
one document or a set of documents, and looks for similar ones (typically the
case of a municipal clerk having to draft a local act and who starts by looking for
similar ones in neighboring municipalities). We can at this stage (without navi-
gation interface) present directly to the user the similar objects, by performing
a simple search (elementary query with the attributes of C164) on the initial
conceptual structure of conventions.
 Answering Complex Queries on Legal Networks 463

In addition, visualizing the conceptual structure built on the top of a given

collection, allows for discovering clusters of objects that share the same proper-
ties (semantic or relational) without having to formulate queries. These clusters
have served as a basis for the design and restructuring of domain ontologies [4].

6 Related Work
Many works have studied the intertextuality between legal sources and acknowl-
edged that network analysis is a powerful way to model legal collections [2,7,10].
However, as in citation and social network analysis [8], focus has been put on the
network level to identify the most strongly connected sub-collections or the most
influential law sources. Less attention has been paid to the detailed analysis of
intertextuality and the semantics of intertextual links.
In major legal access systems, intertextuality is presented as hypertextual
links between texts or as a metadata that could be queried (e.g. Legifrance3 ).
New approaches [3,9] try to enrich legal access systems with contextual net-
works based on link analysis methods. Their proposals consist mainly on the
exploration and visualization of the citations network. They do not take inter-
textuality into account at the query level as we propose to do in exact and
approximate search.

7 Conclusion
This paper shows that legal practitioners’ complex queries could be handled
through modeling and querying legal collections as semantic networks. Network
nodes (documents) are associated with semantic attributes and connected to
each other by various types of semantic links. Queries are themselves modeled
as graphs, which are matched against the collection graph so as to return exact
or approximate answers to the user. We proposed two methods – direct vs.
structured methods –, which have different advantages and drawbacks. Using
one model or the other depends tightly on the application: the granularity of
the description required for the collection, the complexity of queries, the size
of the collection, the relevance of restricted search perspective. Both proposed
approaches are logical ones, they are not evaluated in terms of precision and
recall. Most important criteria are the expressivity of the query language (which
types of queries could be answered) and performance (collection size, processing
time).
Future steps include the implementation of exploration scenarios as sequences
of SPARQL queries, the design of adequate interfaces to help users create com-
plex relational queries and analyze the returned results and the test and valida-
tion of the direct approach on large scale collections.

Acknowledgments. This work has been partially funded by French Single Inter-
Ministry Fund (FUI-9, 2010–2013) and is supported by the Labex EFL supported by
the French National Research Agency (ANR-10-LABX-0083).
3
http://www.legifrance.gouv.fr/.
464 N. Mimouni et al.

References
1. Amardeilh, F., et al.: The légilocal project: the local law simply shared. In: Legal
Knowledge and Information Systems - JURIX 2013: The Twenty-Sixth Annual
Conference, University of Bologna, Italy, 11–13 December 2013, pp. 11–14 (2013)
2. Fowler, J.H., Johnson, T.R., Spriggs, J.F., Jeon, S., Wahlbeck, P.J.: Network anal-
ysis and the law: measuring the legal importance of precedents at the U.S. supreme
court. Polit. Anal. 15, 324–346 (2007)
3. Gultemen, D., van Engers, T.: Graph-based linking and visualization for legislation
documents (GLVD). In: Network Analysis in Law Workshop (NAiL2013@ICAIL)
associated with ICAIL 2013, Rome, Italy, June 2013
4. Rouane-Hacene, M., Valtchev, P., Nkambou, R.: Supporting ontology design
through large-scale FCA-based ontology restructuring. In: Andrews, S., Polovina,
S., Hill, R., Akhgar, B. (eds.) ICCS 2011. LNCS (LNAI), vol. 6828, pp. 257–269.
Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-22688-5 19
5. Mimouni, N., Nazarenko, A., Paul, È., Salotti, S.: Towards graph-based and seman-
tic search in legal information access systems. In: Legal Knowledge and Informa-
tion Systems - JURIX 2014, Krakow, Poland. Frontiers in Artificial Intelligence
and Applications, vol. 271, pp. 163–168. IOS Press (2014)
6. Mimouni, N., Nazarenko, A., Salotti, S.: A conceptual approach for relational IR:
application to legal collections. In: Baixeries, J., Sacarea, C., Ojeda-Aciego, M.
(eds.) ICFCA 2015. LNCS (LNAI), vol. 9113, pp. 303–318. Springer, Cham (2015).
https://doi.org/10.1007/978-3-319-19545-2 19
7. Boulet, R., Mazzega, P., Bourcier, D.: A network approach to the French system
of legal codes- part i: analysis of a dense network. J. Artif. Intell. Law 19, 333–355
(2011)
8. Rubin, R.: Foundations of Library and Information Science. Neal-Schuman Pub-
lishers, Chicago (2010)
9. Winkels, R., Boer, A., Plantevin, I.: Creating context networks in Dutch legislation.
In: Legal Knowledge and Information Systems - JURIX 2013, Italy. Frontiers in
Artificial Intelligence and Applications, vol. 259, pp. 155–164. IOS Press (2013)
10. Winkels, R., de Ruyter, J.: Survival of the fittest: network analysis of dutch
supreme court cases. In: Palmirani, M., Pagallo, U., Casanovas, P., Sartor, G.
(eds.) AICOL 2011. LNCS (LNAI), vol. 7639, pp. 106–115. Springer, Heidelberg
(2012). https://doi.org/10.1007/978-3-642-35731-2 7
11. Wray, T., Eklund, P.: Exploring the information space of cultural collections using
formal concept analysis. In: Valtchev, P., Jäschke, R. (eds.) ICFCA 2011. LNCS
(LNAI), vol. 6628, pp. 251–266. Springer, Heidelberg (2011). https://doi.org/10.
1007/978-3-642-20514-9 19
 Inducing Predictive Models for Decision
Support in Administrative Adjudication

L. Karl Branting(B) , Alexander Yeh, Brandy Weiss, Elizabeth Merkhofer,

and Bradford Brown

The MITRE Corporation, 7515 Colshire Dr, McLean, VA 22102, USA

{lbranting,asy,bweiss,emerkhofer,bcbrown}@mitre.org

Abstract. Administrative adjudications are the most common form of

legal decisions in many countries, so improving the efficiency, accuracy,
and consistency of administrative processes could significantly benefit
agencies and citizens alike. We explore the hypothesis that predictive
models induced from previous administrative decisions can improve sub-
sequent decision-making processes. This paper describes three datasets
for exploring this hypothesis: motion-rulings, Board of Veterans Appeals
(BVA) decisions; and World Intellectual Property Organization (WIPO)
domain name dispute decisions. Three different approaches for predic-
tion in these domains were tested: maximum entropy over token n-grams;
SVM over token n-grams; and a Hierarchical Attention Network (HAN)
applied to the full text. Each approach was capable of predicting out-
comes, with the simpler WIPO cases appearing to be much more pre-
dictable than BVA or motion-ruling cases. We explore several approaches
to using predictive models to identify salient phrases in the predictive
texts (i.e., motion or contentions and factual background) and propose
a design for incorporating this information into a decision-support tool.

1 Introduction
In many countries, the majority of legal adjudications are administrative, typ-
ified by routine licensing, permitting, immigration, and benefits decisions. The
high volume of these administrative adjudications can lead to backlogs, inconsis-
tencies, high resource loads for agencies, and uncertainty for citizens, notwith-
standing the simplicity and uniformity that often characterizes such cases.
This paper presents the hypothesis that predictive models induced from
previous administrative decisions can improve subsequent decision-making pro-
cesses. The first step in establishing this hypothesis is to show the feasibility of
creating models that predict the outcomes of routine administrative cases. The
second step is to demonstrate how such predictive models can be used to improve
decision processes. Our focus is on assisting individual decision makers by using
predictive models to (1) identify the aspects of the instant case that are most
relevant to its outcome and (2) determine the prior cases that share the most
relevant similarities to the instant case. A promising alternative approach to
decision-process improvement not addressed in this paper consists of improved
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 465–477, 2018.
https://doi.org/10.1007/978-3-030-00178-0_32
466 L. K. Branting et al.

case routing and triage, e.g., assigning cases to specialized decision processes
based on likely outcome and duration or apparent complexity. Since this app-
roach depends heavily on the details of a given agency’s decision processes, we
leave it to future work.

2 Prediction in Law

Predictability is a fundamental property of legitimate legal systems. In tra-

ditional jurisprudence, prediction is based on weighing the relative strengths
of arguments formulated by attorneys for and against a given proposition. In
practice, however, attorneys often depend upon intuitive predictions about the
outcome of cases developed from lengthy experience with judges, juries, and
opposing attorneys in prior cases. Insurance companies have long used statisti-
cal models to estimate the settlement value of claims, and in the 1980s expert
systems were developed to model human expertise at this task [12].
More recently, the development of corpus-based techniques for text analysis
together with the increased availability of large legal text corpora has made
feasible induction of models directly from the texts of case records and decisions
[4]. Predictable aspects of cases include such factors as the following:

– The likelihood of success of a given motion (e.g., for dismissal or for extension
of time) or claim (e.g., for veterans disability benefits)
– The expected award amount for a claim, e.g., the amount of a veteran’s
disability award
– The expected return on civil claim, i.e., expected judgment minus expected
litigation cost
– Expected litigation duration
– Recidivism probability

Predictive models of such aspects of legal cases have the potential to improve
access to justice and the efficiency and consistency of case management. How-
ever, such models can be both opaque and susceptible to bias [14]. The work
described in this paper is intended to mitigate these risks by focusing on pre-
dictive models as aids for improving human decision making rather than as
stand-alone substitutes for human discretion.

3 Datasets

In the United States, the agencies responsible for administrative claims, such as
for veterans benefits, Social Security disability, immigration status, and Medi-
care appeals, all suffer from significant backlogs resulting from the inability of
the agencies to handle their growing case loads with the available resources. As
a first step in engagement with these agencies, whose data have privacy and sen-
sitivity issues, we are developing prototypes on less sensitive, but representative,
datasets.
 Inducing Predictive Models for Decision Support 467

– Motion Rulings
Our first dataset consists of 6,866 motion/order pairs drawn from the docket
of a United States federal district court.1 Motions may be granted, denied, or
granted in part and denied in part, and a single order may rule on multiple
motions, potentially granting some and denying others. To obviate these pro-
cedural complexities, our initial dataset is restricted to orders that either rule
on a single motion or that have rulings of the same type for multiple motions,
i.e., all granted or all denied. Each training instance consists of the text of
the motion, which may contain OCR errors (the original filings were in PDF
format), together with a classification as either “granted” or “denied”.
– Board of Veterans Appeals Decisions
Adjudicative bodies vary in the extent to which case facts and decisions are
published. Many adjudicative bodies publish only decisions but not the fac-
tual record on which each decision is based. In many agencies, the original
decisions are not published, but only appellate decisions. The absence of pub-
lished case records can create a cart-and-horse problem in which agencies are
unwilling to share sensitive data for an unproven decision-support tool, but
the decision-support tool can’t be demonstrated because there is no access to
the data on which it must be trained.
A method of finessing this problem exploits the convention that decisions
generally contain statements of the fact of the case. Decisions with clear sec-
tions can be segmented, with the statement of facts treated as a summary of
the actual case record, and the decision treated as the classification of those
facts in terms of legal outcome. This “bootstrapping” approach was used to
demonstrate the feasibility of predicting decisions of the European Court of
Human Rights in [1]. Of course, decision drafters routinely exclude facts that
are irrelevant to the decision and often tailor statements of relevant facts to
fit the intended conclusions. As a result, bootstrapping is merely a proxy for
the actual task of predicting decisions from raw case facts. However, demon-
strating that decisions can be predicted from statements of fact, even if those
statements are filtered, is an essential first step in demonstrating the feasibil-
ity of prediction in more realistic settings.
Board of Veterans Appeals (BVA) cases2 have clear sections: Issues, Intro-
duction, Findings, Conclusions, and Reasons. The Issues and Introduction
sections contain only facts and contentions, and the decision on each issue is
set forth in the Conclusions section. BVA cases often involve multiple issues,
but issues are consistently numbered in Issues, Findings, and Conclusions
sections. We therefore split each published BVA opinion with n issues into n
instances, one for each issue, in which the facts consist of an issue and the
entire Introduction, and the classification is extracted (using regular expres-
sions) from the numbered paragraph of Conclusion that corresponds to the

1
Document filings in US federal courts are “semi-public” in that they are pub-
licly accessible through PACER (https://www.pacer.gov/login.html), but per-page
charges and primitive indexing impede wholesale document mining.
2
https://www.index.va.gov/search/va/bva search.jsp.
468 L. K. Branting et al.

Issue (i.e., that has the same numbering). The possible decisions on each issue
are (1) the requirements for benefits have been met, (2) the requirements have
not been met, (3) the case must be remanded for additional hearings, and
(4) the case must be reopened. Conversion of all published BVA cases in this
fashion yields 3,844 4-class instances or 1605 2-class (met or unmet) instances.
Unfortunately, the Findings section of BVA cases sometimes contain conclu-
sions about facts not discussed in the Issues and Introduction section, so these
sections are an incomplete proxy for the actual case record. This incomplete-
ness makes it impossible in principle to predict the outcome of all BVA cases
from just the Issues and Introduction.
– WIPO Domain Name Dispute Decisions
The World Intellectual Property Organization (WIPO) publishes decisions
resolving complaints brought against the holder of a domain name that
“is identical or confusingly similar” to a trademark belonging to the com-
plainant.3 WIPO cases have only two possible outcomes: the domain name
is transferred to the complainant or it is not. WIPO cases are clearly seg-
mented into seven sections: Parties, Domain Name, History, Background,
Contentions, Findings, and Decision. The facts of each instance consist of the
concatenation of the first 5 sections, and the classification is “transferred”
or “not transferred”. The WIPO dataset consists of 5,587 instances with a
roughly 10-to-1 class skew in favor of “transferred”.

4 Prediction
The first step in confirming the hypothesis that predictive models induced from
previous administrative decisions can improve subsequent decision-making pro-
cesses is to demonstrate that decision outcomes can be predicted. We experi-
mented with 3 predictive techniques: hierarchical attention networks, support
vector machines (SVM), and maximum entropy classification.

4.1 Hierarchical Attention Networks

In our first approach, we extended the hierarchical neural network model pre-
sented in Yang et al. [15] to predict the outcome of legal cases from free-text
sections of their case records. The original model takes as input a sequence of sen-
tences. A sentence representation is built for each sentence with a bidirectional
gated recurrent unit (GRU) layer over word embeddings. An attention mecha-
nism determines the weight of each time-step’s contribution to a sentence vector.
Then, a second GRU layer operates over the sentence vectors, an attention mech-
anism is applied, and the weighted sentence representations are summed to form
a hidden document representation. In prior work the document representation
was used to predict the ratings of Yelp and movie reviews.
The hierarchical model was extended to account for the deeper structure of
legal case documents. With the intuition that human decisions are informed by
3
http://www.wipo.int/amc/en/domains/decisionsx/index-gtld.html.
 Inducing Predictive Models for Decision Support 469

some combination of the text in each section, we adapted the model architecture
to the structure of cases in the two datasets to which the Hierarchical Attention
Network was applied: WIPO and BVA cases. Input for WIPO cases consisted of
three sections: history, background and contentions. We fed each section sepa-
rately into Yang et al.’s document model, sharing weights. The resulting section
representations were combined to create the case representation. The architec-
ture used for BVA cases, shown in Fig. 1, took as input two sections: the issue and
the introduction. The issue is nearly always only one sentence, so was treated as
a single sequence of words. The introduction may be tens of sentences long and
is passed through the hierarchical architecture described in the paper. The case
representation is a learned transformation of the issue and introduction sections.

Fig. 1. Hierarchical neural model architecture for Board of Veterans Appeals cases.
h case is a learned function of h issue, built from the words in the issue section, and
h intro, built from a hierarchical combination of the words-in-sentences and sentences
in the case’s introduction section.

In our experiments, a fully-connected layer appeared to better combine sec-

tions’ hidden representations than a recurrent layer. We therefore used a hidden
layer size of 50 for the WIPO cases and 64 for the BVA cases. We pre-trained
word embeddings using the word2vec algorithm of [9]. For the WIPO cases,
we pre-trained on only the WIPO dataset; for the BVA cases, we used a sepa-
rate dataset of approximately 50,000 appeals. We applied 30% dropout to delay
overfitting these small datasets and used the Adam optimizer. Our models were
trained on 80% of data, developed with an additional 10%, and the remaining
10% was reserved for testing.
470 L. K. Branting et al.

The BVA model achieved a mean F1 of .738 and overall accuracy of 74.7%.
The architecture reached a mean F1 of .944 on the WIPO cases, with an F1 of
.64 for the 10-times-less frequent negative class and overall accuracy of 94.4%.

4.2 Support Vector Machine

The second approach to decision prediction was Support Vector Machine (SVM)
learning. For the WIPO and BVA datasets, text was converted into n-gram fre-
quency vectors for n = 1 − 4, with only those n-grams retained that occurred
at least 8 times. The result was converted into sparse arff format,4 loaded into
WEKA [7], and evaluated in 10-fold cross-validation using WEKA’s implemen-
tation of Platt’s algorithm for sequential minimal optimization [8,13]. Because
of memory issues, the WEKA SVM was run against only a subset of the entire
WIPO dataset consisting of 649 instances from each category.
In 10-fold cross validation the SVM approach achieved a mean F1 of 0.731
on the BVA dataset, with an overall accuracy of 73%. A mean F1 of 0.950 was
achieved on the WIPO dataset, yielding an overall accuracy of 90.5%.

4.3 Maximum Entropy Classification

The third approach to decision prediction that we explored was Maximum

Entropy (Maxent) classification [3] (often termed logistic regression). We used
the jCarafe5 implementation of Maxent, which adds regularization to mitigate
overfitting, to build a model to predict whether a motion will be granted. Our
features consisted of the party filing the motion, the judge ruling on the motion,
the sub-type of motion, and the sequences of 1 to 4 tokens (alphanumeric char-
acter sequences having non-alphanumeric characters on both the left and right
sequence borders) that occur in the text of the motion.
We observed that the motions contain many tokens that appear only in one
motion and seem to be the result of OCR errors (as noted above, the documents
were filed in PDF format, and some were created by scanning images to PDF).
To remove these artifacts, any token that only appeared in only one motion in
the collection was removed.
There are many different sub-types of motions, e.g., for extension of time to
file, for summary judgment, etc. We found that better accuracy was obtained
by training separate models for motion subtypes rather than training a single
model for all subtypes. Accordingly, we split the motions into the following 3
large classes of sub-types and built a separate prediction model for each class:

– Extension-type motions, such as a motion to extend a filing due date, which

tend to have higher grant rates than motions in general
– Motions of the letter sub-type, which tend to have a slightly lower grant rate
than motions in general

4
http://www.cs.waikato.ac.nz/ml/weka/arff.html.
5
https://github.com/wellner/jcarafe.
 Inducing Predictive Models for Decision Support 471

– Motions not included in either of the 2 classes above

We used 10-fold cross validation to build and test separate models for each of the
3 large classes above and then combined the results. The combined results had
an accuracy of 75%, and the recall, precision and balanced F-score for “granted”
were 54%, 66% and 59% respectively.
The predictive results for the three experiments are summarized in Table 1
below:

Table 1. Frequency-weighted mean F1 for predictive algorithms applied to three deci-

sion datasets. Note that the SVM (Support Vector Machine) result on the WIPO
(World Intellectual Property Organization) dataset is on a balanced subset, rather than
the entire WIPO collection, whereas the hierarchical attention network was applied to
the entire skewed set.

Maximum entropy SVM Hierarchical attn. network

Motion-rulings 0.742 0.757
BVA 0.731 0.738
WIPO 0.950 0.944

5 Decision Support
The results of the prediction experiments indicate that routine adjudications and
orders are predictable to some extent from models trained from text representing
the facts of the case (in the WIPO and BVA datasets) or the motion text (for the
order-prediction dataset). Since this approach does not perform argumentation
mining and has no explicit model of the applicable legal issues and rules, there
is a limit to the predictive accuracy that this approach can achieve except in
highly routine and predictable domains, such as WIPO decisions. However, our
objective isn’t replacement of human discretion, but rather support for human
decision making. Our hypothesis is that predictive models can assist human deci-
sion makers by identifying the portions of the predictive text, e.g., statements of
case facts or motion texts, that are most predictive of the outcome. We hypothe-
size that a decision maker may benefit from having the predictive text identified
even when the decision disagrees with the models prediction. This hypothesis is
based on the observation that one of the challenges of decision making is sifting
through irrelevant portions of the case record to locate the most important facts.
We distinguish two uses of predictive text:

– Highlighting the parts of a document most relevant outcome, e.g., granting

or denying a motion, or accepting or rejecting a claim for benefits, so that the
decision maker can quickly identify the facts determinative of the outcome.
– Highlighting the parts of one document most relevant to assessing the sim-
ilarity or difference between the cases. The Common Law doctrine of stare
472 L. K. Branting et al.

decisis, under which a decision in one case is binding on subsequent sim-

ilar cases, is generally inapplicable to administrative adjudications, even in
countries with Common Law legal systems. Nevertheless, we hypothesize that
enabling decision makers to compare the current case to the most similar prior
cases could make decision making faster and more consistent.

We therefore turn to the issue of how predictive texts can be identified. In

the context of algorithms for prediction based on text, identification of the most
predictive text is a special case of the more general problem of feature selection
[6]. We first discuss methods that are independent of the predictive model, then
turn to those that are derived from the predictive model.

5.1 Salient Fact Detection

Information-Theoretic Relevance Measures. A particularly straightfor-

ward approach to predicting relevance is mutual information (sometimes termed
“average mutual information”), which is a measure of how much knowledge
of the value of one variable reduces uncertainty about the value of another
variable [5]:
 P (X = x, Y = y)
I(X, Y ) = P (X = x, Y = y) log
x y
P (X = x)P (Y = y)

For example, in the WIPO domain, the phrases “Complainant has failed to” and
“did not reply” are among the highest information n-grams, that is, occurrence
counts of those phrases are more predictive of the case decision than occurrence
counts of most other variables (i.e., phrases). Mutual information in itself doesn’t
distinguish phrases like “Complainant has failed to”, which is associated with
successful complainant from phrases like “did not reply”, which are associated
with unsuccessful complaints.
Point-wise mutual information (PMI) (sometimes termed “mutual informa-
tion”) [5] measures the extent to which each particular value of a variable is
predictive of a particular value of another variable:

P (x, y)
P M I(x, y) = log
P (x)P (y)

PMI can be either positive or negative, depending on whether the presence of one
value makes the other more or less likely. Thus, the PMI between “Complainant
has failed to” = true and “transferred” = true is positive, whereas the PMI
between “did not reply” = true and “transferred” = true is negative.

Neural Network Attention. Attention mechanisms for neural networks allow

weights to be learned as part of its feature representation. [2] introduced neural
attention for natural language processing, learning a soft alignment for machine
translation such that certain input words contribute most to output words. In
 Inducing Predictive Models for Decision Support 473

the context of text classification, the attention mechanism determines relative

contributions of words in the input sequence to the prediction. This hierarchical
model has attention over the words in each sentence and over the sentences that
make up each section. The attention operates on output from a bidirectional
recurrent layer, meaning that each time-step folds in context from surrounding
words or sentences but is most responsive to the word or sentence at that time-
step.

Fig. 2. Attention weights for words in a representative sentence in the introduction of

a BVA case. These weights reflect the relative importance of the words in determining
the label assigned by the hierarchical attention network to the case.

Extracting the attention weights enables analysis of the model’s predictions.

Figure 2 illustrates the word-level weights of a representative sentence from a
BVA case, showing that words related to the veteran’s medical condition had a
disproportionate weight in the neural network’s prediction.
The use of attention weights to identify the texts most salient to a decision
maker is illustrated in Fig. 3, which shows an excerpt from a BVA decision in
which the highest-attention sentences are highlighted in color. The sentence in
blue received 74% of the attention weight, and the next most important, shown in
yellow, received 9% of the attention weight. The sentence in blue is, in fact, highly
relevant in that it recharacterizes the issue from being moot into something that
can be granted.
While the accuracy of the hierarchical attention network is similar to the
other learning models, it has the significant advantage that attention weights are
474 L. K. Branting et al.

Fig. 3. A portion of a BVA case. The sentence with the highest proportion of attention
weight, 74%, is shown in blue, and the sentence with the next highest weight, 9%, is
shown in yellow.

specific to token instances rather than, as in the case of linear model weights,
global. A phrase that is insignificant in one context can be very significant in
a different context; this distinction can be identified by hierarchical attention
networks but not by models that produce global weights.

Linear Model Weights. An alternative approach to using a predictive model

to identify the most salient case facts makes use of the feature weights learned
during model training. In linear models, such as maximum entropy and SVM
with linear kernels, feature weights are indicative of relevance of features to the
model’s predictions [10], which in our application consists of predicted case deci-
sions. These feature weights are similar to point-wise mutual information (PMI)
(discussed above), with features increasing the chances of a positive prediction
tending to have positive weights and features decreasing the chances tending to
have negative weights. One difference between PMI and the feature weights for
logistic regression with regularization is that a regression feature weight differs
when a feature either occurs only infrequently (weight magnitude is diminished)
or is correlated with other features in the model (weight is adjusted for the effects
of correlated features on the model).
For example, when maximum entropy/logistic regression (with regulariza-
tion) is applied to the domain of motions for an extension of time, phrases
having a relatively large positive feature weight for “granted = true” include
“to dismiss” (feature weight of 0.319), “dismiss” (0.310) and “with the consent”
(0.285), whereas phrases with relatively large negative feature weights include
“a stipulation” (−0.385) and “stipulation” (−0.735).

5.2 User Interface for Decision Support

The data derived from the predictive models is intended to be displayed to
administrative claimants and decision makers using a Graphical User Interface
(GUI). The complexity of the displayed content along with a need to prevent user
error necessitates a usable interface. We are therefore exploring interface designs
to present this information in a manner that best facilitates its use, that is, to
present data in a manner that improves a decider’s speed and accuracy [11].
 Inducing Predictive Models for Decision Support 475

Fig. 4. A prototype decision-support interface design illustrating how the most salient
case facts can by highlighted to assist with analysis of the case record and comparison
between cases. Phrases highlighted in yellow are associated with rulings in favor of
complainants, whereas phrases highlighted in red are associated with rulings in favor
of respondents.

The preliminary design concept shown in Fig. 4 contains several features that
we hypothesize will assist users with deciding on cases efficiently and accurately.
One feature of this design concept allows a user to view the most relevant cases
in multiple ways (i.e., multi-case comparison, high-level comparison, in depth
comparison). Providing multiple formats for case comparison allows a user the
flexibility to decide how in-depth they would like to view the current case and
previous case information. Another design feature provides the user with conve-
nient access to relevant information (e.g., the rules) during the review process.
This design concept leverages the pattern of open/close panels, which allow the
user the ability to customize their view as they go through the evaluation process.
In order to support efficient comparison, this design also provides a highlight-
ing feature that is intended to allow a user to compare the similarities between
current and previous cases. A future evaluation of this design concept will pro-
vide the information needed to iterate on the design patterns and features. The
overall goal is to provide a satisfactory user experience while also assisting the
decision maker to make quick and accurate assessments of cases.
We plan on conducting an initial experimental evaluation to assess the overall
ability of the combined predictive model and user interface to facilitate improved
speed and accuracy in decision making. We hypothesize that the speed and
476 L. K. Branting et al.

accuracy of decision making can be improved by highlighting the phrases in

cases with facts having the greatest weight under a predictive model and by
retrieving prior cases with the strongest similarity to the current case in terms
of the highest weight phrases. The initial evaluation will be performed using
non-lawyers as subjects and WIPO case outcome detection as the predictive
task, since WIPO cases have relatively simple and predictable facts and issues.
After initial evaluation, we plan to conduct future evaluations that validate this
concept using lawyers and other end-users of this type of decision support system.

6 Summary and Future Work

This paper explores the hypothesis that predictive models induced from previ-
ous administrative decisions can improve subsequent decision-making processes.
Three datasets were developed: motion-rulings, BVA issue decisions, and WIPO
domain name dispute decisions. The ability to predict outcomes in these three
domains was demonstrated using three different approaches for prediction: max-
imum entropy over token n-grams, SVM over token n-grams, and a hierarchical
attention network applied to the full text. This initial evaluation did not estab-
lish the superiority of one approach over another, but rather indicated that the
outcome of routine decisions is predictable using any of various alternative mod-
els trained on the text alone and that predictive accuracy varies depending on
the domain and the nature of the predictive texts.
We propose use of feature weights or network attention weights from these
predictive models to identify salient phrases in motions or contentions and case
facts. We have developed an interface design for presenting this information to
improve decision making, and we propose an experimental evaluation to measure
the extent to which speed and accuracy in decision making can be enhanced by
enhancing the salience of phrases based on their weight in the decision model.
The ultimate objective of this work is to improve the efficiency, accuracy, and
consistency of administrative decision making–the form of adjudication that has
the greatest impact on most citizens–by integrating automated decision models
into the human decision process. This work represents an initial step towards
this objective.

Acknowledgments. The MITRE Corporation is a not-for-profit company, chartered

in the public interest, that operates multiple federally funded research and development
centers. This document is approved for Public Release; Distribution Unlimited. Case
Number 17-2336. 2017
c The MITRE Corporation. All rights reserved.

References
1. Aletras, N., Tsarapatsanis, D., Preotiuc-Pietro, D., Lampos, V.: Predicting judicial
decisions of the European Court of Human Rights: a natural language processing
perspective. PeerJ Comput. Sci. (2016). https://peerj.com/articles/cs-93/
2. Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning
to align and translate. arXiv preprint arXiv:1409.0473 (2014)
 Inducing Predictive Models for Decision Support 477

3. Berger, A.L., Pietra, V.J.D., Pietra, S.A.D.: A maximum entropy app-

roach to natural language processing. Comput. Linguist. 22(1), 39–71 (1996).
http://dl.acm.org/citation.cfm?id=234285.234289
4. Branting, L.K.: Data-centric and logic-based models for automated legal problem
solving. Artif. Intell. Law 25(1), 5–27 (2017). https://doi.org/10.1007/s10506-017-
9193-x
5. Gallager, R.G.: Information Theory and Reliable Communication. Wiley, New York
(1968)
6. Guyon, I., Elisseeff, A.: An introduction to variable and feature selection. J. Mach.
Learn. Res. 3, 1157–1182 (2003). http://dl.acm.org/citation.cfm?id=944919.944
968
7. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., Witten, I.H.: The
WEKA data mining software: an update. SIGKDD Explor. 11(1), 10–18 (2009)
8. Keerthi, S., Shevade, S., Bhattacharyya, C., Murthy, K.: Improvements to Platt’s
SMO algorithm for SVM classifier design. Neural Comput. 13(3), 637–649 (2001)
9. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word rep-
resentations in vector space. CoRR abs/1301.3781 (2013). http://arxiv.org/abs/
1301.3781
10. Mladenić, D., Brank, J., Grobelnik, M., Milic-Frayling, N.: Feature selection using
linear classifier weights: interaction with classification models. In: Proceedings of
the 27th Annual International ACM SIGIR Conference on Research and Develop-
ment in Information Retrieval, SIGIR 2004, pp. 234–241. ACM, New York (2004).
https://doi.org/10.1145/1008992.1009034
11. Nielsen, J.: Usability Engineering. Morgan Kaufmann Publishers Inc., San Fran-
cisco (1993)
12. Peterson, M., Waterman, D.: Rule-based models of legal expertise. In: Walters,
C. (ed.) Computing Power and Legal Reasoning, pp. 627–659. West Publishing
Company, Minneapolis (1985)
13. Platt, J.C.: Fast training of support vector machines using sequential minimal
optimization. In: Schölkopf, B., Burges, C.J.C., Smola, A.J. (eds.) Advances in
Kernel Methods, pp. 185–208. MIT Press, Cambridge (1999). http://dl.acm.org/
citation.cfm?id=299094.299105
14. Sidhu, D.: Moneyball sentencing. Boston Coll. Law Rev. 56(2), 672–731 (2015)
15. Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., Hovy, E.: Hierarchical attention
networks for document classification. In: Proceedings of NAACL-HLT, pp. 1480–
1489 (2016)
 Arguments on the Interpretation
of Sources of Law

Robert van Doesburg(&) and Tom van Engers

Leibniz Center for Law, University of Amsterdam, Amsterdam, The Netherlands

robertvandoesburg@uva.nl

Abstract. Many researchers have worked on formalizing legal reasoning and

the representation of law. Particularly in the last decade progress has been made
in creating formal models of argumentation. We aim to develop an approach that
is not only formally correct, but also can be used and understood by common
legal practitioners and IT-staff members. The approach should provide an
instrument that can be used to inform legal experts on relevant issues when
finding a solution to a case at hand. In this paper, we present a real-time case that
is in discussion within the Dutch Tax Administration, the Ministry of Finance, in
court as well as in Parliament. By making a structured interpretation of sources
of norms relevant for the case, and argument schemes to make differences of
opinion explicit, we aim to demonstrate the FLINT method for the interpretation
of norms, and support a process of redesign of administrative procedures.

Keywords: AI and Law
Knowledge acquisition

Legal knowledge-based systems
E-government
Argumentation

1 Introduction

While formalizing law goes back for over decades, one may even argue back to
Leibniz’ days, we still haven’t managed to develop a method that results in a
knowledge representation that can be used in applications that fulfill the demands of
normal software engineering. It is not that there is a lack of examples of legal
knowledge-based systems that are build using some formal representation of law [1, 2].
The problem is the coder-dependency of the suggested approaches, which limits
scalability and has not yet resulted in knowledge engineering to replace the ‘art-of-the-
expert’ that is currently used to acquire knowledge from sources of law in natural
language [3].
In our work, we focus on the analysis based on the original sources of norms in
natural language, in the legal practice. In previous publications, we have described our
approach, which is based upon a revised model introduced by Hohfeld [11]. This
approach has been applied to different legal domains, including migration law, tax law
and labor law. We have developed a first version of a tool supporting modelers and
legal experts [7]. While we see large-scale implementation of legal knowledge-based
systems as a huge challenge, knowledge-based systems seem to have become some-
what a side issue in the AI and Law community.

© Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 478–492, 2018.
https://doi.org/10.1007/978-3-030-00178-0_33
 Arguments on the Interpretation of Sources of Law 479

Not only scholars with a background in law, but also many AI and Law researchers,
seem to be more interested in conflicts and the use of arguments, rather than the
formalization of mainstream massive legal case handling that is typical for adminis-
trative processes. Designing and executing administrative processes also includes the
need for argumentation on the interpretation of the applicable sources of law for
general application. It is in that sense we are interested in argumentation. This means
we are not using argumentation to argue about claims in a specific case, but we are
using argumentation for arguing about an explicit interpretation of sources of law. The
question whether interested parties in a legal argument share the interpretation of the
meaning of sources of law, is assessed separate from arguments about the facts in a
specific case. This type of argumentation serves a specific purpose. One that is not so
commonly addressed in AI and Law literature, namely it should support us in coming
to decisions rather than understanding a decision that is already been made.
In this paper, we present a case study that is the subject of an ongoing debate in the
Dutch Tax Administration, the Dutch Ministry of Finance and the Dutch Parliament.
We will show how we support the interpretation of sources of norms that are relevant
for the case, and the analysis of arguments about those interpretations. Doing so, we
will comment on statements of the State Secretary for Finance concerning the proce-
dures to deal with objections to decisions of the Tax Administration. We also intend to
fuel the ongoing debate on the need for adjustment of the policy guidelines regulating
those procedures.
In this paper, we focus on interpretations of sources of norms to be used in real-time
legal disputes among lawyers, policy advisors and politicians. The representations used
should be comprehensible for experts from the domain of law, public administration
and politics. Therefore, our focus is not on the formalization of the models used, but on
the ability for domain experts to validate the presented knowledge and use it to argue
about cases, to interpret laws and statues, to make decisions about instructions and
support by IT systems.

2 Method

An important aspect of understanding and settling a legal dispute is the legal qualifi-
cation of acts. In our work, acts play a central role. In earlier papers, we have shown
our formalized version of the Hohfeldian concepts [11], our work is based upon [5–7].
In [6] we discussed the differences between rule-based, case-based, logic based and
frame-based approaches to legal knowledge engineering [16] and the differences
between the FLINT frame-based method in relation to earlier frame-based methods
[4, 12]. The quintessence of the frame-based approach is the transformation of a source
of law in natural language into a (semi-) formal interpretation.
In [7], we presented how this approach can be applied on legal sources using an
example from Dutch Immigration Law, resulting in executable knowledge models
expressed in a domain specific language (DSL): FLINT (Formal Language for the
Interpretation of Normative Theories). This DSL is specific in so far that it is targeted
towards the specific way we express norms.
480 R. van Doesburg and T. van Engers

FLINT differs from earlier frame-based methods because it focuses on the specifics
of the legal domain, and the validation of the results of the analysis by legal experts.
Were earlier frame-based models [4, 12] aim at transforming natural language into a
frame-based model, FLINT aims at making a transcription between sources of norms,
and a formal computational model. Van Kralingen explicitly chose not to pay attention
to problems associated with the interpretation of legal knowledge. Breaux focuses on
requirement engineering [6].
In this paper, we will present interpretations of sources of norms as institutional acts
with a precondition and an effect (postcondition). Furthermore, we use argument
schemes to discuss the validity of the presented interpretations. The computational
aspects of the resulting model are not discussed in this paper.

2.1 Interpreting Sources of Norms

The FLINT frame for interpreting sources of norms is presented in Tables 1 and 2.
Table 1 contains the frame for an institutional act, Table 2 contains the frame for a
duty.

Table 1. The institutional act frame

Institutional act Definition
Identifier Description of the institutional act
Action A type of action that causes the transition of an object
Actor An agent that performs an act
Object The thing that makes a transition as a result of an action
Recipient An agent that receives the result of an action by an actor
Precondition Condition that must be fulfilled to allow an act. The precondition consists
of [institutional facts] and Boolean relations between these [facts]
Postcondition The description of the transition of the object caused by the act, consisting
of the creation and termination of institutional facts or duties
References to References to sources
sources

An act is the description of the transformation of an object caused by an act

performed by an acting agent (actor). The result of the act is received by a receiving
agent (recipient).
An act can be legally qualified as ‘allowed’ if that act is described in a source of
norms, and if the precondition for allowing the act, as described in those same sources,
is met. The result of the act, the transition, is specified in the postcondition. In this
paper, the institutional act frames will be shown next to their sources in natural lan-
guage. In our interpretations, we use the original fragments from sources of norms and
put them in a frame structure to specify the conditions that allows the performance of
an acts, and the results of that performance.
A duty is the obligation of an agent to perform an act in the future. It also requires
an agent that has a claim right on the performance of this future act. A duty can only
 Arguments on the Interpretation of Sources of Law 481

Table 2. The frame of a duty

Duty Definition
<Identifier> Description of the duty
Duty holder The agent holding the obligation to perform an institutional act in
the future
Claimant The agent holding the right to claim the termination of the duty
Creating institutional The institutional act that, if performed, creates the duty
act
Termination The institutional act that, if performed, terminates the duty
institutional act
References to sources References to sources

come to existence if it is created by an institutional act, therefore: no duty can exist

without the institutional act that creates it. The creation of a duty implies there is an
institutional act that results in terminating, or fulfilling, the duty. The frame of the duty
is presented in Table 2.

2.2 Argument Schemes for Normative Systems

In this paper, a set of statements of the Dutch State Secretary for Finance is analyzed.
The statements contain interpretations of sources of norms. Disputes on the correctness
of the interpretations of these sources are presented using argument schemes, see
Table 3.

Table 3. Elements of argument scheme

Elements Definition
Claim The issue in an argumentative discourse
Reasons Facts that support the Claim
Rebuttal Attack on the claim that defeats the claim (the claim is false)
Undercutter Attack on the claim that defeats the relation between reasons and claim (it is
unclear whether the claim is true or false)

These argument schemes are based on the Toulmin’s model of argumentation [8,
15]. The statements analyzed in this paper are considered to be claims, based on
reasons. If a claim is attacked successfully, the attack will defeat the claim. This type of
attack is called a rebuttal. If the relation between a claim and the reasons that support
that claim is successfully attacked, that is called a undercutter [13]. So, a successful
rebuttal results in the claim being false, and a successful undercutter results in
uncertainty whether the claim is true or false. Formalization of argument schemes,
using in tools such as ARAUCARIA and ArguMed [10, 14], lies outside the scope of
this paper.
482 R. van Doesburg and T. van Engers

3 Case Study

The case study presented in this paper, is the subject of a Parliamentary debate on the
quality of instructions used by the Dutch Tax Administration to process the withdrawal
of objections to their decisions based on oral announcements during a telephone
conversation. After receiving an objection, the Tax Administration calls the submitter
of an objection within 3 days to discuss the objection, aiming to set peace. Explaining
the reasons that support the decision can convince the taxpayer that the Tax Admin-
istration took the right decision and that insisting on reconsidering the objected deci-
sion is a waste of time and resources. In such cases the taxpayer may want to withdraw
his/her objection. Roughly 20.000 of the annual 400.000 objections received by the
Tax Administration are withdrawn after telephonic contact.
In January 2016, the withdrawal of one specific objection was disputed in a court of
law. Following the publication of the verdict the State Secretary for Finance received a
request to disclose all instructions and correspondence on policies concerning the
withdrawal of objections based on telephone conversations. The disclosed documents
raised questions by Parliamentarians. During the Parliamentary debate, the State Sec-
retary announced that the procedures for dealing with objections would be revised. The
analysis presented in this paper, is made to support the redesign of the procedures to
deal with objections.

3.1 Setting the Scene

In the case at hand, Ms. Jones lodges an objection to a decision of the Tax Admin-
istration. The relevant acts in the case are the following.
1. In the beginning of January 2016 Ms. Jones lodges an objection against an order of
the Tax Administration.
2. A tax inspector calls Mr. Jones, the husband of Ms. Jones, within 3 days to discuss
the contents of the objection and the continuation of the procedure. During the
telephone call Mr. Jones orally states he wants to withdraw the objection of Ms.
Jones.
3. On January 22, 2016, the tax inspector sends a letter to Ms. Jones that her objection
is withdrawn following a telephone conversation with her husband.
4. Ms. Jones lodges an appeal to the administrative court against the withdrawal of
her objection.
5. On March 10, 2016, a lawyer of the Tax Administration sends the administrative
court a letter that it considers the appeal inadmissible because the objection is
withdrawn.
6. On April 12, 2016, the administrative court rules that the written confirmation that
the objection is withdrawn, can be seen as a written refusal by the Tax Adminis-
tration to make an order. The court orders the Tax Administration to decide on the
objection within 6 weeks. The verdict is published on July 6, 2016.
 Arguments on the Interpretation of Sources of Law 483

7. On July 29, 2016, the Tax Administration receives a request for disclosure of
instructions regarding the withdrawal of objections based on telephone conversa-
tions and internal correspondence on this subject. On September 27, 2016, the
dossier is disclosed.
8. On October 20, 2016, Members of Parliament ask the State Secretary for Finance
questions on the quality of policy guidelines for the withdrawal objections to
decisions of the Tax Administration.
9. On November 3, 2016, the State Secretary for Finance sends his reply to
Parliament.
10. On December 15, 2016, the State Secretary for Finance announces that the pro-
cedures for dealing with objections will be reconsidered.
From the questions posed to the State Secretary we selected the two main questions
for analysis:
1. Can an objection only be withdrawn in writing, or orally during a hearing?
2. Is the Instruction ‘Call in Case of an Objection’ contra legem?
In the following subsections, we will demonstrate how our method can be used to
create interpretations of sources of norms. Furthermore, we use argument schemes to
comment the answers given by the State Secretary. The results of our analysis are
presented such that they can be used to support the redesign of the administrative
procedures for the reconsideration of objected decisions.
For the analysis, we will use English versions of the statues which are unofficial
translations made by the authors in absence of an official version, and the official
English version of the General Administrative Law Act (GALA) [9].

3.2 Can an Objection Only Be Withdrawn in Writing, or Orally During

a Hearing?

The Reply of the State Secretary. In the opinion of the State Secretary, according
article 6:21 GALA and case law, the only way to withdraw an objection, is in writing or
orally during a hearing.
Relevant Sources of Law. The withdrawal of an objection to a decision made by an
administrative authority, e.g. the Tax Administration, is described in article 6:21
GALA. The article is explicitly mentioned in both the question, and in the answer of
the State Secretary.
Article 1:5, paragraph 1, GALA, is added because it contains the definition of an
objection and the fact that an objection should be lodged with the administrative
authority that took the objected decision. Case law, e.g. the verdict in the case of Ms.
Jones, explicitly states that only the person that submits an objection, has the power to
withdraw it. These sources where added because the frame of the institutional act is not
complete without an actor and a recipient. The agent roles are not mentioned in article
6:21 GALA.
As a result of a search for sources that describe the requirements to make an oral
withdrawal undisputable, article 7:7 GALA was included. This article is about the duty
484 R. van Doesburg and T. van Engers

to make a report of a hearing. Additionally, the Memorandum of Explanation GALA

states that the oral withdrawal of the objection should be mentioned in the report of the
hearing.
The files that were made public by the State Secretary also refer to case law that
does allow for withdrawal that is not in writing, or orally at a hearing. It concerns the
withdrawal during a telephone conversation that is not a hearing. This source is not
discussed in the answers of the State Secretary, nor in the Parliamentary debates that
from those answers.
The Interpretation of Sources of Law. Table 4 shows the institutional act of with-
drawing an objection in writing. It shows the action of ‘withdrawing in writing’ of the
objection by the person that lodged the objection. The recipient of the result of the act
is the administrative authority that took a decision. There are no additional conditions
for performing this act. As long as there is an objection, the person that lodged the
objection can withdraw it in writing by sending a withdrawal in writing to the ad-
ministrative authority that took a decision. As a result of this act, the objection is
terminated.

Table 4. Interpreting sources of law on the withdrawal of objections in writing

Sources of Law Interpretation of sources of law
Article 1:5 Section 1 GALA withdraw objection in
“‘Lodging an objection’ means exercising the right writing
conferred by law to request the administrative authority Action: [withdraw in writing]
that took a decision to reconsider it” Object: [objection]
Article 6:21 GALA Actor: [person that lodged the
“An objection or appeal may be withdrawn in writing objection]
It may also be withdrawn orally at a hearing” Recipient: [administrative
Case law authority that took a decision]
ECLI:NL:RBDHA:2016:6098 Precondition: (-)
“The power to withdraw an objection lies exclusively Creating postcondition:
with the person that lodged the objection” [withdrawal in writing]
Terminating postcondition:
[objection]

The oral withdrawal of an objection during a hearing is described separately in

Table 5, because it concerns a different act, with different pre- and postcondition. The
oral withdrawal of an objection at a hearing can be performed by the person that lodged
the objection towards the administrative authority that took decision. The precondition
of this act is that a hearing concerning the objection is taking place. The act can only be
performed during a hearing, whereas the written withdrawal can be performed any
time.
The postcondition of the oral withdrawal is the creation of a duty to <mention the
withdrawal in the report of the hearing>. The administrative authority that took
decision is the holder of this duty, and the person that lodged the objection is the
claimant. The duty can be created by the act withdraw objection orally at a
 Arguments on the Interpretation of Sources of Law 485

Table 5. Interpreting sources of law on the oral withdrawal of objections during a hearing
Sources of Law Interpretation of sources of law
Article 1:5 Section 1 GALA withdraw objection orally at a
“‘Lodging an objection’ means exercising the hearing
right conferred by law to request the Act: [withdraw orally at a hearing]
administrative authority that took a decision to Object: [objection]
reconsider it” Actor: [person that lodged the objection]
Article 6:21 GALA Recipient: [administrative authority that
“An objection or appeal may be withdrawn in took a decision]
writing. Precondition: [hearing]
It may also be withdrawn orally at a hearing” Creating postcondition: <mention the
Article 7:7 GALA withdrawal in the report of the hearing>
“A record shall be drawn up of the hearing” Terminating postcondition: (-)
Case law
ECLI:NL:RBDHA:2016:6098
“The power to withdraw an objection lies
exclusively with the person that lodged the
objection”
Memorandum of Explanation GALA
Withdrawal of an objection or appeal during the
session [hearing] is also possible. (…). In that
case it is required that the withdrawal becomes
sufficiently indisputable. It should therefore be
mentioned in the report of the hearing

hearing. It can be terminated by mentioning of the withdrawal in the report of the

hearing. There may be possibilities for the claimant to terminate the duty by waiving
the claim right, but those possibilities lay outside the scope of this case study.
The institutional act withdraw objection orally at a hearing does not have a
terminating postcondition. The objection is not considered to be withdrawn at the
moment that the withdrawal is uttered during a hearing, but only after the report of the
hearing, that contains the mentioning of the withdrawal, has been established, and can
no longer be disputed.
Table 6 shows the institutional act withdraw objection orally during a telephone
conversation. The source that enables this possibility is a lawsuit against the Tax
Administration. The relevant fragment of the court order is included in the sources of
norms of Table 5. The act to withdraw an objection orally by telephone can be per-
formed by the person that lodged the objection, also referred to as the submitter. The
recipient of the act is the administrative authority that took decision. The action must
be performed during a telephone conversation. The result of the act is the creation of
the duty to <send a written confirmation of the withdrawal>. We assume that the
objection is only terminated after the moment the duty to <send a written confirmation
of the withdrawal> is terminated.
Argumentation. The argumentation scheme below shows that State Secretary is
wrong. There is a possibility to orally withdraw an objection in a telephone conver-
sation that is not a hearing. Although the withdrawal legally lies within the scope of
486 R. van Doesburg and T. van Engers

Table 6. Case law on the withdrawal of objections by telephone

Sources of law Interpretation of sources of law
Article 1:5 Section 1 GALA withdraw objection orally during a
“‘Lodging an objection’ means exercising the right telephone conversation
conferred by law to request the administrative Action: [withdraw orally by telephone]
authority that took a decision to reconsider it” Object: [objection]
Article 6:21 GALA Actor: [person that lodged the objection|
“An objection or appeal may be withdrawn in submitter]
writing. Recipient: [administrative authority that
It may also be withdrawn orally at a hearing” took decision]
Case law Precondition: [telephone conversation]
ECLI:NL:RBDHA:2016:6098 Creating postcondition: <send a written
“The power to withdraw an objection lies confirmation of the withdrawal>
exclusively with the person that lodged the Terminating postcondition: (-)
objection”
ECLI:NL:GHSGR:2009:BI4839
“It must be assumed that a withdrawal of an
objection made by telephone [in a conversation that
is not a hearing], lies within the scope of Article
6:21, paragraph 2, GALA, provided the
administrative authority sends a written
confirmation to the submitter, the submitter agrees
and the procedure is sufficiently carefully
performed”
Memorandum of Explanation GALA
“Withdrawal of an objection or appeal during the
session [hearing] is also possible. (…). In that case
it is required that the withdrawal becomes
sufficiently indisputable. It should therefore be
mentioned in the report of the hearing”

Article 6:21, paragraph 2, GALA, the act that causes the withdrawal is not performed
during a hearing. The reply of the State Secretary is false.
Claim An objection can only be withdrawn in writing, or orally during a hearing.
Data Article 6:21 GALA: An objection or appeal may be withdrawn in writing
(1). It may also be withdrawn orally at a hearing (2).
Rebuttal An objection can be withdrawn during a telephone conversation that does
not qualify as a hearing, provided the administrative authority sends written
confirmation of the withdrawal and the submitter agrees with the written
confirmation

3.3 Is the Instruction ‘Call in Case of an Objection’ Contra Legem?

The Reply of the State Secretary. The State Secretary claims the instruction ‘Call in
Case of an Objection’ is contra legem, i.e. contrary to law. No arguments are given for
 Arguments on the Interpretation of Sources of Law 487

this claim, other than the claim that an objection can only be withdrawn in writing, or
orally during a hearing.
Relevant Sources of Law. On page 4 and 5 of the instruction ‘Call in Case of an
Objection’, version 4 of April 26, 2011, the procedure to withdraw an objection based
on a telephone conversation is described. The instruction contains no references to
other sources of norms.
The Interpretation of Sources of Norms. The fact that an objection can be with-
drawn orally in a setting that is not a hearing, as is shown in Sect. 3.2, does not
necessarily lead to the conclusion that the instruction ‘Call in case of an objection’ is
not contra legem. In Table 6 the interpretation of the instruction ‘Call in Case of an
Objection’ is presented.
Table 7 shows the institutional act withdraw objection orally during a telephone
conversation. The act to withdraw an objection orally during a telephone conver-
sation can be performed by the submitter. The recipient of the act is the tax inspector.
The act must be performed during a [telephone conversation]. The result of the act is
the creation of the duty to <consider the objection not to be withdrawn in case the
withdrawer expresses regrets concerning the withdrawal>. We assume that the objec-
tion is terminated at the time the speech act is performed. If at some point in time the
withdrawer expresses regrets concerning the withdrawal, the tax inspector is assumed
to reverse the withdrawal and to reconsider the objected decision under the regular
procedure.

Table 7. Case law on the withdrawal of objections by telephone

Sources of Law Interpretation of sources of law
Instruction making a telephone call in case of withdraw objection orally during a
an objection telephone conversation
“Following the telephone call the Action: [withdraw orally by telephone]
citizen/entrepreneur can withdraw the Object: [objection]
objection. The call is confirmed by a letter of Actor: [submitter]
the tax inspector, upon request of the Recipient: [tax inspector]
submitter of the objection” Precondition: [telephone conversation]
“Note: In the letter, the withdrawer is given a Creating postcondition: <consider the
14-day period to express regrets related to the objection not to be withdrawn in case the
withdrawal. This 14-day period has an internal withdrawer expresses regrets concerning the
meaning only. If someone expresses regrets withdrawal>
concerning the withdrawal after the 14-day Terminating postcondition: [objection]
period, the tax administration will consider the
objection to be not withdrawn, and will
reconsider the objected decision under the
regular procedure”

Argumentation. The instruction ‘Call in Case of an Objection’ allows the Tax

Administration to refrain from sending a written confirmation of the withdrawal. Case
law only allows for the withdrawal by telephone if a written confirmation of the
withdrawal is send to the submitter of the objection.
488 R. van Doesburg and T. van Engers

Claim The instruction ‘Call in Case of an Objection’ is contra legem.

Data Based on the instruction ‘Call in Case of an Objection’ an objection can
be orally withdrawal without written confirmation if the withdrawer
does not request written confirmation
Undercutter The objection is considered not to be withdrawn, in case the withdrawer
expresses regrets concerning the withdrawal

The claim that the instruction ‘Call in Case of an Objection’ is contra legem is based on
the fact that the instruction made the written confirmation of the oral withdrawal during
a telephone conversation only required if the withdrawer explicitly requests one. The
written confirmation is necessary to make the withdrawal sufficiently indisputable.
Since the objection is reversible by the expression of regret by the withdrawer, the
absence of a written confirmation is no longer a valid support for the claim.

3.4 Reasons for Policy Change

The central matter in this case study is not the case of Ms. Jones, but the fact that the
debate that was sparked by this case lead to the announcement of the State Secretary for
Finance to reconsider the procedures for dealing with objections. The files that were
made public in before and during the Parliamentary debate on this matter, showed
disagreements on legal interpretations within the Tax Administration and the Ministry
of Finance. In the documentation concerning the case, and in the Parliamentary debates
on this matter, these disagreements are not properly addressed. We have used the
FLINT method to make an explicit representation of the positions in the case. Now we
will assess the case of Ms. Jones, using these interpretations, and use the results of that
assessment to deduce aspects that should be addressed during the procedures to pro-
pose policy changes.
What Caused the Conflict in the Case of Ms. Jones? It was not the unlawfulness of
the instruction ‘Call in Case of an Objection’ that caused the Tax Administration to lose
the lawsuit against Ms. Jones. It was the fact that the tax inspector spoke not to Ms.
Jones about her objection and the withdrawal thereof, but to her husband. The
importance of speaking directly to the submitter of the objection is undisputed, and is
mentioned in the instruction.
Additionally, when Ms. Jones lodged an appeal against the withdrawal of the
objection, this was not interpreted as expressing regrets concerning the withdrawal of
the rejection. The case of Ms. Jones is not about improper instructions, but on not
following instructions. It was the publication of the instructions of the Tax Adminis-
tration that sparked a Parliamentary debate on the lawfulness of these instructions.
The only reason to include the case of Ms. Jones in the assessment of the policy
guidelines for dealing with objections is to find out why instructions were not followed
in the case of Ms. Jones.
The Right to Regret a Withdrawal by Telephone Was Kept a Secret. The
instruction to make a telephone call in case of an objection, was not made public. As a
result of that, taxpayers did not know they had the right to regret the oral withdrawal of
 Arguments on the Interpretation of Sources of Law 489

an objection during a telephone conversation. Also, the tax inspector that dealt with the
appeal of Ms. Jones, was unaware of the possibility to regret a withdrawal if it (the
withdrawal) was withdrawn orally during a telephone conversation. Instead of
reversing the withdrawal he decided to declare the appeal of Ms. Jones inadmissible. If
citizens are granted rights, whether by law or some other source of norms, the existence
of these rights should be publicly announced.
Why Is a Written Confirmation of an Objection so Important? The withdrawal of
an objection cannot be reversed. It is for that reason that the Memorandum of
Explanation GALA requires administrative organizations to make it sufficiently
indisputable that the submitter of an objection performed the act of withdrawing it. The
reason to claim that the instruction ‘Call in case of an objection’ is contra legem, is that
the possibility to withdraw an objection, without written confirmation, can harm the
legal position of the submitter. Since, according to the instruction, the submitter can
express regret concerning the withdrawal indefinitely, and the expression of regret is
supposed to be followed by a reconsideration of the objected decision under the regular
procedure, the legal position of the submitter of an objection is not harmed because of
the instruction itself. The possibility to reverse an objection, should be properly
addressed during the redesign of the procedures for dealing with objections.
Is it Desirable to Allow for a Reversal of the Withdrawal Indefinitely? The reason
to limit the period that objections can be made regarding decisions of an administrative
authority, is the desire to make legal decision incontestable after a fixed period of time,
usually 2–6 weeks. By allowing regrets to the withdrawal of objections indefinitely, the
objected decision remains contestable forever. When a policy change is proposed, the
period in which the withdrawal of an objection can be revoked, should be addressed. At
present, written withdrawals cannot be reversed. The oral withdrawal during a hearing
becomes definitive if the submitter does not dispute the report of the hearing within
several weeks.
Legal Protection by procedures Versus the Accessibility of those Procedures for
Everyone. The instruction ‘Call in case of an objection’ is a derivative of the policy
notes “Pleasant relations with the Government” and “Professional Handling of
Objections”, published by the Ministry of the Interior, and the Ministry of Justice in
2011 and 2014. The goal of these policy notes was to give guidelines for informal
procedures to the assessment applications, objections, and complaints by government
agencies. The notes describe the margins for deviating from formal procedures.
A proposal for redesign of the procedures for dealing with objections should address
the tradeoff between the value of legal protection by formal procedures and the need for
informal conversations to give laymen access to their rights.

4 Conclusion

In this paper, we show how the FLINT representation formalism that we developed for
creating explicit interpretations of legal sources can be used to create an explicit
representation of the data of a case study, e.g. on the redesign of procedures for the
490 R. van Doesburg and T. van Engers

withdrawal of objections. The representation is superior to the representation in natural

text as normally used in legal disputes, because the frameworks enforces completeness
and by doing so making relations between different sources of norms explicit. The
method to represent institutional acts resulted in adding relevant sources. Sources that
were included in the disclosed dossier of instructions regarding the withdrawal of
objections based on telephone conversations and internal correspondence on this
subject, but were not used in arguments by the State Secretary in his correspondence
with Parliament.
The case study presented in this paper, was originally assessed by legal experts of
the Tax Administration and the Ministry of Finance. This resulted in two legal state-
ments that are successfully attacked in our representation. These attacks were not made
during the parliamentary debates about the case.
The claim that the withdrawal of an objection can only be withdrawn in writing, or
orally during a hearing, is successfully attacked by a rebuttal. The claim is false, an
objection can, under specific conditions, also be withdrawn orally during a telephone
conversation that is not a hearing.
The claim that the instruction ‘Call in case of an objection’ is contra legem is
attacked by an undercutter. Because the objection is considered not to be withdrawn in
case the withdrawer expresses regrets concerning the withdrawal, the ground for the
requirement to send a written confirmation is defeated. But since the right to reverse a
withdrawal was kept a secret, it is unclear whether the claim that the instruction is
contra legem is true or false.
The representation presented in this paper also resulted in three specific aspects to be
discussed in the coming reconsideration of the procedures for dealing with objections and
the withdrawal thereof: regulating possibilities to reverse the withdrawal of objections
made by telephone by expressing regrets (1), regulating the period in which a withdrawal
can be reversed (2), the public announcement of rights granted to citizens (3).
Furthermore, the explicit representation of the reasons to support and attack the
claims in the case study, makes them open for critical questions.

5 Future Work

In our ongoing research on developing a formal method that enables automated legal
reasoning, we have developed a domain specific language, FLINT, to express the
frames as described in this paper. The FLINT expressions are close enough to the
natural language of the sources of law, allowing legal experts to validate these inter-
pretations. FLINT furthermore enables automated reasoning about the consequences of
such interpretations. Validating the genericity and testing scalability of the DSL and
specification and execution environment are currently done in various projects. In those
projects, we are also working on instruments to support the transition of sources of
norms into a formalized representation.
Our approach puts much more focus on the correct interpretation of the sources of
law compared to traditional knowledge based systems approaches. With our approach,
the debate on the correctness of an interpretation consists of separate debates on the
correct selection of sources of norms (1), and the correct mapping of concepts retrieved
 Arguments on the Interpretation of Sources of Law 491

from those norms onto the elements in the institutional act frame, or the duty frame (2).
Furthermore disputes on interpretations of sources of law are separated from disputes
on the correct application of rules in a specific case.
The frame-based approach presented here, has been tested with experts in the Dutch
government, but, as stated before, broader validation is planned. The development of
tool support for our method is ongoing and we aim to further develop our method and
supporting tools in collaboration with practitioners and academics in the years to come.
As we strive for large scale application of our method and supporting tools we will also
focus on the knowledge acquisition aspects that come with our formal models. This
includes knowledge capturing using natural language processing techniques, and
visualizations that are understood by a broader audience, may be even laymen. Inter-
preting sources of law may not be the easiest task, but being able to present inter-
pretations in a clear way is in the interest of all.

References
1. Bench-Capon, T., et al.: A history of AI and Law in 50 papers: 25 years of the international
conference on AI and Law. Artif. Intell. Law 20(3), 215–319 (2012)
2. Bench-Capon, T., Sartor, G.: A model of legal reasoning with cases incorporating theories
and values. Artif. Intell. 150(1–2), 97–143 (2003)
3. Boella, G., et. al.: A critical analysis of legal requirements: engineering from the perspective
of legal practice. In: IEEE 7th International Workshop on Requirements Engineering and
Law (RELAW), Karlskrona, Sweden, pp. 14–21 (2014)
4. Breaux, T.D.: Legal requirements acquisition for the specification of legally compliant
information systems. Ph.D. thesis, North Carolina State University, Raleigh (NC) (2009)
5. van Doesburg, R., et. al.: Towards a method for a formal analysis of law. In: Study case
Report ICT with Industry Workshop 2015, NWO (2016). http://www.nwo.nl/over-nwo/
organisatie/nwo-onderdelen/ew/bijeenkomsten/ict+with+industry+workshop/proceedings.
Accessed 28 Apr 2017
6. van Doesburg, R., van Engers, T.: Perspectives on the formal representation of the
interpretation of norms. In: JURIX 2016, pp. 183–186. IOS Press Amsterdam (2016)
7. van Doesburg, R., van Engers, T., van der Storm, T.: Calculemus: towards a formal language
for the interpretation of normative systems. Artif. Intell. Justice 1, 73 (2016)
8. van Eemeren, F.H., et al.: Handbook of Argumentation Theory. Springer, Dordrecht (2014)
9. General Administrative Law Act from Dutch National Government. http://archief06.
archiefweb.eu/archives/archiefweb/20180224072222/https://www.rijksoverheid.nl/binaries/
rijksoverheid/documenten/besluiten/2009/10/01/general-administrative-law-act-text-per-1-
october-2009/gala18-11-09.pdf. Accessed 30 Sep 2018
10. Hitchcock, D., Verheij, B.: Arguing on the Toulmin Model: New Essays in Argument
Analysis and Evaluation, 1st edn. Springer, Berlin (2010). https://doi.org/10.1007/978-1-
4020-4938-5
11. Hohfeld, W.N., Cook, W.W.: Fundamental Legal Conceptions as Applied in Judicial
Reasoning, and Other Legal Essays. Yale University Press, New Haven (1919)
12. van Kralingen, R.W.: Frame-Based Conceptual Models of Statute Law. Kluwer, Dordrecht
(1995)
492 R. van Doesburg and T. van Engers

13. Pollock, J.: Defeasible reasoning. Cognit. Sci. 11(4), 481–518 (1987)
14. Reed, C., Rowe, G.: ARAUCARIA: software for argument analysis, diagramming and
representation. Int. J. Artif. Intell. Tools 13, 961 (2004)
15. Toulmin, S.E.: The Uses of Argument. Cambridge University Press, Cambridge (1958)
16. Valente, A.: Legal Knowledge Engineering: A Modeling Approach. IOS Press, Amsterdam
(1995)
Courts, Adjudication and Dispute
Resolution
 Dynamics of the Judicial Process
by Defeater Activation

Martı́n O. Moguillansky(B) and Guillermo R. Simari

Institute for Computer Science and Engineering (ICIC) – CONICET,

Universidad Nacional del Sur (UNS), Bahı́a Blanca, Argentina
{mom,grs}@cs.uns.edu.ar

Abstract. We present a novel activating approach to Argument Theory

Change (ATC) for the study of the dynamics of the judicial process. ATC
applies belief change concepts to dialectical argumentation for altering
trees upon which the semantics for reasoning are defined. The activating
approach to ATC considers the incorporation of arguments to define a
revision operator for studying how to provoke change to the semantics’
outcome. Our objective is to contribute to the discussion of how to deal
with circumstances of the judicial process like hypothetical reasoning for
conducting investigations of a legal case, and for handling the dynamics
of the judicial process. We finally observe the behavior of our proposal
upon the sentences of two different real criminal procedures.

1 Introduction and Motivation

Argumentation provides a theoretic framework for modeling paraconsistent rea-
soning, a subject of utmost relevance in areas of research like medicine and law.
For instance, legal reasoning can be seen as the intellectual process by which
judges draw conclusions ensuring the rationality of legal doctrines, legal codes,
binding prior decisions like jurisprudence, and the particularities of a deciding
case. This definition can be broaden to include the act of making laws. Observe
that the evolution of a normative system –for modeling promulgation of laws–
would imply the removal/incorporation of norms for ensuring some specific pur-
pose but keeping most conflicts from the original framework unaffected. As an
example, we briefly refer to the Argentinean broadcasting media law reformed
during 2009. The previous media law, promulgated by the latter de facto regime,
empowered the government to regulate the different media allowing total control
of news. When democracy was restituted, the regulation of media was extended
to private investment groups. As years went by, these groups took over majorities
of types of media, conforming monopolies in some cases. This brought excessive
power to groups with partial interests, allowing them to manipulate the social
opinion about the actual government, and even to condition politicians, thus
striking to national sovereignty. Article 161 of the new media law became one of
the most controversial points, since it forces monopolistic enterprises to get rid
of part of their assets in a maximun period of 1 year. Some enterprises warned
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 495–512, 2018.
https://doi.org/10.1007/978-3-030-00178-0_34
496 M. O. Moguillansky and G. R. Simari

that they would be forced to sell off their assets at very low prices. This opposes
to article 17 of the National Constitution which speaks about private property
rights. Moreover, some members of the Supreme Court think that article 161
recalls the control over the media exercised by totalitarian regimes, which would
violate article 1 of the National Constitution. In fact, such situation could evolve
to a distrust state on the principle of legal security. Hence, an informal analy-
sis of the case imply working with conflicts among arguments in a sense like:
art 17 counter-argues art 161, art 1 and principle of legal security
counter-argues art 161, and so on. (The new media law keeps being discussed
in Argentina nowadays.) Consequently, the objective of this article is to study
the bridge between the dynamics of the legal reasoning and the dynamics of argu-
mentation. To that end, we rely upon abstract argumentation for constructing
a theory of argumentation dynamics, and afterwards we study the dynamics of
real legal procedures by analyzing the behavior of the theory upon the legal
case’s argumentation. However, the arguments that follow from a legal sentence
are constructed in an informal manner, since argument mining from legal texts
is an independent area that lies beyond the scope of this article.
Argument Theory Change (ATC) [18–20] relies upon the theory of belief revi-
sion [3,16] to define models of argumentation change which are devoted to the
alteration of dialectical trees pursuing warrant of a particular argument. Dialec-
tical trees are constructed from a graph of arguments where nodes stand for
arguments and arcs for the attack relation between pairs of arguments. The root
of a dialectical tree is the argument standing for the main issue in dispute of the
reasoning process. The dialectical argumentation semantics analyzes the dialec-
tical tree and as a result, the root argument may end up warranted, proposing a
reason to believe that the issue in dispute should prevail. This form of argumenta-
tion reasoning brings a theoretical perspective that adapts well for modeling the
dynamics of the judicial process. The dynamic abstract argumentation frame-
work [21] extends Dung’s framework [14] in order to consider (1) subarguments
(internal, necessary parts of an argument that are arguments by themselves),
and (2) a set of active arguments (those enabled to be used in the reasoning
process). The main contribution provided by ATC is a revision operator that
revises a theory by an argument seeking its warrant. For achieving warrant, a
revision operator can be defined through deactivation of arguments [20]: the
reasoning process disconsiders some arguments for ensuring warrant of the argu-
ment at issue. A complementary approach to such revision operation involves
only activation of arguments [18]. This approach complements the one based
on deactivation because sometimes it is either not feasible to deactivate argu-
ments or the activation of another provokes less change. For instance, studying
change in a normative system may imply the derogation of norms through the
incorporation of a derogative norm. In this case, the model for derogation is
achieved through the activation of an argument standing for a derogative norm
in contrast to the proposal of the deactivating approach in which the old norm
could only be disconsidered or removed from the system. The activating app-
roach increases the amount of active arguments to achieve warrant. However,
 Dynamics of the Judicial Process by Defeater Activation 497

since defeaters might be unavailable to be activated, the revision could be not

always successful. Moreover, it will be clear that each activation of an argument
results in the possible activation of more arguments, thus provoking preexistent
arguments to play additional roles.
In this article we present an original model of ATC referred as global activat-
ing approach which develops the objective of the activating ATC presented in [18]
in a different perspective more related to the deactivating model for logic-based
argumentation presented in [19]. The advantage of the global activating model
includes reducing the theoretical complexity towards an appropriate axiomatic
characterization. Our main objective is to contribute to the discussion of virtues
and shortcomings of the activating approach as a theoretical resource for ana-
lyzing the current epistemic state of a judicial procedure under the presumption
that such a legal practice is dynamic by nature. This means that in order to
consider dialectical argumentation as a theoretic model of judicial procedures
we necessarily must rely upon a theory with capabilities for handling dynam-
ics of arguments. It is important to mention that we will only concentrate on
the particularities related to the argumentation dynamics and not on legal con-
cepts –like burden or standard of proof– related to the specific court to which
a judicial procedure corresponds (see [6]). However, in Sect. 3, we analyze the
sentences of two different real legal cases: a Supreme Court’s for analyzing the
dynamics throughout appellate instances of the judicial process, and secondly,
the argumentation related to a process of a criminal court. To that end, we build
the corresponding arguments and show that the judicial process, either from the
judge’s or from the lawyer’s viewpoint, is dynamic by nature by applying the
global model of argumentation dynamics presented here. However, we firstly
present the change operation from an abstract perspective (Sect. 2), following a
practical take on this matter. There, the usual principles of change from belief
revision are considered. The axiomatization of the proposed operator should con-
form to postulates adapted to dialectical argumentation (see [19]) by following
such principles. This part of the work is underway.

2 A Model of Abstract Argumentation Dynamics

2.1 Fundamentals for the Argumentation Framework
Arguments, in the usual sense, are interpreted as a reason for a certain claim from
a set of premises. In abstract argumentation [14], these features are abstracted
away; hence we will work with arguments as “black boxes of knowledge” which
may be divided in several smaller arguments, referred to as subarguments. In the
dynamic framework used here we assume a universal set of arguments holding
every conceivable argument that could be used by the inference machinery, from
which a subset of active arguments can be distinguished. These arguments rep-
resent the current state of the world and are the only ones to be considered to
compute warrant. Therefore, at a given moment, those arguments that are inac-
tive represent reasons that, though valid, cannot be taken into consideration due
to the current context. For instance, the presentation of an appeal outside the
498 M. O. Moguillansky and G. R. Simari

acceptable period of time would render this argument inactive. That is, although
the appeal constitutes a valid reason, it is incompatible with the current state
of the world. A similar concept to inactive arguments is given in [23] to inad-
missible arguments. Another configuration for inactiveness can be explained in
the context of legal reasoning when some norm can be inferred from multilateral
treaties and therefore, although it can be considered to be part of the norma-
tive system, its application depends on the decision of the Supreme Court for
establishing jurisprudence regarding its applicability. An example in this sense
is developed in Sect. 3.1. In this section, we will abstract away from any compre-
hensive mechanism for activating arguments. Afterwards, the ATC’s activating
global model will handle activation of arguments in a proper manner, seeking
for a concrete objective. These changes will be performed at a theoretical level,
i.e., any inactive argument could be eventually activated. Nonetheless, a practi-
cal implementation for legal reasoning should restrict which inactive arguments
are currently available to be activated according to interpretation of norms,
which lies beyond the scope of this article. The abstract dynamic framework is
defined next.
Definition 1 (DAF). A dynamic abstract argumentation framework
(DAF) is a tuple U, →, [A], where U is a finite set of arguments called uni-
versal, A ⊆ U is called the set of active arguments, → ⊆ U × U denotes the
attack relation, and  ⊆ U × U denotes the subargument relation.
As said before, an argument B ∈ U may be composed by other arguments.
Since an argument is considered a subargument of itself, it holds that BB,
and therefore an argument always has at least one subargument. To illustrate
the use of arguments and subarguments, consider for instance an argument B:
“the fact that Steven did not turn in the report indicates that he is irresponsible;
since Steven is irresponsible, he should be fired ”. There we can identify two
subarguments, B1 : “failing in turning in reports suggests irresponsibility” and
B2 : “irresponsible people should be fired ”.
The subset of inactive arguments is identified as I = U \ A. This set contains
the remainder of arguments (within the universal set) that is not considered by
the argumentative process at a specific moment. The principle characterizing
argument activation is:

(Activeness Propagation) B ∈ A iff B  ∈ A for any B  B.

A set of active arguments determines that an argument containing them and

only them is also active. Furthermore, an argument becoming active makes all of
its subarguments to become active. An operator Cap : P(U) → P(U) is assumed
to implement the closure under activeness propagation. Thus, U, →, [A] is
closed iff A = Cap (A). To represent change over the set of active arguments we
assume an activation operator ⊕ : P(U) × P(U) → P(U) such that Ψ1 ⊕Ψ2 =
Cap (Ψ1 ∪ Ψ2 ), with Ψ1 , Ψ2 ⊆ U.
Usually, the acceptability analysis over an argumentation framework is made
over the graph of arguments implicit from the framework. In this article, we
 Dynamics of the Judicial Process by Defeater Activation 499

build and evaluate a dialectical tree rooted in the argument under study in
order to determine whether it is warranted. This approach allows us to analyze
only the relevant portion of the arguments graph since we are evaluating the
warrant status of a single argument. A dialectical tree is conformed by a set of
argumentation lines; each of which is a non-empty sequence λ of arguments from
a DAF, where each argument in λ attacks its predecessor in the line. The first
argument is called the root, and the last one, the leaf of λ. We refer to a DAF as a
dynamic argumentation theory (DAT) if it is closed under activeness propagation
and whose argumentation lines are built according to the dialectical constraints
(DCs) [15]: different restrictions on the construction of argumentation lines which
ensure every line is finally acceptable. We assume DCs to avoid constructing
circular argumentation lines. The domain of all acceptable argumentation lines in
a DAT T, is noted as LinesU A U
T , while LinesT ⊆ LinesT will be the domain enclosing
every acceptable line containing only active arguments. The root argument of a
line λ from a DAT T will be identified through the function root : LinesU T → U,
while the leaf of λ through a function leaf : LinesU T → U. From now on, given
a DAT T, to refer to an argument A belonging to a line λ ∈ LinesA T , we will
overload the membership symbol and write “A ∈ λ”, and will refer to λ simply as
argumentation line (or just line) assuming it is acceptable. Since argumentation
lines are an exchange of opposing arguments, conceptually we could think of it
as two parties engaged in a discussion: one standing by the root argument and
the other arguing against it. Consequently, given a line λ, we identify the set of
pro (resp, con) arguments containing all arguments placed on odd (resp, even)
positions in λ, noted as λ+ (resp, λ− ).

Definition 2 (Upper Segment). Given a DAT T and a line λ ∈ LinesU T such

that λ = [B1 , . . . , Bn ], the upper segment of λ wrt. Bi (1 ≤ i ≤ n) is defined
as λ↑ [Bi ] = [B1 , . . . , Bi ]. The proper upper segment of λ wrt. Bi (i = 1) is
λ↑ (Bi ) = [B1 , . . . , Bi−1 ].

We refer to both proper and non-proper upper segments simply as “upper

segment” and either usage will be distinguishable through its notation (round or
square brackets respectively). As stated next, the upper segment of an argument
in a line constitutes a (non-exhaustive) argumentation line by itself. (Proofs are
omitted due to space reasons.)
Proposition 3. For any λ ∈ LinesU ↑ U
T , if B ∈ λ then λ [B] ∈ LinesT .

As said before, the warrant status of an argument will be determined by

analyzing the dialectical tree rooted in it. On the other hand, a dialectical tree
rooted in an argument A will be built from a set of argumentation lines rooted
in A.
Definition 4 (Dialectical Tree). Given a DAT T, a dialectical tree TT (A)
rooted in A is built by a set X ⊆ LinesU T of lines rooted in A, such that an
argument C in TT (A) is: (1) a node iff C ∈ λ, for any λ ∈ X; (2) a child of a
node B in TT (A) iff C ∈ λ, B ∈ λ , for any {λ, λ } ⊆ X, and λ↑ [B] = λ↑ (C). A
leaf of any line in X is a leaf in TT (A).
500 M. O. Moguillansky and G. R. Simari

However, the acceptability of a dialectical tree will depend on the set X

of lines used to build such tree. Hence, an acceptable dialectical tree will be
constructed from a bundle set ST (A) which –given a DAT T– contains all the
acceptable and exhaustive argumentation lines from LinesA T rooted in A. (We
refer to a line as exhaustive when no more arguments can be added to it.) Thus,
following Definition 4, TT (A) is acceptable if it is built from a set X = ST (A).
The domain of all acceptable dialectical trees from T is noted as TreesT . We
will refer to acceptable dialectical trees simply as dialectical trees (or trees),
unless the contrary is explicitly stated. Besides, we will overload the membership
symbol and write “λ ∈ TT (A)” when the line λ belongs to the bundle set ST (A)
associated to the tree TT (A) ∈ TreesT .
Dialectical trees allow to determine whether the root node of the tree is war-
ranted or not. This evaluation will weigh all the information present in the tree
through a marking criterion to evaluate each argument in the tree by assigning
them a mark within the domain {D, U }, where U (resp., D) denotes an unde-
feated (resp., defeated) argument. We adopt a skeptical marking criterion: (1)
all leaves are marked U ; and (2) every inner node B is marked U iff every child
of B is marked D, otherwise, B is marked D. This marking has been also used
to implement DELP [15], a rule-based argumentation formalism. Using this par-
ticular criterion reduces the conceptual complexity of the approach presented
here, allowing us to focus exclusively on the change mechanism. Finally, warrant
is specified through a marking function Mark : TreesT → {D, U } returning the
mark of the root. Thus, given a DAT T, an active argument A ∈ A is warranted
iff Mark(TT (A)) = U . Whenever A is warranted, the dialectical tree TT (A) is
called warranting tree; otherwise, it is called non-warranting tree.
The digraph depicted in figure (a) describes a DAT, where triangles are
arguments and arcs denote attack. The sets A of active and I of inactive argu-
ments (dashed triangles) compose the universal set U. Subarguments were drawn
inside their superarguments. A’s super-
argument is inactive because it con-
tains an inactive subargument. Figure
(b) shows a dialectical tree spanning
the graph from A. Note that inactive
arguments are not considered in the
tree, which is non-warranting. This sta-
tus would change by activating, for
instance, a defeater for the root’s left
defeater.

2.2 A Global Activating Approach

ATC defines a change operator that revises a DAT by an argument, making the
necessary modifications to the theory to warrant that argument, by analyzing the
dialectical tree rooted in it. The core of the change machinery involves the alter-
ation of some lines in such dialectical tree when it happens to be non-warranting.
Therefore, the objective of altering lines is to change the topology of the tree
 Dynamics of the Judicial Process by Defeater Activation 501

containing them in order to turn it to warranting. Alteration of lines comes from

changes applied to the set of active arguments in the theory; that is, arguments
cannot be simply removed/added to/from the tree. Hence, since an argument
could appear in different positions in several lines in a tree, an alteration of a line
could result in collateral alterations of other lines. In this activating approach,
by alteration we refer to the addition of an argument (i.e., a defeater) into a
line. This may end up extending the line and even incorporating new lines to
the domain. Next, we introduce the notion of attacking lines [19]: argumentation
lines that are somehow “responsible for a non-warranting dialectical tree”.

Definition 5 (Attacking Line). Given a tree TT (A); a line λ ∈ TT (A) is

referred as attacking iff every B ∈ λ+ (resp., B ∈ λ− ) is marked D (resp., U ).

We refer as attacking set of TT (A) to the set Att(TT (A)) ⊆ LinesU

T containing
all the attacking lines from TT (A). Note that a tree without attacking lines is
warranting.
Theorem 6. A tree TT (A) is warranting iff Att(TT (A)) = ∅.
In order to achieve warrant for an argument A, we would need to alter the
tree TT (A) by altering every attacking line in it, with the objective of obtain-
ing a new tree free of attacking lines. The following example shows the impor-
tance of identifying the precise argument –in a line to be altered– for which a
defeater needs to be activated. Trees are drawn with gray/white triangles denot-
ing defeated/undefeated arguments.
Example 7. Consider a DAT T yielding the tree TT (A) on the right
with lines λ1 = [A, B1 , B3 ] and λ2 = [A, B2 , B4 , B5 ]. There is a
single attacking line within the attacking set Att(TT (A)) = {λ2 }.
If we activate a defeater for B2 in T, we would generate a new
line within a new tree having no attacking lines. If we instead add
a defeater for B5 , the line that was attacking is “extended” and
again, in the resulting tree, there would be no attacking lines. On
the other hand, if we activate a defeater for B4 , we would generate
another attacking line.
We call effective alteration to the alteration of an attacking line that turns
it to non-attacking. In the ATC deactivating approach, an effective alteration
is achieved by deactivating a con argument in the given line, thus truncating
the line. Thus, deactivating a con argument in a line leaves a pro argument as
the leaf of the new line. However, in the ATC activating approach presented
here, the alteration of a given attacking line is done by activating a defeater
D for a con argument B in the line. This would imply B to end up marked as
defeated since a new line will sprout activating the undefeated argument D as
its leaf. Afterwards, the resulting altered line would not be attacking. Note that,
any effective alteration of an attacking line has necessarily to be done upon a
con argument in the line. To that end, we identify the set of inactive defeaters
associated to each argument in order to choose appropriate activations.
502 M. O. Moguillansky and G. R. Simari

Definition 8 (Set of Inactive Defeaters). Given a DAT T = U, →, [A],

the set of inactive defeaters of an argument B ∈ λ, is a function idefsT :
A × LinesA
T → P(I):

idefsT (B, λ) = {D ∈ I | exists λ ∈ LinesU  ↑ ↑

T such that D ∈ λ and λ [B] = λ (D)}

Each inactive defeater D of B ∈ λ belongs to a line λ ∈ LinesU T which

means that λ is acceptable but contains inactive arguments. Requiring λ↑ (D)
to coincide with λ↑ [B] implies not only the segments from the root to B in both
λ and λ to be equal, but also D→B. The following notion allows to anticipate
the effect of changes over a DAT T by identifying a hypothetical tree HT (A, Ψ ),
as the tree rooted in A that would result from T by the hypothetical activation
of the arguments in a set Ψ ⊆ U. We refer to these trees as hypothetical given
that they do not appear within the domain TreesT .

Definition 9 (Hypothetical Tree). Given a DAT T = U, →, [A], A ∈ A,

and a set Ψ ⊆ U; the hypothetical tree HT (A, Ψ ) is the tree built from the set
of lines:
{λ↑ [B] | ∀λ ∈ LinesU ↑
T , ∀C ∈ λ [B] : C ∈ (A⊕Ψ ) holds, where root(λ) = A and
either (leaf(λ) = B) or (∃D ∈ λ : λ↑ (D) = λ↑ [B] and D ∈ / (A⊕Ψ ))}

The activation of defeaters may bring new shortcomings towards an effective

alteration of lines. The addition of a defeater D in a line λ provokes a line exten-
sion: if D attacks the leaf of λ, the whole line ends up extended; but if D attacks
an argument placed strictly above that leaf, a new argumentation line arises by
extending an upper segment of λ. The activation of D not only attaches D to λ,
but also includes the addition of D’s (active) defeaters, and these defeaters bring
their (active) defeaters, and so on. An entire subtree rooted in D sprouts from
the activation of D. This subtree could contain arguments that were already
active, as well as arguments that are activated by virtue of activeness propaga-
tion. For instance, an argumentation line λ = [A, B1 , . . . , D, . . . , Bn ] may have
all its arguments active excepting for D. Thus, a tree rooted in A would have a
line λ↑ (D). Afterwards, if we activate D, the resulting hypothetical tree would
have the complete line λ. Hence, an effective alteration in the activating app-
roach requires ensuring that the new line appearing from an alteration ends
up as non-attacking, or equivalently, ensuring that the new defeater D ends up
undefeated within the resulting tree for ensuring that the selected con argument
(attacked by D) finally ends up defeated. Moreover, the activation of a defeater
may involve activating other argument/s which may extend other argumentation
line/s in the evaluated tree. These shortcomings describes different alternatives
of what we have called collateral alterations. A fortuitous case results from the
collateral alteration of an attacking line which turns the line into non-attacking
by chance. However, it may also happen to have a non-attacking line which ends
up collaterally altered and turned into attacking, or even to have an effectively
altered line which turns into attacking once again due to some collateral alter-
ation occurring above the argument from which the original effective alteration
has been done. The following example clarifies this matter.
 Dynamics of the Judicial Process by Defeater Activation 503

Example 10. We have two attacking lines λ1 = [A, B1 , B3 , B5 ] and

λ2 = [A, B2 , B4 , B6 ] from the tree depicted on the right, warranting
A implies two effective alterations, one per attacking line. Line λ1
can be altered upon B5 by activating a defeater D1 ∈ idefsT (B5 , λ1 ).
However, assuming that C1 D1 is also inactive then A⊕{D1 } will
also activate C1 . Afterwards, if it is the case that C1 →B4 , then
the effective alteration of λ1 produces a collateral alteration
over λ2 . This situation impose a condition for achieving an effective
alteration of λ2 . For instance, in figure
(a), by activating D2 ∈ idefsT (B6 , λ2 ),
the line is intended to be altered below
a collaterality which will affect the cur-
rent alteration through the activation of
D2 , avoiding its effectiveness. The incon-
venience is that the collateral alteration
extends the segment λ↑2 [B4 ], incorporat-
ing an attacking line [A, B2 , B4 , C1 ] to the
tree, leaving ineffective the alteration of
λ2 . This shows the importance of avoid-
ing collateral alterations occurring above
the specific alteration in the line. In other words, the only safe collateral alter-
ations are those taking place below the selected con argument in the same line. In
this case, the alternative will be either to change the alteration of λ1 to avoid the
collaterality over λ2 , or to change the alteration of λ2 forcing it to occur above
the collaterality, i.e., upon the con argument B2 . The latter case is illustrated
in figure (b). Assuming D3 ∈ idefsT (B2 , λ2 ), the resulting tree ends up warrant-
ing its root argument A, given that all the attacking lines would be effectively
altered.
From the example above, if no inactive defeater for a con argument provides
an effective alteration then a different con argument should be selected in order
to obtain an appropriate inactive defeater for it, ensuring the effective alter-
ation in a different part of the line. Such dynamics allows to overcome from bad
collateralitites. The global defeating function triggers a set of inactive defeaters
which would effectively alter the necessary lines, controlling collateralities, and
ensuring a resulting warranting tree through the analysis over hypothetical trees
free of attacking lines. However, when no (appropriate) defeater is available for
any con argument in the same line, the revision operation will not succeed, and
the global defeating will trigger the empty set.
Definition 11 (Global Defeating). Given a DAT T = U, →, [A] and a
tree TT (A) ∈ TreesT ; the function σ : TreesT → P(I) is a global defeating
function iff it holds:
1. σ(TT (A)) = ∅ iff (Att(TT (A)) = ∅, or there is no non-empty global defeating).
2. if σ(TT (A)) = ∅ then HT (A, σ(TT (A))) has no attacking lines.
3. if D ∈ σ(TT (A)) then the following conditions simultaneously hold
(a) there is some attacking λ ∈ HT (A, σ(TT (A)) \ {D}),
504 M. O. Moguillansky and G. R. Simari

(b) there is some B ∈ λ− such that D ∈ idefsT (B, λ),

(c) for any D ∈ idefsT (B, λ), if D D then D = D,
(d) there is some non-attacking λ ∈ HT (A, σ(TT (A))) such that λ↑ [B] =
λ↑ (D),
(e) if the following conditions simultaneously hold
i. there is some λ ∈ HT (A, σ(TT (A))),
ii. some C ∈ λ such that C ∈ A⊕σ(TT (A)) \ {D},
iii. some D ∈ σ(TT (A)),
iv. some λ ∈ HT (A, σ(TT (A)) \ {D }),
v. some B  ∈ λ such that D ∈ idefsT (B  , λ ), and
vi. some B  ∈ λ such that λ↑ [B  ] = λ↑ (C)
then B  ∈ λ↑ (B  ).

The defeating function is the fundamental key of the global activating model,
its three conditions are necessary and sufficient for ensuring that it is impossible
(resp. of, possible) to turn the non-warranting tree into an altered warranting
one, in which case, the function triggers an empty (resp. of, non-empty) set.
On the contrary, whenever the original tree is warranting, the defeating func-
tion would also trigger an empty-set for ensuring that no unnecessary change
will be made. This is controlled by condition (1). Condition (2) states that if
the defeating function is non-empty then the activation of the arguments that
it maps would result in a tree with no attacking lines. Condition (3) controls
the behavior of the function when the revision is possible. It states that the
activation of each D ∈ σ(TT (A)) alters a specific line that otherwise would be
attacking (cond. (3a). The alteration is done over some selected con argument
B for which an inactive defeater D is available (cond. 3b), ensuring that D does
not contain any strict subargument that would serve for such a purpose (cond.
3c) –this is necessary for avoiding activations that could trigger unnecessary
collateralities. Through condition 3d, it is ensured that the alteration is effec-
tive. This is done by identifying the new line λ that appears as a result of the
alteration of λ. And finally, condition 3e takes care of collateralitites, ensuring
that when they occur, they never will take place above the selected argument
for altering the line. We will see in detail this case, by referring to Example 10,
figure (a): λ –from Definition 11– is instantiated with λ1 , D with D1 , B with B5 ,
λ with the alteration of line λ1 , i.e., with the extended line [A, B1 , B3 , B5 , D1 ]
(which ends up effectively altered satisfying cond. 3d), λ with the extended line
[A, B2 , B4 , C1 ] which appears from the collaterality that activates C (instantiated
with C1 ) from the activation of D (instantiated with D1 ) (cond. 3(e)i and 3(e)ii).
Then, D is instantiated with D2 , λ with λ2 , and B  with B6 which is the
con argument that would be defeated by D in λ (cond. 3(e)iii, 3(e)iv, and
3(e)v). Finally, cond. 3(e)vi identifies B  (instantiated with B4 ) as the argument
that was defeated by the collaterality C (instantiated with C1 ). This will allow
to verify if B  is above B  in λ , which is violated in Fig. (a) since it would
require B6 to be above B4 for ensuring that the alteration of λ2 occurs above
the collaterality produced by the alteration of λ1 . Observe that the case of Fig.
(b) satisfies the conditions of the global defeating function, which will map to a
 Dynamics of the Judicial Process by Defeater Activation 505

set {D1 , D3 }. Note that Fig. (1) depicted

on the right, shows a special case of that
presented in Example 10, Fig. (a), where
the bad collaterality is produced by the
alteration of the same line. The analy-
sis of Definition 11 can be similarly done,
where λ will coincide with λ , D with
D, λ with λ, and B  with B. On the
other hand, the case presented in Fig. (2)
shows a different anomaly: the activation
of D2 brings its own active defeater C2 to
extend the segment [A, B2 ]. In this spe-
cific case, the new line [A, B2 , D2 , C2 ] will be attacking and thus the alteration
does not change the mark of the selected con argument B2 . This situation will
violate condition 3d from Definition 11.
When selecting a con argument from a line λ ∈ TT (A), a selection criterion
could lead the mapping to an argument B ∈ λ− by assuming an ordering among
arguments. Moreover, deciding among several alternatives for activating inac-
tive defeaters would also require some specific activation criterion which, once
again, could be founded upon some ordering among arguments. Both criteria
are a necessary condition for determinism. We avoid going further into this sub-
ject, abstracting away from specific criteria. Next, we define the condition for
a warranting defeating function, as a defeating whose outcome ensures a suc-
cessful revision. This will depend on the availability of all the necessary inactive
defeaters for warranting the tree.

Definition 12 (Warranting Defeating). A defeating function “σ” is said to

be warranting for TT (A) iff (σ(TT (A)) = ∅ iff Att(TT (A)) = ∅).

The activating argument revision is defined upon a warranting defeating.

Whenever no such warranting function exists, the revision operation ends up
undefined.

Definition 13 (Argument Revision). Given a DAT T = U, →, [A] and

A ∈ U; the operator “∗” stands for an activating argument revision of T by
A iff
T∗A = U, →, [A ⊕σ(TT (A))]
where T = U, →, [A ] with A = A⊕{A}, and the global defeating function σ
is warranting for TT (A).

Example 14. Consider a DAT T = U, →, [A], where A = {B1 , B2 , B3 , B4 , B5 }

and → = {(B3 , B1 ), (B4 , B2 ), (B5 , B4 )}, and assume we need to revise T by an
inactive argument A ∈ I. The revision operator will analyze the construction
of a global defeating function upon the DAT T = U, →, [A⊕{A}], whose
resulting dialectical tree TT (A) coincides with the one depicted in Example 7.
Assuming the selection over λ2 = [A, B2 , B4 , B5 ] is the con argument B2 , and
506 M. O. Moguillansky and G. R. Simari

D1 ∈ idefsT (B2 , λ2 ); the mark of B2 turns to D.

Nonetheless, by assuming C1 ∈ A⊕{D1 } and C1 →B3 ,
line λ1 = [A, B1 , B3 ] is collaterally altered, and even
more, such collateral alteration turns the line to attack-
ing (triggering a new line λ3 = [A, B1 , B3 , C1 ]) in the
context of the hypothetical tree HT (A, {D1 }). This
situation forces the alteration of λ3 in order to turn
it to non-attacking. Recall that this will require the
selection of a con argument placed above the point of
the collaterality, i.e., an argument placed above C1 . The only alternative is
B1 . Afterwards, assuming D2 ∈ idefsT (B1 , λ3 ), we obtain a hypothetical tree
HT (A, σ(TT (A))) free of attacking lines, where σ(TT (A)) = {D1 , D2 }. Finally,
since σ is a warranting defeating for TT (A), the revision ends up as T∗A =
U, →, [A⊕{A, D1 , D2 }].
Theorem 15. If T∗A relies on a warranting defeating function then T∗A war-
rants A.

3 Dynamics of Judicial Procedures

3.1 An Example of Dynamics Through Appellate Instances
Under certain circumstances a court can let a criminal defendant off on proba-
tion1 . However, this special benefit may not be given to a defendant involved in
a gender-based violence case, since probation may end up dismissing the crim-
inal case and thus, no fair nor effective legal procedure would be granted for
the woman, contradicting multilateral treaties on the matter like the Vienna
Convention on the law of treaties (or Vienna, for short) and the Convention
of Belem do Pará on the prevention, punishment, and eradication of violence
against women (or BdP, for short). Such exceptional case was not considered by
the Argentinean Law until the Supreme Court dictated the definitive sentence
on Gongora’s case, on April 23rd, 2013 [11]. In this example, we analyze such
a sentence which revoked the prior sentence dictated by the previous appellate
instance, referred as Cámara Federal de Casación Penal, from now on, Casación.
In short, Casación’s sentence referred to the Penal Code (PC) for conceding the
benefit of probation for the case, despite the state of affairs included gender
violence evidence as shown in the lowest appellate instance. We will resume the
case, by including only the argumentation provided by the Supreme Court, with-
out the specific details of the argumentation given by Casación in the previous
appellate instance. We assume T = U, →, [A] with arguments B, B1 , B2 , B  in
A, where B1 B and B2 B, such that:
B: Probation for Gongora is ordered according to arguments B1 and B2 .
B1 : Article 76 bis (PC), grants probation for sentences of under 3 years
mandatory prison time.
1
Humanitarian effort to give second-chance to minor offenders instead of serving time
in prison.
 Dynamics of the Judicial Process by Defeater Activation 507

B2 : A mandatory sentence for Gongora’s should not be over 3 years prison

time.
B  : Evidence of gender violence.

The argumentation presented through the sentence dictated by the Supreme

Court relies mainly upon Vienna and BdP treaties, specifically:

Article 31(1) (Vienna). “A treaty shall be interpreted in good faith in

accordance with the ordinary meaning to be given to the terms of the treaty
in their context and in the light of its object and purpose.”
Article 7.b (BdP). “Apply due diligence to prevent, investigate and impose
penalties for violence against women.”
Article 7.f (BdP). “Establish fair and effective legal procedures for women
who have been subjected to violence which include, among others, protective
measures, a timely hearing and effective access to such procedures.”

The Supreme Court’s argumentation includes the following arguments:

A1 : Interpreting BdP (due to Vienna, Art. 31(1)) in the light of its objec-
tive and purpose (to eradicate violence against women); for gender violence
cases, a legal procedure should be ensured (Art. 7.f ) without delay (Art. 7.b),
discarding any possibility of granting probation.
A: According to A1 and B  , probation cannot be granted for Gongora’s.

The consideration of Supreme Court’s arguments can be understood as a

dynamic process in which the legal system evolves towards restricting probation
grant under the exceptional circumstance of gender violence. We can identify
such a dynamic situation by referring to a framework T∗A, whose intermediate
stage is defined through a framework T = U, →, [A ], where A = A⊕{A}.
Observe that as a consequence, from activeness propagation, A1 ends up con-
tained in A . Note however, that A would not be warranted given that Casación’s
sentence (argument B) would appear as a counterargument of A. At this stage,
the theory of ATC, through the evaluation of the dialectical tree TT (A) ∈ TreesT
built with the unique argumentation line λ = [A, B] (which is an attacking line),
is indicating that although argument A is a valid supporter against probation
for Gongora’s, we still need to defeat Casación’s previous sentence. Afterwards,
the unique selection will be B ∈ λ− and therefore, the theory indicates that a
defeater D, such that idefsT (B, λ) ⊇ {D} (see Definition 8), should be activated
looking for a warranting hypothetical tree HT (A, {D}) (see Definition 9). The
activation of argument D seems to be identified by the Supreme Court, which
develops a second part of the sentence, providing an appropriate justification
for their decision exposed through argument A. Recall that the activation of D
would imply activating each of its subarguments, where D D, D1 D , D2 D ,

A1 D2 , D2.1 D2 , D2.2 D2 , and D2.2 D2.2 :

D: The application of Art. 76 bis (PC) is restricted under evidence of gender

violence.
508 M. O. Moguillansky and G. R. Simari

D1 : Principles of the legal doctrine like law as integrity and coherence posse
obligations for considering the complete legal system, including international
treaties, and the ability to reinterpret the law in order to avoid contrary deci-
sions.

D2.2 : Article 75(22) (NC) incorporates the international treaties to the con-
stitutional block.

D2.2 : From D2.2 , both Vienna and BdP have constitutional hierarchy.
D2.1 : Art. 31 (NC) establishes the supremacy of the National Constitution.
D2 : From D2.1 and D2.2 , A1 is over Art. 76 bis (PC).
D : According to D1 and D2 , Art. 76 bis (PC) should be reinterpreted in
the light of A1 .

Finally, T∗A = U, →, [A⊕{A, D}] ends up warranting the Supreme
Court’s argument A since T(T∗A) (A) is a warranting tree with a unique non-
attacking argumentation line [A, B, D]. From the viewpoint of law, the case is
interesting given that argument D can be seen as an instance of creation of law
by an administrative operator of the judicial organ, which in the Argentinean
Law is an exceptional case normally reserved for the legislature. From a different
theoretical perspective, when creation of law is not admitted for judges, argu-
ment D can be seen as a case of law discovery, by assuming that the norm was
already in the system, but it simply has never been applied before. In this case,
the activation of D can be understood as a case of jurisprudence [8].

3.2 An Example of Dynamics on a Criminal Procedure

We analyze a small part of a criminal case sentenced on April, 16th 2007 by the
Tribunal en lo Criminal No. 1 (Bahı́a Blanca) [22]. This was a resonant case due
to the great astonishment it caused in society. In short, the prosecutor claims
guilt by attempting to prove that Cuchán had killed and burned his girlfriend,
Lucy, in a grill at his house, whereas the defense’s case is that the girl died
from a cocaine overdose and was afterwards burned by the man in a state of
despair. In order to make a practical example, we have extracted a small but
representative amount of allegations, i.e., arguments, and organized them within
a dialectical tree. The tree’s root is the initial argument posed by the defense,
which ends up defeated, and then we will apply ATC’s logical machinery to bring
arguments that turn the root argument into a warrant. In this way, ATC proves
to be a useful tool to perform hypothetical reasoning. We will assume a DAT
T = U, →, [A], where the set U will include the following arguments:

A: Lucy died of cocaine overdose, so Cuchán did not kill her.

B1 : Cuchán burned Lucy’s remains; the burning objective was to hide the
homicide.
B2 : According to friends, Lucy never did drugs and she was afraid of cocaine.
B3 : No traces of violence were found at the crime scene.
B4 : No records of previous incidents between Cuchán and Lucy implies no
motive for a homicide.
 Dynamics of the Judicial Process by Defeater Activation 509

B5 : An overdose might happen the first time a person consumes cocaine.

B6 : Threads of hair belonging to Lucy were found nearby the crime scene
implying violence.
B7 : Psychological and psychiatric studies discovered Cuchán’s psychopathic
features.
D1 : From the autopsy report, no cause of death could be determined.
D2 : If the cause of death cannot be determined, there is no proof for a homi-
cide.
D3 : If the cause of death cannot be determined, there is no proof for an over-
dose.
D4 : Recorded phone calls suggest Lucy’s friends were threatened to declare in
her favor.

We will assume that every argument is active, except for

D1 and D4 . Let also consider two more inactive arguments:
D and D , where D1 D and D2 D; D1 D and D3 D .
The attack relation among active arguments is depicted on
the right. Additionally, we have that D→B1 , D →B5 and
D4 →B2 . Note that, in each argumentation line, arguments
placed at odd positions are those posed by the defense,
whereas the ones in even positions correspond to the pros-
ecution. Now suppose the defense accounts with this infor-
mation –representing the current state of the trial– and
wishes to know how to proceed to change the outcome, i.e., which allegations
it should present. In order to warrant the defense’s initial argument, we need to
effectively alter the attacking lines λ1 = [A, B1 , B3 , B6 ] and λ2 = [A, B1 , B4 , B7 ].
We select B1 as the con argument in λ1 , then, a
defeater D ∈ idefsT (B1 , λ1 ) appears. The activation of D
requires to activate D1 (since D2 is already active). Note
that, D itself is interpreted as a collaterally activated
defeater for B1 but in λ2 . Thus, the global defeating func-
tion restricts the selection in λ2 to occur in the segment
λ↑2 [B1 ]. This will provoke the selection of con arguments
in both λ1 and λ2 to coincide, triggering B1 in both cases.
Hence, the activation of D turns both attacking lines into
non-attacking, as depicted on the right.
Nonetheless, the revision is not complete, since an
undesirable side-effect has occurred: the activation of
D1 has also activated argument D , which attacks B5 ,
since it dismisses B5 ’s overdose hypothesis. This col-
lateral activation is controlled by the global defeating
function in its condition 3e. Afterwards, the collateral
alteration provokes λ3 to turn into a new attacking line
λ = [A, B2 , B5 , D ] in the context of the hypotheti-
cal tree HT (A, {D}). From cond. 3a in Definition 11,
the new extended line λ is effectively altered through
510 M. O. Moguillansky and G. R. Simari

the activation of the inactive defeater D4 ∈ idefsT (B2 , λ ) for the selected
con argument B2 , turning the line into non-attacking. The figure depicted
on the right shows that A ends up warranted from the revision operation
T ∗A = U, →, [A⊕{A, D, D , D4 }].

4 Conclusions, Related and Future Work

We have presented an approach for studying dynamics of the judicial process

through an operator of argument revision which considers activation (incorpo-
ration) of arguments. This approach is comprehended within Argument Theory
Change [18,20] and provides another standpoint to change the status of warrant
of an argument in dialectical trees. The method is particularly interesting in the
context of legal reasoning given that it could bring the theoretical foundations
for recommender systems for the study of a legal case prior to the corresponding
judicial procedure. This would render different critical points in the legal argu-
mentation of the case for which further investigations could be directed towards
discovering implicit norms that have not been applied before (judge’s viewpoint),
or towards discovering new pieces of evidence (lawyer’s viewpoint).
Several recent articles (like [5,7,9,10,12,13]) deal with the dynamics of argu-
ments. Some consider change by removing attacks between arguments, others
by enriching logics with dynamic features. However, our perspective of change
pursued here is different: we provoke change upon the morphology of arguments’
graph for studying dynamics. In this sense, [17] is more similar since it studies
dynamics upon graphs. However, the main difference is that the argumentation
semantics that we follow in this article is necessarily the one derived from dialec-
tical trees, since they probably constitute the most appropriate argumentation
model for the analysis of the judicial process.
Regarding future work, the activating approach is the stepping stone towards
the definition of a hybrid revision for deciding between the activation or deacti-
vation in order to choose the most appropriate of both. This choice could apply
activation in some attacking lines and deactivation in others. From the stand-
point of the legal domain, the hybrid method would bring a theoretic model
for studying promulgation of laws. A process of promulgation may involve the
derogation of invalid norms (by disuse) that break the coherence of the legal
ordering when considering the new law. Additionally, it may be the case that
some articles of the new law should be modified for avoiding it to be declared
inconstitutional due to contradictions –for instance, with some principles of the
doctrine. A hybrid operator could model such a situation by activating argu-
ments standing for derogative norms and deactivating arguments standing for
conflicting articles of the new law. This would render a recommender system for
identifying which previous norms should be derogated and which parts of the
new law should be reformulated towards a successful process of promulgation.
 Dynamics of the Judicial Process by Defeater Activation 511

References
1. KR 2014, Vienna, Austria, 2014. AAAI Press (2014)
2. IJCAI 2015, Buenos Aires, Argentina, 2015. AAAI Press (2015)
3. Alchourrón, C., Gärdenfors, P., Makinson, D.: On the logic of theory change: partial
meet contraction and revision functions. J. Symb. Logic 50, 510–530 (1985)
4. Baroni, P., Cerutti, F., Giacomin, M., Simari, G.R. (eds.): Computational Models
of Argument. In: Proceedings of COMMA 2010. IOS Press, Amsterdam (2010)
5. Baumann, R., Brewka, G.: AGM meets abstract argumentation: expansion and
revision for dung frameworks. In: IJCAI 2015, Buenos Aires, Argentina, 2015 [2],
pp. 2734–2740
6. Bex, F., Verheij, B.: Legal shifts in the process of proof. In: Ashley, K.D., van
Engers, T.M. (eds.) ICAIL 2011, Pittsburgh, PA, USA, pp. 11–20. ACM (2011)
7. Booth, R., Kaci, S., Rienstra, T., van der Torre, L.: A logical theory about dynamics
in abstract argumentation. In: Liu, W., Subrahmanian, V.S., Wijsen, J. (eds.) SUM
2013. LNCS (LNAI), vol. 8078, pp. 148–161. Springer, Heidelberg (2013). https://
doi.org/10.1007/978-3-642-40381-1 12
8. Bulygin, E.: Sentencia Judicial y Creación de Derecho. La Ley 124, 355–369 (1966)
9. Coste-Marquis, S., Konieczny, S., Mailly, J.-G., Marquis, P.: A translation-based
approach for revision of argumentation frameworks. In: Fermé, E., Leite, J. (eds.)
JELIA 2014. LNCS (LNAI), vol. 8761, pp. 397–411. Springer, Cham (2014).
https://doi.org/10.1007/978-3-319-11558-0 28
10. Coste-Marquis, S., Konieczny, S., Mailly, J., Marquis, P.: On the revision of argu-
mentation systems: minimal change of arguments statuses. In: KR 2014, Vienna,
Austria, 2014 [1]
11. CSJN: Góngora, Gabriel Arnaldo s/causa No. 14092 (2013). http://riom.jusbaires.
gob.ar/sites/default/files/gongora csjn.pdf
12. Diller, M., Haret, A., Linsbichler, T., Rümmele, S., Woltran, S.: An extension-based
approach to belief revision in abstract argumentation. In: IJCAI 2015, Buenos
Aires, Argentina, 2015 [2], pp. 2926–2932
13. Doutre, S., Herzig, A., Perrussel, L.: A dynamic logic framework for abstract argu-
mentation. In: KR 2014, Vienna, Austria, 2014 [1]
14. Dung, P.M.: On the acceptability of arguments and its fundamental role in non-
monotonic reasoning and logic programming and n-person games. AIJ 77, 321–357
(1995)
15. Garcı́a, A.J., Simari, G.R.: Defeasible logic programming: an argumentative app-
roach. TPLP 4(1–2), 95–138 (2004)
16. Hansson, S.O.: A Textbook of Belief Dynamics: Theory Change and Database
Updating. Springer, Dordrecht (1999)
17. Moguillansky, M.O.: A study of argument acceptability dynamics through core and
remainder sets. In: Gyssens, M., Simari, G. (eds.) FoIKS 2016. LNCS, vol. 9616,
pp. 3–23. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-30024-5 1
18. Moguillansky, M.O., Rotstein, N.D., Falappa, M.A., Garcı́a, A.J., Simari, G.R.:
Argument theory change through defeater activation. In: Baroni et al. [4], pp.
359–366
19. Moguillansky, M.O., Wassermann, R., Falappa, M.A.: Inconsistent-tolerant base
revision through argument theory change. Logic J. IGPL 20(1), 154–186 (2012)
20. Rotstein, N., Moguillansky, M., Falappa, M., Garcı́a, A., Simari, G.: Argument
theory change: revision upon warrant. In: COMMA, pp. 336–347 (2008)
512 M. O. Moguillansky and G. R. Simari

21. Rotstein, N.D., Moguillansky, M.O., Garcı́a, A.J., Simari, G.R.: A dynamic argu-
mentation framework. In: Baroni et al. [4], pp. 427–438
22. SCBA: Cuchán, Pablo Victor (2007). http://www.scba.gov.ar/prensa/Noticias/17-
07-07/Mat%F3.htm
23. Wyner, A., Bench-Capon, T.: Modelling judicial context in argumentation frame-
works. In: COMMA, pp. 417–428 (2008)
 Claim Detection in Judgments of the EU
Court of Justice

Marco Lippi1(B) , Francesca Lagioia2 , Giuseppe Contissa2,3 ,

Giovanni Sartor2,3 , and Paolo Torroni4
1
DISMI, Università degli Studi di Modena e Reggio Emilia, Reggio Emilia, Italy
marco.lippi@unimore.it
2
EUI – European University Institute, Fiesole, Italy
{francesca.lagioia,giovanni.sartor}@eui.eu
3
CIRSFID, Università degli Studi di Bologna, Bologna, Italy
giuseppe.contissa@unibo.it
4
DISI, Università degli Studi di Bologna, Bologna, Italy
paolo.torroni@unibo.it

Abstract. Mining arguments from text has recently become a hot

topic in Artificial Intelligence. The legal domain offers an ideal scenario
to apply novel techniques coming from machine learning and natural
language processing, addressing this challenging task. Following recent
approaches to argumentation mining in juridical documents, this paper
presents two distinct contributions. The first one is a novel annotated
corpus for argumentation mining in the legal domain, together with a
set of annotation guidelines. The second one is the empirical evaluation
of a recent machine learning method for claim detection in judgments.
The method, which is based on Tree Kernels, has been applied to context-
independent claim detection in other genres such as Wikipedia articles
and essays. Here we show that this method also provides a useful instru-
ment in the legal domain, especially when used in combination with
domain-specific information.

Keywords: Claim detection · Argumentation mining

Legal arguments

1 Introduction
One of the most traditional yet lively research sub-areas at the intersection
of Artificial Intelligence and Law is the study of argumentation in the legal
context [5,6]. Argumentation is a wide research field that spans across several
different areas, having its roots in logic, philosophy, and linguistics, as it basically
studies how different theses and opinions are proposed, debated and evaluated,
taking into account their relations and inter-dependencies. The legal domain
This work was done while Marco Lippi was at DISI – University of Bologna and
Francesca Lagioia was at CIRSFID – University of Bologna.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 513–527, 2018.
https://doi.org/10.1007/978-3-030-00178-0_35
514 M. Lippi et al.

thus offers a natural scenario for the application of different argument models,
in order to perform legal reasoning [26], to build specific ontologies [2], or to
support the teaching of jurisprudence [3,10].
From the Artificial Intelligence viewpoint, many contributions have been
made in the context of building computational and logic models for legal argu-
ments [26,27], in case-based reasoning [1,7], in the full semantic interpretation
of judicial opinions [20], in yielding the syntactic structure of sentences using
a rule-based parser in order to detect legal modifications [8], and also in the
automatic extraction of arguments (or part thereof) from legal documents, as
an application of the recent discipline of argumentation mining [21,22]. Some
recent works [36] are focused on the discovery and analysis of the internal struc-
ture of an arguments, the identification of its premises and the conclusions, and
the internal syntactical and grammatical structure of each statement.
Building tools capable of automatically detecting arguments in legal texts
would produce a dramatic impact on many disciplines related to Law, providing
invaluable instruments for the retrieval of legal arguments from large corpora,
for the summarization and classification of legal texts, and finally for the devel-
opment of expert systems supporting lawyers and judges. Mochales Palau and
Moens [22] gave an influential contribution in this domain, providing the first
system for mining arguments from legal documents. Their system, specifically
designed by experts in the legal domain, combines highly engineered feature con-
struction, machine learning approaches, and a hand-crafted context-free gram-
mar to infer links between arguments. Their results are yet hard to reproduce,
since the dataset they used, made up of documents extracted from legal texts
of the European Court of Human Rights (ECHR) and from the AraucariaDB,
is currently not available, whereas their methodology exploits plenty of context-
dependent information that was specifically extracted from that corpus.
In this work, we aim to contribute to the budding field of argumentation
mining in legal documents by moving in two different directions. First, we present
a novel, freely available, annotated corpus for argumentation mining in the legal
context, accompanied by a set of guidelines that have been followed during the
document labeling process. Second, we consider a machine learning approach for
the extraction of claims from legal documents that has recently been applied to
context-independent argument mining [18], and we evaluate it in this new genre.
Our preliminary results show that context-independent claim detection is also
helpful in the legal domain, thus providing a powerful framework that can be
used in combination with domain-specific information.

2 Background
Argumentation mining is concerned with the automatic extraction of arguments
from generic textual corpora. This has a self-evident application potential in
a variety of domains. IBM recently funded a multi-million cognitive comput-
ing project called Debater, whose core technology is argumentation mining, and
which aims to retrieve pro and con arguments concerning a given controversial
 Claim Detection in Judgments of the EU Court of Justice 515

topic.1 If we consider user-generated content available on the Web, argumen-

tation mining can be seen as the evolution of sentiment analysis: its goal is to
detect not only opinions, but also the reasons behind them [15].
One of the pioneering applications of argumentation mining was in the legal
domain. Building on Teufel’s work on argumentation zoning [30], Hachey and
Grover proposed a system for legal document summarization [16]. Mochales
Palau and Moens [22] proposed the first argumentation mining system, focussing
on the extraction of claims and their supporting premises from a collection of
structured legal documents. They worked on two datasets: the European Court
of Human Rights (ECHR) [21] and AraucariaDB, a collection maintained by the
University of Dundee.2 More recently, the Vaccine/Injury Project (V/IP) [4] was
carried out, with the goal of extracting arguments from a set of judicial decisions
involving vaccine regulations.
Although the general idea of argumentation mining is clear, a precise defi-
nition of the problem is difficult to obtain, because the problem is complex in
itself (there are many sub-tasks that can be identified), and the very notion of
argument is still a matter for discussion, and is somehow genre-dependent. An
argument in law is different from an argument in an online discussion, although
they do share commonalities.
A simple and intuitive characterization of an argument is given by Walton
as a set of statements consisting of three parts: a conclusion, a set of premises,
and an inference from the premises to the conclusion [33]. Aside from this basic
premise/conclusion argument model, there are other noteworthy models are due
to Toulmin [31] and Freeman [14]. A rather comprehensive account of argumen-
tation models from an argument analysis perspective is given by Peldszus and
Stede [25]. To add to the terminological complexity, in the literature conclu-
sions are sometimes referred to as claims, premises are often called evidence or
reasons, and the link between the two, i.e., the inference, is sometimes called
the argument itself. The task of detecting the premises and conclusion of an
argument, as found in a text of discourse, is typically referred to as detection
or identification [33]. More specific sub-tasks are claim detection and evidence
detection [17,18,28].
Even the targets of argumentation mining vary widely. Some research aims
at extracting the arguments from generic unstructured documents, which is a
fundamental step in practical applications [17], whereas other research starts
from a given set of arguments and focuses on aspects such as the identifica-
tion of attack/support [11] or entailment [9] relations between them, or on the
classification of argument schemes [13] in the sense of Walton et al. [34]. The
aforementioned approach by Mochales Palau and Moens [22] represents, to date,
one of the few works whose goal was to implement a full-fledged argumentation
mining system, albeit specific to a single genre.

1
More about IBM Debating Technologies at http://researcher.watson.ibm.com/
researcher/view group.php?id=5443.
2
http://corpora.aifdb.org/.
516 M. Lippi et al.

Diverse methodologies have been developed which involve aspects of natural

language processing and understanding, information extraction, feature discov-
ery and discourse analysis. In general, all the argument mining frameworks pro-
posed so far can be described as multi-stage pipeline systems [19], whose input
consists of natural, free text documents, and whose output is a marked-up doc-
ument, where arguments (or parts of arguments, such as claims) are marked up.
Each stage addresses a sub-task of the whole argumentation mining problem,
by employing one or more machine learning and natural language processing
methodologies and techniques.
A first stage usually consists in detecting which sentences in the input doc-
ument are argumentative, which means that they contain an argument, or part
thereof. Once argumentative sentences are singled out, one needs to detect the
boundaries between the various argument components. Finally, a last stage in
the pipeline considers these components in order to predict links between them
and/or between the arguments they are part of. For an overview of the state
of the art in argumentation mining, the reader can refer to [19]. In this work,
we focus on the first stage of the pipeline, considering concluding claims as the
target. By the concluding claim of an argument we mean its conclusion, namely,
the proposition that is affirmed on the basis of the argument’s premises. The
concluding claims affirmed in the opinion conclusions are crucial for its under-
standing, as they highlight the core legal grounds supporting the judgment.
We choose to focus on judicial opinions since they are highly structured legal
texts, in which an account of the arguments by different parties is first presented,
and then the justification for the judge’s decision is provided. Both aspects are
extensively developed in the opinions of the European judges.

3 Corpus
The source corpus consists of fifteen relevant European Court of Justice (ECJ)
decisions from 2001 to 2014 extracted from the EUR-LEX database, all related to
data protection. These documents were manually labeled following a procedure
that we describe in detail in the following subsections. These annotations will
represent the ground truth for our claim detection system.

3.1 Data Collection

This source was selected because: (a) ECJ decisions contain different types of
legal arguments by different actors (e.g. arguments appealing to statues, princi-
ples or precedents, according to different interpretive canons, e.g. arguments by
the parties, the Advocate General or the judges); (b) ECJ decisions have a stan-
dard (although not fixed) structure, in which the complex and highly variable
structure of arguments is embedded; (c) the selected decisions come from the
same domain (data protection), in which the annotators have some expertise.
ECJ decision usually include the following sections:
1. preamble: information on the parties and the main object of the judgment;
2. legal context: listing of all the legal instruments used in the judgment;
 Claim Detection in Judgments of the EU Court of Justice 517

3. background of the case: the procedural history of the case and the question
referred to the court;
4. consideration on the question(s) referred: the observations submitted to the
court by the parties and other actors such as the Governments of Member
States, plus the responses of the Court;
5. costs: the attribution of costs;
6. ruling: the final decision and the orders to the parties.

3.2 Annotation Guidelines

In analyzing the ECJ decisions, we did not consider Sects. 1 (preamble), 2 (legal
context) and 3 (background of the case), because they contain only legal and fac-
tual information, but no arguments are put forward. The most interesting part
for our aims is Sect. 4 (consideration on the question(s) referred), which contains
all argumentative steps leading to the final ruling. Section 5 (costs) was taken
in to account. We did not consider Sect. 6 (ruling), since it usually repeats the
top claims of Sect. 4, completed with orders to the parties. The text is divided
into numbered paragraphs. In selecting arguments we proceeded as follows: for
each paragraph, if arguments were present, we considered the chaining of argu-
ments [32], identified the top-level argument, that is, the ultimate argument in
the chain, and we annotated the claim corresponding to its conclusion, as well
as the keywords signaling or introducing such argument. Highlighting keywords
and markers in the text was useful for the purpose of keeping the annotations
uniform.
In order to detect an argument we first considered the grammatical and syn-
tactical structure of the text, looking for occurrences of conclusion indicators [12]
such as “as a result”, “therefore”, “consequently”, “thus”, “for this reason”.
Nevertheless, sometimes the grammatical and the syntactical structures were
not sufficient to detect arguments, and it was necessary to take into account the
semantics and the legal context. For instance, consider the following statements
taken from judgment C-301/06, paragraphs 28 and 38, respectively:

Ireland submits that the choice of Article 95 EC as the legal basis

for Directive 2006/24 is a fundamental error.

Article 4 of Directive 2006/24 provides that the conditions for access to and
processing of retained data must be defined by the Member States subject
to the legal provisions of the Union and international law.

We can say that the first statement introduces a claim that can be evaluated
as the conclusion of an argument, while the second simply repeats the content
of a legal provision and it is not part of an argumentative claim.
The contextual analysis also helped us to distinguish two uses of precedents
and other legal sources: (1) an argumentative use, where the court refers to a
precedent or source in order to reinforce and bolster its arguments supporting the
decision; (2) a non-argumentative way, in which the court incidentally mentions
518 M. Lippi et al.

a previous decision or an argument of other parties, as a contextual element.

Only in the first case did we annotate the sentence as containing a claim. As
an example, consider the following statements of the Court from judgment C-
275/06, paragraph 36 and 62:

It should be recalled that it is solely for the national court before

which the dispute has been brought, and which must assume
responsibility for the subsequent judicial decision, to determine
in the light of the particular circumstances of the case both the
need for a preliminary ruling in order to enable it to deliver
judgment and the relevance of the questions which it submits
to the Court(Case C 217/05 Confederación Española de Empresarios de
Estaciones de Servicio [2006] ECR I 11987, paragraph 16 and the case-law
cited).

It should be recalled that the fundamental right to property and the fun-
damental right to effective judicial protection constitute general principles
of Community law (see respectively, to that effect, Joined Cases C 154/04
and C 155/04Alliance for Natural Health and Others [2005] ECR I 6451,
paragraph 126 and the case-law cited, and Case C 432/05 Unibet [2007]
ECR I 2271, paragraph 37 and the case-law cited).

This first statement is an example of how the Court refers to a previous

ruling to strengthen its argument, while the second one refers to a contextual
element.
When a paragraph includes two or more top-level arguments, we highlighted
the conclusions of both arguments and also marked the keywords or the expres-
sions linking the arguments, such as “moreover”, “furthermore”, “additionally”,
“it should be added that”, etc. As an example, consider the following two state-
ments taken from judgment C-131/12, paragraph 22 in which we marked both
the top-level claims, highlighted in bold in the text:

According to Google Spain and Google Inc., the activity of search

engines cannot be regarded as processing of the data which
appear on third parties web pages displayed in the list of search
results.

Furthermore, even if that activity must be classified as “data processing”,

the operator of a search engine cannot be regarded as a “con-
troller” in respect of that processing since it has no knowledge
of those data and does not exercise control over the data.

When contiguous arguments are chained together, so that the preceeding

ones are only meant to provide the premises for the last one, we have marked
only the last argument (highlighted in bold in the text). Look at the following
example:
 Claim Detection in Judgments of the EU Court of Justice 519

As regards Article 12(b) of Directive 95/46, the application of which is

subject to the condition that the processing of personal data be incompatible
with the directive, it should be recalled that, as has been noted in paragraph
72 of the present judgment, such incompatibility may result not only from
the fact that such data are inaccurate but, in particular, also from the fact
that they are inadequate, irrelevant or excessive in relation to the purposes
of the processing, that they are not kept up to date, or that they are kept for
longer than is necessary unless they are required to be kept for historical,
statistical or scientific purposes.

It follows from those requirements, laid down in Article 6(1)(c) to (e) of

Directive 95/46, that even initially lawful processing of accurate data may,
in the course of time, become incompatible with the directive where those
data are no longer necessary in the light of the purposes for which they
were collected or processed. That is so in particular where they appear to
be inadequate, irrelevant or no longer relevant, or excessive in relation to
those purposes and in the light of the time that has elapsed.

Therefore, if it is found, following a request by the data subject pursuant

to Article 12(b) of Directive 95/46, that the inclusion in the list of results
displayed following a search made on the basis of his name of the links to
web pages published lawfully by third parties and containing true informa-
tion relating to him personally is, at this point in time, incompatible with
Article 6(1)(c) to (e) of the directive because that information appears,
having regard to all the circumstances of the case, to be inadequate, irrel-
evant or no longer relevant, or excessive in relation to the purposes of the
processing at issue carried out by the operator of the search engine, the
information and links concerned in the list of results must be
erased.

When a party in the case mentions an argument by another to endorse that

argument, we considered the conclusion of the argument as a claim by the party.
As an example, look at the structure of the following statement, taken from
judgment C-101/01, paragraph 31:

The Swedish Government submits that, when Directive 95/46 was imple-
mented in national law, the Swedish legislature took the view that pro-
cessing of personal data by a natural person which consisted in
publishing those data to an indeterminate number of people, for
example through the internet, could not be described as “a purely
personal or household activity” within the meaning of the second
indent of Article 3(2) of Directive 95/46. However, that Government
does not rule out that the exception provided for in the first indent of that
paragraph might cover cases in which a natural person publishes personal
data on an internet page solely in the exercise of his freedom of expression
and without any connection with a professional or commercial activity.
520 M. Lippi et al.

In this example, the claim put forward by the Swedish legislature, introduced
by the expression “took the view that”, is endorsed by the Swedish Government,
an endorsement expressed by the locution “submits that”.
The annotated corpus is available at http://argumentationmining.disi.unibo.
it/aicol2015.html.

4 Methods
In this work, we focus on the first stage of the pipeline sketched in Sect. 2, i.e.,
on argumentative sentence classification, and in particular on claim detection.
Our goal is thus to detect sentences that contain a claim.
The most common approaches to this task typically employ machine learn-
ing systems, whose aim is to construct a classifier that is capable of associating
a given sentence x with a label y that indicates whether or not the sentence
contains a claim. There is a wide variety of methods for building such a clas-
sifier. They differ by their chosen machine learning algorithm and by how they
represent sentences. Some techniques simply represent the sentence with the
well-known bag-of-words model, in which a sentence x is just represented by the
set of its words, regardless of their order, encoded into a linear vector. Advanced
variants of that model also consider bigrams and trigrams of words. The most
common existing approaches to claim detection rely on large sets of sophisticated
features, that are very often domain-dependent and designed by hand to address
the task of interest. While simple machine learning algorithms are typically used
as off-the-shelf tools [19], a lot of effort is dedicated in these approaches to the
development of such highly engineered features. This is the case, for example,
in the work by Mochales Palau and Moens [22] on judicial decision, the works
by the IBM Haifa research team in the context of the Debater project [17,28],
and the approach presented by Stab and Gurevych on persuasive essays [29].
Such works use, for example, the following inputs: pre-determined lists of spe-
cial keywords that are usually highly indicative of the presence of an argument;
the output of external classifiers that compute the sentiment or the subjectivity
score of a sentence; semantic information coming from thesauri and ontologies
like WordNet.
A recent work by Lippi and Torroni [18] has shown that the structure of a
sentence is very often highly informative on the presence of argumentative com-
ponents, such as claims. The key idea is that information coming from natural
language processing, as in the case of parse trees, can be employed to measure
similarity between sentences, and thus to detect fragments and structures that
typically encode claims. As an example, consider the following two sentences:
The Netherlands Government submits that Directive 95/46 does not
preclude Member States from providing for greater protection in
certain areas.
The Parliament also argues that reliance on Article 95 EC as the
legal basis is not invalidated by the importance attributed to com-
bating crime.
 Claim Detection in Judgments of the EU Court of Justice 521

NP VP .

DT NNP NNP VBZ SBAR .

The Netherlands Government submits IN S

that NP VP

NNP CD VBZ RB VP

Directive 95/46 does not VB NP PP

preclude NNP NNPS IN S

Member States from VP

VBG PP

providing IN NP

for NP PP

JJR NN IN NP

greater protection in JJ NNS

certain areas

NP ADVP VP .

DT NNP RB VBZ SBAR .

The Parliament also argues IN S

that NP VP

NP PP VBZ RB ADJP

NN IN NP is not JJ PP

reliance on NP PP invalidated IN NP

NNP NNP NNP IN NP by NP VP

Article 95 EC as DT JJ NN DT NN VBN PP

the legal basis the importance attributed TO S

to VP

VBG NP

combating NN

crime

Fig. 1. Constituency trees for two sentences containing claims. Boxed nodes highlight
the common structure of a subordinate introduced by a third-person verb (VBZ) and
the that preposition (IN). Examples are taken from judgments collected in our corpus

The first sentence is taken from judgment C-101/01 (paragraph 93), while the
second one is taken from judgment C-301/06 (paragraph 37) and they have both
been labeled as containing a claim (highlighted in bold) in our corpus. The parse
trees for these sentences are shown in Fig. 1, where boxed nodes highlight the
common structures, in this case consisting of a subordinate introduced by a third-
person verb (VBZ) and the preposition “that”. The two verbs introducing such
subordinates (submit, argue) are also indicative of the presence of an argument.
Other patterns are frequently observed in sentences containing claims.
The structure of a sentence is thus highly indicative of the presence of a
claim, and constituency parse trees represent a very powerful instrument to
capture such information. Based on this observation, Lippi and Torroni [18]
built a claim detection system that employs a Support Vector Machine (SVM)
classifier capturing similarities between parse trees through Tree Kernels [23].
522 M. Lippi et al.

The Tree Kernel approach has been shown to outperform competitors exploiting
classic, handcrafted features, widely used in NLP, such as bag-of-words, bigrams,
trigrams, part-of-speech tags and lemmas, while achieving results comparable to
highly sophisticated systems, specifically designed for context-dependent claim
detection.
Kernel methods, and in particular Tree Kernels, have a quite long tradi-
tion in natural language processing applications, including relation extraction,
named entity recognition, or question classification [24]. A kernel machine clas-
sifier learns a function f : X → Y where X is the input space, usually a vector
space representing features, and Y is the output space representing the set of
labels, or categories, to be distinguished (in our case, claim vs. other). To learn
function f , a loss function is minimized over a set of N given observations, which
is a dataset D = {(xi , yi )}N
i=1 . Examples xi ∈ X are not necessarily represented
by vectors of features, but they can also exploit structured data, in order to
encode relational information, as it happens with trees or graphs. A Tree Kernel
(TK) can be basically thought of as a similarity measure between two trees,
that evaluates the number of their common substructures, sometimes also called
fragments. According to the definition of fragments, different TK functions can
be constructed.
For example, one could consider only complete subtrees as allowed fragments,
as well as define more complex fragment structures. Intuitively, each possible tree
fragment is associated with a different feature in a high-dimensional vectorial
space, where the j-th feature simply counts the number of occurrences of the j-
th tree fragment: the TK can therefore be computed as the dot product between
two such representations of different trees. A kernel machine is then defined,
which exploits the structured information encoded by the tree kernel function
K(x, z):
N N

f (x) = αi yi φ(xi ) · φ(x) = αi yi K(xi , x) (1)
i=1 i=1

where φ is the feature mapping induced by the tree kernel K, and N is the
number of support vectors. In general, the kernel between two trees Tx and Tz
can be computed as:
 
K(Tx , Tz ) = Δ(nx , nz ) (2)
nx ∈NTx nz ∈NTz

where NTx and NTz are the set of nodes of the two trees, and Δ(·, ·) measures the
score between two nodes, according to the definition of the considered fragments.
In this work we consider the Partial Tree Kernel (PTK) [23], which allows the
most general set of fragments (called Partial Trees), being any possible portion
of subtree at the considered node. The higher the number of common fragments,
the higher the score Δ between two nodes.
The representation power behind tree kernels is evident. Basically, a kernel-
like PTK is capable of automatically generating a very rich feature set, that
captures structured representations without the need of a costly, hand-crafted
 Claim Detection in Judgments of the EU Court of Justice 523

feature engineering process. Nevertheless, it is very interesting to notice that

the TK framework allows us to include in the representation of each example
also a plain vector of features, which can enrich the description of the considered
instance given by the parse tree. In this case, the final kernel would be computed
as the combination between a classic kernel between feature vectors KV and
the kernel between trees KT , e.g., with a weighted sum or product of the two
contributions. Also note that the use of a TK by itself does not take into account
information about the context of the considered documents, which, in our case,
is the legal domain, nor the whole structure of the document. For this reason,
context-dependent information could be appended to the feature vector to be
combined with the TK, thus exploiting both context-dependent and context-
independent information, whereas relational learning algorithms could be applied
in order to capture the relations across different sentences.

5 Results
For our experiments, we employed the new corpus of judgments of the ECJ
related to data protection described in Sect. 3. We employed a leave-one-out
procedure (LOO), as customary in machine learning experimental evaluation.
The procedure dictates that in turn, each judgment be considered as a test case,
while all the other documents constitute the training and validation sets. With
N documents, training and testing are thus independently performed N times,
and results are finally averaged across all the runs.
We used the Stanford CoreNLP suite3 both to split each document into sen-
tences, and to compute the parse tree for each sentence. We obtained in this way
a total of 1,096 sentences, of which 435 were labeled as positive (i.e., containing a
claim) and the remaining 661 as negative (i.e., not containing any claim). Aside
from computing the parse trees, we also extracted from each sentence a vector
of features, that have been extensively used in a variety of Natural Language
Processing applications: bag-of-words, bag-of-bigrams and bag-of-trigrams for
words, stems and part-of-speech tags.
We thus trained three distinct classifiers: (1) an SVM based on the PTK
described in Sect. 4 as the kernel computed over the parse trees (we exploited a
combination of both stemmed and not-stemmed constituency parse trees); (2)
an SVM trained with a linear kernel over the vector of features only; (3) an SVM
combining (summing) the PTK and the linear kernel over the feature vector. We
will refer to these three classifiers as PTK, FV, and PTK+FV. For all classifiers,
we selected the SVM regularization parameter C by employing three documents
as a validation set. We relied on default values for the other PTK parameters.
Table 1 shows the results obtained by the three classifiers with the LOO pro-
cedure, macro-averaged on the 15 test documents, and also the results obtained
with a random baseline classifier. Since our problem is a binary classification task
(with two classes only), we define as True Positives (TP) the correctly detected
elements of the positive class, as False Positives (FP) the negative examples that
3
http://nlp.stanford.edu/software/corenlp.shtml.
524 M. Lippi et al.

Table 1. Results obtained on our ECJ corpus, macro-averaged over the 15 documents.

Classifier P R F1
Random 39.5 39.5 39.5
FV 40.0 63.1 47.9
PTK 39.7 80.5 52.4
PTK + FV 44.3 78.3 55.4

are wrongly classified as positives, and as False Negatives (FN) the positive cases
that are not retrieved. Standard classification measurements for these kind of
tasks include Precision (P = T PT+F P TP 2P R
P ), Recall (R = T P +F N ), and F1 = P +R ,
i.e., the harmonic mean between Precision and Recall.
The results show that all approaches perform much better than a random
predictor, even with a relatively small training set. It is interesting to note that
the combination of PTK and FV achieves the best performance, thus indicat-
ing that the information exploited by the two distinct approaches is somehow
complementary, and that PTK could be conveniently used also in combination
with more context-dependent information. In order to assess the statistical sig-
nificance of these results, we run a Wilcoxon paired test [35] on the F1 values
obtained on each document, which produced a p-value < 0.01 for the PTK+FV
classifier with respect to FV.
Finally, consider also that the approaches used here do not take into account
the whole document structure, which is instead a crucial piece of knowledge for
retrieving the concluding claims, as explained by the guidelines illustrated in
Sect. 3. Therefore, it is clear that there is a large margin of improvement for the
considered task, especially if including in the model contextual and relational
information.

6 Conclusion
The ECJ decisions related to data protection are a small-sized but novel anno-
tated corpus for argumentation mining in the legal domain. We are actively
working on its extension. Nevertheless, we hope that it can represent a use-
ful benchmark for future work in this domain. We have demonstrated that a
context-independent methods such as the Tree Kernel-based classifier proposed
in [18] could be a valuable asset for claim detection in this genre, especially when
used in combination with domain-specific information. We are aware that this
is only a first step, and there is certainly room for improvement. In the future,
we plan to expand our corpus, and to extend the analysis to the labeling and
prediction of premises, the labeling and prediction of support/attack links, and
the extraction of argument maps directly from text. One important remark is
that labeling was made using a considerable amount of information from the
document and discourse structure. Not only ECJ decisions are structured in
well-defined sections, only some of which contain argumentative content, but
 Claim Detection in Judgments of the EU Court of Justice 525

the structure itself of the argumentation within each section was analyzed and
captured by the used labeling (top-level claims, embedded arguments, arguments
referring to other arguments, such as strengthening arguments or repeated argu-
ments, etc.). Often, the labeler is able to correctly identify a concluding claim
thanks to the structure of the argumentation. By contrast, our classifier—which
is nevertheless able to provide acceptable results—does not use this structure.
This opens an avenue for further work that should go in the direction of exploit-
ing the discourse structure. Statistical relational learning could thus represent
a perfectly suitable framework for exploiting relational information across sen-
tences. Another key contribution could also come from deep learning, which has
recently achieved breakthrough results in a variety of tasks related to natural
language processing.
In future work, we will also compare our approach with other classifiers and in
particular with a simple rule-based one, that implements basic pattern recogni-
tion rules. We expect that our system and the classifier using pattern recognition
rules will deliver similar outcomes with regard to small source corpora, but that
our approach will deliver better results when applied to larger sets of documents.

References
1. Aleven, V.: Using background knowledge in case-based legal reasoning: a computa-
tional model and an intelligent learning environment. Artif. Intell. 150(1), 183–237
(2003)
2. Alexander, B.: LKIF core: principled ontology development for the legal domain.
In: Law, Ontologies and the Semantic Web: Channelling the Legal Information
Flood, vol. 188, p. 21 (2009)
3. Ashley, K.D., Desai, R., Levine, J.M.: Teaching case-based argumentation concepts
using dialectic arguments vs. didactic explanations. In: Cerri, S.A., Gouardères, G.,
Paraguaçu, F. (eds.) ITS 2002. LNCS, vol. 2363, pp. 585–595. Springer, Heidelberg
(2002). https://doi.org/10.1007/3-540-47987-2 60
4. Ashley, K.D., Walker, V.R.: Toward constructing evidence-based legal arguments
using legal decision documents and machine learning. In: Francesconi, E., Verheij,
B. (eds.) ICAIL 2012, Rome, Italy, pp. 176–180. ACM (2013)
5. Bench-Capon, T., Freeman, J.B., Hohmann, H., Prakken, H.: Computational mod-
els, argumentation theories and legal practice. In: Machines, A. (ed.) Reed C, Nor-
man TJ. Argumentation Library, vol. 9, pp. 85–120. Springer, Dordrecht (2003).
https://doi.org/10.1007/978-94-017-0431-1 4
6. Bench-Capon, T., Prakken, H., Sartor, G.: Argumentation in legal reasoning. In:
Simari, G., Rahwan, I. (eds.) Argumentation in Artificial Intelligence, pp. 363–382.
Springer, Boston (2009). https://doi.org/10.1007/978-0-387-98197-0 18
7. Bench-Capon, T., Sartor, G.: A model of legal reasoning with cases incorporating
theories and values. Artif. Intell. 150(1), 97–143 (2003)
8. Brighi, R., Lesmo, L., Mazzei, A., Palmirani, M., Radicioni, D.P.: Towards semantic
interpretation of legal modifications through deep syntactic analysis. In: Proceed-
ings of the 2008 conference on Legal Knowledge and Information Systems: JURIX
2008: The Twenty-First Annual Conference, pp. 202–206. IOS Press (2008)
9. Cabrio, E., Villata, S.: A natural language bipolar argumentation approach to
support users in online debate interactions. Argum. Comput. 4(3), 209–230 (2013)
526 M. Lippi et al.

10. Carr, C.S.: Using computer supported argument visualization to teach legal argu-
mentation. In: Kirschner, P.A., Buckingham Shum, S.J., Carr, C.S. (eds.) Visualiz-
ing Argumentation. Computer Supported Cooperative Work, pp. 75–96. Springer,
London (2003). https://doi.org/10.1007/978-1-4471-0037-9 4
11. Chesñevar, C.I., et al.: Towards an argument interchange format. Knowl. Eng. Rev.
21(4), 293–316 (2006)
12. Copi, I.M., Cohen, C., McMahon, K.: Introduction to Logic: Pearson New Inter-
national Edition. Pearson Higher Education (2013)
13. Feng, V.W., Hirst, G.: Classifying arguments by scheme. In: Lin, D., Matsumoto,
Y., Mihalcea, R. (eds.) The 49th Annual Meeting of the Association for Computa-
tional Linguistics: Human Language Technologies, Proceedings of the Conference,
Portland, Oregon, USA, 19–24 June 2011, pp. 987–996. ACL (2011)
14. Freeman, J.B.: Dialectics and the Macrostructure of Arguments: A Theory of Argu-
ment Structure, vol. 10. Walter de Gruyter (1991)
15. Habernal, I., Eckle-Kohler, J., Gurevych, I.: Argumentation mining on the web
from information seeking perspective. In: Cabrio, E., Villata, S., Wyner, A. (eds.)
Proceedings of the Workshop on Frontiers and Connections Between Argumen-
tation Theory and Natural Language Processing. Forlı̀-Cesena, Italy, 21–25 July
2014. CEUR Workshop Proceedings, vol. 1341. CEUR-WS.org (2014)
16. Hachey, B., Grover, C.: Extractive summarisation of legal texts. Artif. Intell. Law
14(4), 305–345 (2006)
17. Levy, R., Bilu, Y., Hershcovich, D., Aharoni, E., Slonim, N.: Context dependent
claim detection. In: Hajic, J., Tsujii, J. (eds.) COLING 2014, Dublin, Ireland, pp.
1489–1500. ACL (2014)
18. Lippi, M., Torroni, P.: Context-independent claim detection for argument min-
ing. In: Yang, Q., Wooldridge, M. (eds.) Proceedings of the Twenty-Fourth Inter-
national Joint Conference on Artificial Intelligence, IJCAI 2015, Buenos Aires,
Argentina, 25–31 July 2015, pp. 185–191. AAAI Press (2015)
19. Lippi, M., Torroni, P.: Argumentation mining: state of the art and emerging trends.
ACM Trans. Internet Technol. 16(2), 10:1–10:25 (2016)
20. McCarty, L.T.: Deep semantic interpretations of legal texts. In: Proceedings of
the 11th International Conference on Artificial Intelligence and Law, pp. 217–224.
ACM (2007)
21. Mochales Palau, R., Ieven, A.: Creating an argumentation corpus: do theories
apply to real arguments? A case study on the legal argumentation of the ECHR.
In: Proceedings of the Twelfth International Conference on Artificial Intelligence
and Law (ICAIL 2009), Barcelona, Spain, 8–12 June 2009, pp. 21–30. ACM (2009)
22. Mochales Palau, R., Moens, M.F.: Argumentation mining. Artif. Intell. Law 19(1),
1–22 (2011)
23. Moschitti, A.: Efficient convolution kernels for dependency and constituent syn-
tactic trees. In: Fürnkranz, J., Scheffer, T., Spiliopoulou, M. (eds.) ECML 2006.
LNCS (LNAI), vol. 4212, pp. 318–329. Springer, Heidelberg (2006). https://doi.
org/10.1007/11871842 32
24. Moschitti, A.: State-of-the-art kernels for natural language processing. In: Tutorial
Abstracts of ACL 2012, ACL 2012, p. 2. Association for Computational Linguistics,
Stroudsburg (2012)
25. Peldszus, A., Stede, M.: From argument diagrams to argumentation mining in
texts: a survey. Int. J. Cogn. Inf. Nat. Intell. (IJCINI) 7(1), 1–31 (2013)
 Claim Detection in Judgments of the EU Court of Justice 527

26. Prakken, H., Sartor, G.: A dialectical model of assessing conflicting arguments in
legal reasoning. In: Prakken, H., Sartor, G. (eds.) Logical Models of Legal Argu-
mentation, pp. 175–211. Springer, Dordrecht (1997). https://doi.org/10.1007/978-
94-011-5668-4 6
27. Prakken, H., Sartor, G.: The role of logic in computational models of legal argu-
ment: a critical survey. In: Kakas, A.C., Sadri, F. (eds.) Computational Logic:
Logic Programming and Beyond. LNCS (LNAI), vol. 2408, pp. 342–381. Springer,
Heidelberg (2002). https://doi.org/10.1007/3-540-45632-5 14
28. Rinott, R., Dankin, L., Perez, C.A., Khapra, M.M., Aharoni, E., Slonim, N.: Show
me your evidence - an automatic method for context dependent evidence detec-
tion. In: Màrquez, L., Callison-Burch, C., Su, J., Pighin, D., Marton, Y. (eds.)
Proceedings of the 2015 Conference on Empirical Methods in Natural Language
Processing, EMNLP 2015, Lisbon, Portugal, 17–21 September 2015, pp. 440–450.
The Association for Computational Linguistics (2015)
29. Stab, C., Gurevych, I.: Identifying argumentative discourse structures in persuasive
essays. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) EMNLP 2014, Doha,
Qatar, pp. 46–56. ACL (2014)
30. Teufel, S.: Argumentative zoning. Ph.D. Thesis, University of Edinburgh (1999)
31. Toulmin, S.E.: The Uses of Argument. Cambridge University Press, Cambridge
(1958)
32. Walton, D.: Fundamentals of Critical Argumentation. Critical Reasoning and
Argumentation. Cambridge University Press, Cambridge (2006)
33. Walton, D.: Argumentation theory: a very short introduction. In: Simari, G., Rah-
wan, I. (eds.) Argumentation in Artificial Intelligence, pp. 1–22. Springer, Boston
(2009). https://doi.org/10.1007/978-0-387-98197-0 1
34. Walton, D., Reed, C., Macagno, F.: Argumentation Schemes. Cambridge University
Press, Cambridge (2008)
35. Wilcoxon, F.: Individual comparisons by ranking methods. Biom. Bull. 1(6), 80–83
(1945)
36. Wyner, A., van Engers, T.: From argument in natural language to formalised argu-
mentation: components, prospects and problems. In: Proceedings of the Worskhop
on Natural Language Engineering of Legal Argumentation, Barcelona, Spain (2009)
 A Non-intrusive Approach to Measuring
Trust in Opponents in a Negotiation
Scenario

Marco Gomes1(B) , John Zeleznikow2 , and Paulo Novais1

1
ALGORITMI Centre, University of Minho, Braga, Portugal
{marcogomes,pjon}@di.uminho.pt
2
College of Business, Victoria University, Melbourne, Australia
john.zeleznikow@vu.edu.au

Abstract. There is a consensus that trust in one’ opponent plays a sig-

nificant role in promoting parties to engage in the conflict management
process. Trust is an important yet complex and little-understood rela-
tion among parties in conflict. In general, trust can be seen as a measure
of confidence that an entity or entities will behave expectedly. Without
trust, the instruments to prevent or manage the conflict, such as nego-
tiation, are handicapped and cannot reach their full potential for pro-
moting an end to or a mitigation of a conflict. Hence, our motivation to
examine trust is three-fold. First, the present study aims to address and
expand on this line of research by investigating the possibility of mea-
suring trust based on quantifiable behavior. To do so, we provide a brief
review of the existing definitions of trust and define trust in the context
of a negotiation scenario. Further, we propose a formal definition so that
the analysis of trust in this kind of scenarios can be developed. Thus, it
is suggested the use of Ambient Intelligence techniques that use a trust
data model to collect and evaluate relevant information based on the
assumption that observable trust between two entities (parties) results
in certain typical behaviors. Third, this work aims to study the partic-
ular connection between relational aspects of trust and parties’ conflict
styles based on two dimensions: cooperativeness and assertiveness. The
main contribution of this work is the identification of situations in which
trust relationships influences the negotiation performance. To do so, an
experiment was set-up in which we tried to streamline all the relevant
aspects of the interaction between the parties and its environment that
occur in a sensory rich environment, to measure trust. To simulate a
conflict situation, a web-based game was developed. It was designed to
enable test participants to engage in a conflict experience induced by the
presence of Ambient Intelligence systems. Several tests were performed.
We then engaged in rigorous assessment, post- processing and analysis
of results. We validated the results comparing them with trust measures
obtained through the use of a questionnaire (carefully adapted) from
social networks.

Keywords: Ambient intelligence · Conflict handling styles

Negotiation · Trust measurement
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 528–542, 2018.
https://doi.org/10.1007/978-3-030-00178-0_36
 A Non-Intrusive Approach to Measuring Trust 529

1 Introduction
Trust is understood as a complex phenomenon, and it is widely accepted as play-
ing a significant role in human social relationships. Analysing trust definitions
in the main disciplines concerned with trust relationships (psychology, sociology,
etc.) lead us to a profusion of interpretations containing different trust types or
facets, with different properties, which require different models for analysis. This
abundance of meanings leads to a degree of uncertainty about what is meant
by trust, creating conceptual and terminological confusion. While trust in the
system and your adversaries is vital in negotiating a dispute, it is not clear what
trust is, nor how trust can be enhanced. In fact, the most potent concern about
trust seems to be when it is absent [1]. Although a variety of definitions of the
term trust have been suggested, this paper will use the definition proposed by
Castelfranchi who saw trust as a rich and complex mental attitude of x towards
y as for a given action and goal [2]. This attitude consists of evaluations of y
and the situation, and of expectations about y’s mind, behaviour and possible
results. This includes, of course, the assumption that trust should be understood
and interpreted by framing its natural subjectivity and the information needed
at a particular time and for a specific context in a computer-based model: an
abstraction that has the power to represent data in terms of entities and rela-
tionships relevant to a domain of inquiry (trust).
Scholars and practitioners agree that trust and trust scarcities play a central
role in conflict and conflict resolution [3]. From empirical studies, the absence
of trust binds parties to conflict. Further, the presence of trust is stated as a
necessary condition for parties engaging in a conflict mitigation process. These
considerations can be distinctly observed in a typical conflict scenario (buyer
versus seller) in which the need for trust emerges where the eventual outcome
is dependent on the actions of both parties and also from strategies within the
conflict management process itself. In other words, trust is involved where, to
get what you want, you are dependent on the other party not exploiting the
situation to your eventual cost. For example, a seller who is faced with a claim
by a buyer may realize that to resist the claim would only encourage the buyer
leaving the negotiation process for lower-price sellers elsewhere. By agreeing to
the claim, this exodus would be (apparently) prevented. With this outcome,
both seller and buyer appear to receive some benefit but the seller has to trust
that the buyer will not then claim a similar benefit to other sellers. Zeleznikow
used his Family Winner system to analyse the Israeli-Palestinian dispute [4]. His
results mirrored the outcomes of Oslo accords signed twenty years previously.
Arafat, Peres, and Rabin won a Nobel Peace Prize for their efforts. Nevertheless,
the accord unravelled. Zeleznikow claimed while a logical, a mutually beneficial
solution exists, the proponents will not take it up, because of a lack of trust in
each other. Further, every act of violence further increases this distrust.
By investigating this link, one can better understand the potential effects
of trust in conflict and be better prepared to prevent conflict from escalating
[20,21]. When researchers and practitioners understand which conditions lead
to trust, it seems plausible to state that they can follow different avenues toward
530 M. Gomes et al.

trust building, and therefore toward conflict mitigation. Furthermore, being able
to measure existence, emergence or dissolution of trusting relationships in a con-
flict situation will give a new set of information that can be used to improve
processes and interventions, enabling the characterization of individuals and
enhancing negotiation performance. It is also important to understand whose
factors of parties’ trust relationships can be captured algorithmically to spec-
ify the relationship between trust and conflict and to advance significantly the
conflict mitigation process. Despite the shared understanding that trust plays a
crucial role in conflict, the relation of trust to conflict is poorly studied. This is
partly due to the complexity (multi-dimensionality) of the trust. Consequently,
there is an evident lack of research on this link (trust-conflict) as well as to a
lack of instruments to measure trust in such context.
The multi-dimensionality of trust (meanings that trust is not determined by
one single component of the relationship, but by multiple components) makes
conflict interventions more complex since they have to meet different conditions.
But it also allows more entry points for the interventions. From a computer
science perspective, it becomes attractive to adopt an approach that reduces
this multidimensionality to something that computer systems can handle. This
is a point of view aligned with the dominant tradition in Economics and Game
Theory which suggests that trust as a concept is reducible to “subjective prob-
ability” (a quantitative and opaque view). Although differences of opinion still
exist, there appears to be some agreement that the reduction of trust to a man-
ageable number, quantity or probability is highly satisfactory. This is because,
after realizing the difficulty in interpreting the human perception subjectivity of
trust prevents us from defining objective evaluation criteria. However, we still
need to find a way of evaluating (measuring) trust to guide the research effort in
the correct direction. Facing this issue, we must raise a critical question: how to
balance the subjective nature of trust with the objectivity-dependent nature of a
computer system? In other words, how can a computer system deal intelligently
with this kind of subjectivity? Well, some glimpses into the ways that computer
systems deal with subjectivity can be found in the “expert” literature. This is
exemplified in the work undertaken by Rosalind Picard in which she stressed
that this endeavour is challenging but achievable [5]. Afterwards, she outlined a
strategy for computer systems to cope with subjectivity issues. Specifically, it is
proposed a three-fold approach: that they [computer systems] will need to (1)
share some of the common sense of the user, (2) observe and model the user’s
actions, and (3) learn from these interactions. To summarize, this approach sug-
gests that subjectivity is expressed as the user interacts with the system during
a succession of queries. These involving inputs of the user can be tracked, mod-
elled, and use to retrieve data consistent with changing requests. In resume, this
strategy seems to be suitable to be applied, so we planned to follow it in our
work in ways to overcome the further issue.
We describe, in the remainder of this article, the progress we have made
towards achieving these goals. The rest of the article is organized as follows.
In the following section, we provide insight into conflict and conflict handling
styles along with an explanation of the mathematical model used to classify the
 A Non-Intrusive Approach to Measuring Trust 531

parties’ conflict style. Then a formal definition of the problem of measuring trust
is presented. The conditions of the negotiation game, the main findings, and their
analysis are provided next. The final section details the main conclusions drawn
from this study.

2 Conflict and Conflict Handling Styles

It is difficult to define and to reach a universally accepted definition of conflict.

We define it as a disagreement through which the parties involved perceive a
threat to their needs, interests or concerns [6]. Similarly, it can also be seen as
an opposition of interests that disrupts or blocks an action or decision-making
process [7]. Thus, a conflict may be seen as a process that begins when a party
feels that the other parties have or are about to negatively affect their interests.
In order to enhance conflict resolution, we believe it is important to: (1) provide
the parties and manager with valuable knowledge about the conflict and (2)
to assist parties throughout all the process. Having such information can lead
to a reduction in the severity of a conflict. Conflicts can develop in stages and
consequently may involve many different responses as the conflict proceeds. Peo-
ple involved develop various strategies, solutions or behaviours, to deal with the
conflict. In order to classify the conflict style, the proposals must be analysed,
namely regarding of their utility. In that sense, in each stage of the negotiation
the parties’ proposals are analysed according to their utility value and a range

Fig. 1. Relationship between the utility of a proposal and the personal conflict handling
style.
532 M. Gomes et al.

of possible outcomes defined by the values of the Worst Alternative to a Nego-

tiated Agreement (WATNA) and Best Alternative to a Negotiated Agreement
(BATNA) of each party. Lodder and Zeleznikow provide a detailed account of
such approaches [8]. This approach uses a mathematical model [9], which classi-
fies a party’s conflict style considering the range of possible outcomes, the values
of WATNA and BATNA as boundaries, and the utility of the proposal (Fig. 1).
The utility quantifies how good a given outcome is for a party, it is acceptable
to argue that a competing party will generally propose solutions that maximize
its own utility in expense of that of the other party (the utility of the proposal
is higher than the WATNA of the other party), whereas, for example, a compro-
mising party will most likely search for solutions in an intermediary region (the
utility of the proposal falls within the range of the zone of possible agreement, the
range of overlapped outcomes that would benefit both parties). Essentially, we
were able to classify the personal conflict style of a party by constantly analysing
the utility of the proposals created. Once the styles are identified, strategies can
be implemented that aim to improve the success rate of procedures for resolu-
tion and conflict management. This information and other insights about how
the individual’s conflict style is classified can be found in [9].

3 Measuring Trust Using an Algorithmic Approach

As suggested in [10], trust literature can be categorized based on three crite-

ria: (i) trust information collection, (ii) trust value assessment, and (iii) trust
value dissemination. Each in turn can be further classified: trust information
gathering from three sources, namely (i) attitudes, (ii) behaviors, and (iii) expe-
riences; trust value assessment according to the data model, namely (i) graph,
(ii) interaction, and (iii) hybrid; and trust value dissemination into trust-based
recommendation and visualization models. With regard to this work, and taking
into account Sherchan’s strategy [10], we will highlight the trust definitions, trust
types, properties, and measurement models from the perspective of the Com-
puter Science and Economics disciplines. Such approach seeks to focus on some
particular aspects of trust measurement, which are simplified, so that these prop-
erties can be captured algorithmically. We wish to obtain and/or quantify trust
by detecting statistically significant realizations of these trust-like behaviours in
computer-based systems. Therefore, the basis for our study is a proposition that
trust results in characteristic interaction behaviour patterns that are statistically
different from random interactions in a computerized conflict management sys-
tem. Another important decision was to reduce the study of the trust domain to
the study of interpersonal trust, transforming the challenge of measuring trust
to make it slightly more accessible. Furthermore, this type of trust is common
in most conflict situations and it plays a key role in conflict dynamics [11].
Facing these challenges, how, from a Computer Science point of view, can
trust can be classified? According to [10], in general, it can be divided into two
broad categories: “user” and“system”. For the scope of this present work, we
will consider only the notion of “user” trust is derived from Psychology and
 A Non-Intrusive Approach to Measuring Trust 533

Sociology, with a standard definition as “a subjective expectation an entity has

about another’s future behavior” [10]. This implies that trust is inherently per-
sonalized. In this sense, trust is relational. As two individuals interact with each
other frequently, their relationship strengthens, and trust evolves based on their
experience. Following the Adali proposition [12] related to interpersonal trust,
the main type or facet of trust understudy will be the basis for our proposal.
Hence, the focus will be on the following proposition: is it possible to observe
that trust between two entities A and B and will this result in certain typical
behaviours that can be statistically captured. In other words, our aim is to quan-
titatively measure dyadic trust (trust between two entities) based on observed
behaviour in a negotiation process. These behaviours are not only an expression
of trust but can also facilitate the development of further trust. The simplest
such behaviour is just interaction, in which an action occurs as two or more enti-
ties have some kind two-way effect upon one another. Regarding trust evaluation
models, our approach to trust computation will be based on an interaction-based
trust model. In this case, our interaction-based model evaluates trust based on
the interactions performed in a computer-based conflict management system.
Then trust-related information will be captured and assessed following the
strategy aforementioned. The sources and the means to assess trust can be
resumed to:
– Source of trust information gathering: Past Behaviour. Why? Because
user behaviour is an important aspect of trust. Another reason is that past
behavior is identified by patterns of interactions that can be captured by an
algorithm running on a computer system. Therefore, this will be our main
source of information gathering. For example, in a negotiation scenario if a
party is an extremely active participant and suddenly stops participating, this
change in behaviour (the interaction is interrupted) is noticeable and might
imply that this party’s trust in the other party or with the party with whom
he/she has been frequently interacting with has decreased.
– Trust value assessment: Regarding the techniques used to measure (com-
pute) trust, they can be broadly classified into statistical and machine learning
techniques, heuristics-based techniques, and behaviour-based techniques. In
the present work, we will use behaviour-based techniques to assess trust. The
reason for this choice lies in the fact that our measure of trust is based on
quantifiable behaviour. So it seems obvious to choose assessment techniques
based on this proposition.
After we have defined which directions and techniques we will add to measur-
ing trust, let us then formally define our notion based on Adali’s approach. This
formalization is applied to the context of a negotiation, a process for two (or
more) parties to find an acceptable solution to a conflict. Within a negotiation
process, each party can make several proposals and exchange an unlimited num-
ber of messages. For the context of the interaction in a negotiation, the input is
the proposal stream in a negotiation process, specified by a set of 4-tuples,

sender, receiver, message, time (1)

534 M. Gomes et al.

note that we pretend to study the problem of trust purely from the observed
interaction statistics, using no semantic information. Meanwhile, in our formal-
ization, in an interaction between parties some semantic aspects (e.g. message)
are considered in order to provide posterior semantic analysis. The output consid-
ered here is a set T induce from these inputs. The participants of the negotiation
are represented by the elements of this set.
There is evidence that parties anticipating an online negotiation expect less
trust before the negotiations begin and have less desire for future interaction
with the other party [13]. Assuming that, we postulate in this work that the
longer and more balanced an interaction is between two parties (meaning that
the average number of times that two entities interact within the process), the
more likely is less trust each other and, also, the less tightly connected they are.
In other words, assuming that parties negotiating electronically are characterized
by lower levels of pre-negotiation trust, they will need to interact more to develop
trust and so increasing the likelihood that negotiation will proceed in a positive
direction. So, the fundamental task is first to identify when two elements of T
set are interacting. Let A and B be a pair of users, and let P = {t1 , t2 , ..., tk }
be a sorted list of the times when a message was exchanged between A and B.
Therefore the average time between messages is defined as τ = (tk − t1 )/k.
The measure of trust will be based on the interactions in I, obeying the fol-
lowing postulates: (1) Longer interactions imply less trust; (2) More interactions
imply less trust; and (3) Balanced participation by A and B implies less trust.
We define the relational trust RI (A, B) as follows:
l
RI (A, B) = i=l Ii  · H(Ii ) (2)

Where H(Ii ) is a measure of the balance in the i interaction contained in I. We

use the entropy (measure of the amount of information that is missing in the
flow of interaction) function to measure balance:

H(Ii ) = −p log p − (1 − p) log(1 − p), (3)

where p(Ii) is the fraction of messages in the interaction Ii that were per-
formed by A. The complexity of the algorithms for computing relational trust is
Θ(|D| log |D|), where |D| is the size of the interaction stream.

4 Case-Study: A Negotiation Scenario

A negotiation scenario is commonly used to describe the setting of a negotiation.

It consists of a domain, an outcome space, preference profiles, and negotiation
styles (adapted from [14]). A negotiation scenario should specify everything we
know about the problem being addressed in order to frame the interactions that
occur within. The detailed scenario that frames our experiments is presented in
the following subsection (Subsect. 4.2). Taking into account the domain of the
negotiation (which specifies the issues that the negotiators need to agree upon),
 A Non-Intrusive Approach to Measuring Trust 535

in our case-study, the single issue negotiators need to agree on is the price of
the item being sold. The negotiation space represents all the outcomes achiev-
able i.e. all possible issue value combinations within the domain. Once again,
the negotiation space is restricted following the mathematical model presented
in Sect. 2. Moreover, the negotiation space represents all the outcomes achiev-
able i.e. all possible issue value combinations with the domain. Once again, the
negotiation space is restricted following the mathematical model presented in
Sect. 2. Regarding the negotiation styles being used in a negotiation process,
this refers to how a participant behaves, and how she is expected to behave.
And even within the process, negotiating behaviours add a layer of complex-
ity to the negotiation. Again, this issue will be addressed applying the model
introduced in Sect. 2.
With this purpose in mind, we adapted a technological framework (presented
in Subsect. 4.1) that aims to support the decision-making of the conflict manager
by facilitating access to information such as the conflict handling style of the
parties or their social context. In this work, we introduce a new module that
takes into account the context using trust analysis. The development of such
a framework previews a set of services or functionalities that will support the
work of the conflict manager. The underlying aim is to release her so that she
can have a more informed and effective approach to deal with complex issues
such as the improvement of interpersonal communication and relationships.
Moreover, the main objective of this research work is to identify and measure
the users’ interpersonal trust, to correlate to their negotiation performance and
how it can be used in a simple conflict (zero-sum) situation. To demonstrate
this, an experiment was set up (See Subsect. 4.2) in which we tried to estimate
all the relevant aspects of the interaction between the individual that occur in
a sensory rich environment (where contextual modalities were monitored). The
participants of the experiment were volunteers socially connected with our lab
members. Twenty individuals participated, both female and male, aged between
22 and 36. The first step of the experiment was to ask the volunteers to fill in
a small individual questionnaire (depicted in the following section). The follow-
ing step was the monitoring of the individuals’ interaction with the developed
web-based negotiation game. During the experiments, the information about
the user’s context was provided through a monitoring framework, which is cus-
tomized to collect and treat the interaction data. The participants played the
web-based game through computers that allowed the analysis of the described
features.

4.1 An Intelligent Conflict Support Environment

An Ambient Intelligent (AmI) system consists of a series of interconnected com-
puting and sensing devices which surround the user pervasively in his environ-
ment and are invisible to him. It is clear that when one designs a system with
these characteristics there are some challenges that are raised by the heterogene-
ity and number of devices and technologies present in this kind of environment.
Is not just a bunch of devices interconnected but instead a group of devices
536 M. Gomes et al.

Fig. 2. A conceptual framework to support the decision-making of the conflict manager.

working together and sharing information towards a common goal. To cover

it the system must ensure that these devices interact and exchange informa-
tion, and compatibility must be guaranteed between the different technologies
and components. Our aim is to provide a service that is dynamically adapted
to the interaction context so that users (parties and the conflict manager) can
naturally take advantage of interacting with the system and thus enhance the
conflict management process. It can be stated that our underlying intent is to
extend the traditional technology-based conflict resolution/management meth-
ods, in which a user simply interacts with the system, with a new component, an
Intelligent Environment (IE). These environments should be made in a pervasive
and transparent way since when people are aware that they are being monitored,
they tend to behave differently. Therefore, towards constructing an intelligent
conflict support system the following ambient intelligence system was developed
(Fig. 2).
The aim of the system is to sense conflict context, acquire knowledge and
then make reasoning on the acquired context and thus act on the parties’ behalf.
To achieve this, the system builds a profile of each party and can link that pro-
file subsequently with the individual performance within the conflict process
that is monitored by the system. In other words, while the user conscientiously
interacts with the system and takes his/her decisions and actions, a parallel and
transparent process takes place in which contextual and behavioural information
 A Non-Intrusive Approach to Measuring Trust 537

is sent in a synchronized way to the conflict support platform. Upon converting

the sensory information into useful data, the platform allows the conflict man-
ager to obtain a contextualized analysis of the user’s data. The contextualized
analysis of user’s data is critical when the data is from heterogeneous sources of
diverse nature like sensors, user profile, and social media and also at different
timestamps. To overcome some of these problems, the features are extracted
from multiple sensor observations and combined into a single concatenated fea-
ture vector that is introduced into different classification modules (conflict styles,
trust analysis, etc.). To integrate all the multimodal evidence we used a decision
level integration strategy. Examples of decision level fusion methods used in this
work include weighted decision methods and machine-learning techniques and
are detailed in previous work [15,16].

4.2 Experimental Set-Up

As stated before, to simulate a conflict situation in real-life environments a web-
based game was developed. It was designed to enable test participants to have a
conflict experience induced by the presence of an Ambient Intelligence system. In
that sense, the game simulates a business situation (a conflict) where each party
wishes to achieve a desired result in the negotiation or else go bankrupt. The
desired outcome was a win/win situation for both parties. A performance-based
reward was settled to increased participants’ intrinsic motivation. The game
starts with the application randomly giving one of the predetermined roles to
each party. The instructions to win the game were to negotiate a satisfactory
deal and ensure that the party in question didn’t go bankrupt. Each party’s
instructions were presented, visible to them through the application interfaces.
The objectives and the persona for each party are depicted as:
– Role A - party A was a piano seller who specialized in selling cheap pianos. He
was not the only supplier of this kind of pianos. In order to stay in business,
he needed to sell each piano at 1000 Euro or more, knowing that piano prices
vary greatly depending on the locale and the particular situation. If A did
not achieve this result, she would go bankrupt. Party A was also given the
information that Party B needed to make this deal.
– Role B - Party B represented a musician that need a piano urgently. He had
recently received a contract to give a concert and needed a piano to play. The
contractors were prepared to pay 1200 Euro per concert. If Party B did not
manage to negotiate with Party A to buy the piano at 1200 Euro or less, then
he would go bankrupt. Party B was told that party A was in a little financial
trouble and needed to make the deal to survive.
Regarding the conflict styles analysis, the ZOPA (Zone Of Possible Agree-
ment) was bounded by the BATNA (1000 Euro) and WATNA (1200 Euro) val-
ues. The range of possible agreement is 200, but the parties were not aware of
this detail. Moreover, it does it in an entirely transparent and non-intrusive way,
i.e., rather than relying on traditional self-reporting mechanisms such as ques-
tionnaires to infer behaviours, it analyses the actions of the parties, in real-time.
538 M. Gomes et al.

To validate our trust measurements, we will assume that trust phenomenon

in some way increases or decrease in magnitude and strength within a relation-
ship. In other words, it is assumed that, at early stages of a relationship, an
individual will carefully calculate how the other party is likely to behave in a
given situation depending on the rewards for being trustworthy and the deter-
rents against untrustworthy behaviour [17]. From this perspective, in our study
we hypothesised that the degree of a relationship between two individuals could
be a suitable measure of their trust relationship. To do so, in this work, we used
a social network analysis method to map and measure the experiment’s partici-
pants relationships. This approach was based on the intuitive notion that these
patterns are important aspects of trustworthy relations of the individuals who
display them through their interaction, namely in a conflict situation. In this
manner, the data was obtained through a questionnaire (carefully adapted) of
social networks. Since the interest of the research was to analyse the strength of
relationships, we use responses on a 5-point Likert Scale from “not at all” (1) to
“very much” (5). With the data obtained, we measured the in-degree centrality
scores for each disputant. The in-degree score is an indication of centrality, which
measures the number of nominations received by a person. This variable was cal-
culated for all individuals in the Advice, Hindrance and Friendship Network. At
the individual level, we can also calculate the position of individuals who may be
intermediaries (Betweenness Centrality), influential (Eigenvector Centrality) or
conflicting (In-degree Centrality in Hindrance Network). The questionnaire was
sent to a network of twenty individuals, the data was collected in binary format
and was stored in 20 × 20 square matrices. We calculated dependent variables
with the software program UCINET VI. The meta-analytic techniques used to
analyse the data obtained were as follows: Descriptive statistics, both standard
deviations, etc.; Correlations between the independent variables (centrality and
density) and dependent variables (hindrance); Regressions to test the model.
Gender was used as a control variable.

4.3 Results
The data containing the required information was acquired in digital form
through the players’ monitoring framework, which is customized to collect and
treat the interaction data. This data is combined and synchronized to describe
several important aspects of the behavior of the user. The participants played the
web based game through computers that allowed the analysis of the described
features. In the preliminary data analysis, the experimental data is organized
into two groups based on the social network analysis. One group contains the
collection of data obtained through the application of the questionnaire (measur-
ing the participant’s relationships). This enables the establishment of a baseline
for comparison with the second group, gathered through the web-based negotia-
tion game (applying the trust measuring algorithm previously presented), which
comprises the data gathered from the parties negotiating. To statistically deal
with data concerning to the utility values of the parties’ proposals, it was nec-
essary to convert data to an arbitrary numeric scale (zero is the least favorable
 A Non-Intrusive Approach to Measuring Trust 539

style for the resolution and four the most favorable style). This type of scale
ensures that the exact numeric quantity of a particular value has no significance
beyond its ability to establish a ranking over a set of data points. Therefore, it
was built rank-ordering which describes order, but not the relative size or degree
of difference between the items measured. This was a necessary step to make the
data suitable for statistical and machine-learning techniques. Further, according
to our postulates (see Sect. 3), the basis for this analysis was the assumption that
if A and B have a strong relationship then RI (A, B) value (the relational trust)
is below the median of the calculated trust for all the pairs within the interac-
tions data set. In other words, this implies that if one pair of participants has a
high degree of relationship then the same pair of participants, during the game,
will perform fewer interactions than the median of total interactions per game.
At this point, it should be highlighted that to apply non-parametric statistical
analysis the raw data was firstly pre-processed and secondly, it was subjected to
tests. The outcomes were compared using the Mann-Whitney U test (compares
the central tendencies of two independent samples), given the fact that most of
the distributions are not normal. The null hypothesis is thus: H0 = The medians
of the two distributions are different. For each pair of distributions compared,
the test returns a p-value, with a small p-value suggesting that it is unlikely
that H0 is true. For each parameter (a pair of participants), data from both
samples is compared. In all the tests, a value of α = 0.05 is used. Thus, for every
Mann-Whitney test whose p-value < α, the difference is considered to be sta-
tistically insignificant, i.e., H0 is rejected. Consequently, the results have shown
that no (statistically) important difference between data from the two samples
were found. In other words, it means that our assumptions were valid. So it is
possible to infer that trust measurements using our algorithmic approach are
valid to a certain extent.
To analyse if trust relationships influence the negotiation performance, in the
preliminary data analysis, the experimental data is organized into two groups
based on the analysis of the trust measurements. One group contains the collec-
tion of experimental data about how a user (A) behaves when he/she negotiates
with someone (B) in which RI (A, B) has a low value (RI (A, B) < median).
This approach enables the establishment of a baseline for comparison with the
second group, which comprises the data gathered from parties that negotiate
with someone that has high RI values (RI ≥ median). To statistically deal with
data concerning the utility values of the parties’ proposals, it was necessary to
convert the data to an arbitrary numeric scale (0 is the least favourable style for
the resolution and four the most favourable style). This type of scale means that
the exact numeric quantity of a particular value has no significance beyond its
ability to establish a ranking over a set of data points. It was constructed using
in rank-ordering (which describes order), but not the relative size or degree of
difference between the items measured. This was a necessary step to make the
data suitable for statistical and machine-learning techniques.
From the data analysis regarding the evolution of the conflict handling styles
evidenced by the parties, we concluded that the conflict style is on average
540 M. Gomes et al.

more favourable (mutually beneficial) when the parties have a lower value of
RI than average. The final value of the negotiation process occurs when RI is
lower, the parties reaches mutually satisfactory solutions, i.e., solutions that are
closer to the optimum result. Also interesting is the conclusion that participants
who have a high RI value need more rounds and exchange longer messages to
achieve a successful outcome. This outcome is aligned with our postulates, in
which we have hypothesized that more interaction implies less trust in a conflict
situation. However, it can be assumed that when the trust relationship is weak,
the development of trust between the opponents need more steps than otherwise.
Indeed, events at the beginning and end of a negotiation sequence play a major
role in building and maintaining trust [18].
Moreover, the analysis shows that there is an apparent difference between
the two groups regarding the conflict styles exhibited during the game. One
conclusion is that when participants share a significant trust relationship (low
RI value) the frequency of collaborative behaviours is far superior (49%) than
otherwise (24%). A conclusion that corroborates the Malhotra and Murnighan
study, in which they found in their study that an individual’s trust will increase
as the number of positive interactions between parties increases [19]. In a similar
analysis, but now concerning the roles played by participants, we conclude that
the sellers are much more competitive than buyers (57% vs. 29%) while buyers
are primarily collaborative. To interpret the significance of these results it is
important to recall that participants were asked to negotiate a favourable deal
in a competitive and win-lose scenario. Nevertheless, it is shown that when par-
ticipants have a significant trust relationship they are more likely to transform
it into a win/win situation. This is especially visible in the final results of the
negotiations. On the one hand, we find that 100% of the agreements made by
parties with an important trust relationship accomplished a successful deal, i.e.,
between the range of solutions that would benefit both. On the other hand, only
50% of negotiations that occurred between untrusted opponents (low RI value)
reached a mutual benefits agreement. It may be that they assumed they had
to negotiate and get the best price (win/loose). But that was not the objective.
Their objective was to negotiate a deal so they would not go bankrupt (win/win).
The preliminary evidence suggests a basis for expecting a connection between
the trust relationship and the use of conflict styles. Despite these results, we
still do not know much about how this kind of influence might facilitate (or
inhibit) positive conflict outcomes. Therefore, we will perform more and deeper
experiments to understand how to collect and analysis relational ties that can
influence negotiation performance.

5 Conclusion

The starting point for the development of trust measurements lies in the fact
that current computer-based conflict assessment tools do not consider (or are
unable to measure it) the underlying trust relations between parties in a par-
ticular scenario. Thus, within this work, we aim firstly to identify and apply
 A Non-Intrusive Approach to Measuring Trust 541

an algorithm for measuring interpersonal trust; secondly, to validate this app-

roach opposing data collected from a questionnaire with data gathered from a
web-based negotiation game and to statistically study the correlation between
mutual trust and conflict styles. From the experiment outcomes, the findings
highlight the potentially quantifiable measurements of trust which further the
understanding of conflict dynamics. They indicated relationships between the
features being monitored and the participants’ relationships elicited through a
small questionnaire. These findings have the potential to enable the characteriza-
tion of individuals and enhance negotiation performance. The main contribution
of this work is thus the identification of trust relations between opponents in a
negotiation scenario.
These results are preliminary in the sense that there is more information that
one can retrieve from the collected data namely through a deeper semantic anal-
ysis. This type of analysis could considerably enhance the trust measurements.
Furthermore, due to the small sample size used in the current study, some cau-
tion must be taken when interpreting the results of the statistical analysis pre-
sented and underpinning the conclusions. In that sense, additional limitations of
the current research must be pointed out. First, the participants were recruited
from a particular population (that are socially related to our lab members)- a
population that may limit the generalizability of the results. Admittedly, the
participants of the experiment may not be representative of conflict parties in
general. Consequently, we are unable to demonstrate the causality of the vari-
ables conclusively. Moreover, it is possible that individual differences (i.e., per-
sonalities) might have influenced the impact of the results. Another drawback
of our study was that we tested all of the variables at the individual level. This
limited us from conducting global level analysis, which could provide us more
variance of the data. Also, the data was collected through self-reported surveys
at one time, which is subject to common method variance problem. Finally, the
computational facet of this work should provide an understanding of the diffi-
culties in algorithmically capturing and computing interpersonal trust. A more
comprehensive and in-depth study to provide theoretical advances, as well as
implement technological solutions, is yet to be developed.

Acknowledgements. This work has been supported by COMPETE: POCI-01-0145-

FEDER-007043 and FCT - Fundação para a Ciência e a Tecnologia (Portuguese Foun-
dation for Science and Technology) within the Project Scope UID/CEC/00319/2013.

References
1. Fells, R.E.: Developing trust in negotiation. Empl. Relat. 15(1), 33–45 (1993)
2. Castelfranchi, C., Falcone, R.: Trust is much more than subjective probability:
mental components and sources of trust. In: Proceedings of the 33rd Hawaii Inter-
national Conference on System Sciences, Washington, DC, USA, HICSS 2000, vol.
6, p. 6008. IEEE Computer Society (2000)
3. Kappmeier, M.: Its all about trust how to assess the trust relationship between
conflict parties. In Proceedings of the IACM 24th Annual Conference Paper (2011)
542 M. Gomes et al.

4. Zeleznikow, J.: Comparing the Israel–Palestinian Dispute to Australian Family

Mediation. Group Decis. Negot. 23(6), 1301–1317 (2014)
5. Picard, R.W.: Computer learning of subjectivity. ACM Comput. Surv. 27(4), 621–
623 (1995)
6. Deutsch, M.: The Resolution of Conflict: Constructive and Destructive Processes.
Carl Hovland Memorial Lectures. Yale University Press, New Haven (1977)
7. Robbins, S.P.: Organizational Behavior. Prentice Hall, Upper Saddle River (2001)
8. Lodder, A.R., Zeleznikow, J.: Enhanced Dispute Resolution Through the Use of
Information Technology. Cambridge University Press, Cambridge (2010)
9. Carneiro, D., Gomes, M., Novais, P., Neves, J.: Developing dynamic conflict reso-
lution models based on the interpretation of personal conflict styles. In: Antunes,
L., Pinto, H.S. (eds.) EPIA 2011. LNCS (LNAI), vol. 7026, pp. 44–58. Springer,
Heidelberg (2011). https://doi.org/10.1007/978-3-642-24769-9 4
10. Sherchan, W., Nepal, S., Paris, C.: A survey of trust in social networks. ACM
Comput. Surv. 45(4), 47:1–47:33 (2013)
11. Ebner, N.: Online dispute resolution and interpersonal trust. In: ODR: Theory and
Practice (2012)
12. Adali, S., et al.: Measuring behavioral trust in social networks. In: 2010 IEEE
International Conference on Intelligence and Security Informatics (ISI), pp. 150–
152, May 2010
13. Naquin, C.E., Paulson, G.D.: Online bargaining and interpersonal trust. J. Appl.
Psychol. 88(1), 113–120 (2003)
14. Marsa-Maestre, I., Klein, M., Jonker, C.M., Aydoan, R.: From problems to proto-
cols: towards a negotiation handbook. Decis. Support Syst. 60, 39–54 (2014)
15. Gomes, M., Oliveira, T., Carneiro, D., Novais, P., Neves, J.: Studying the effects
of stress on negotiation behavior. Cybern. Syst. 45(3), 279–291 (2014)
16. Carneiro, D., Novais, P., Andrade, F., Zeleznikow, J., Neves, J.: Online dispute
resolution: an artificial intelligence perspective. Artif. Intell. Rev. 41(2), 211–240
(2012)
17. Lewicki, R., Brinsfield, C.: Measuring trust beliefs and behaviours. In: Lyon, F.,
Möllering, G., Saunders, M. (eds.) Handbook of Research Methods on Trust, p.
29. Edward Elgar Pub. (2012)
18. Lewicki, R.J., Polin, B.: Trust and negotiation. In: Handbook of Research on Nego-
tiation, chap. 7, p. 161 (2013)
19. Malhotra, D., Murnighan, J.K.: The effects of contracts on interpersonal trust.
Adm. Sci. Q. 47(3), 534–559 (2002). 9
20. Ebner, Noam, Zeleznikow, John: Fairness, trust and security in online dispute
resolution. Hamline Univ. Sch. Law’s J. Public Law Policy 36(2), (2015)
21. Han, G., Harms, P.D.: Team identification, trust and conflict: a mediation model.
Int. J. Confl. Manag. 21(1), 20–43 (2010)
 Network, Visualization, Analytics. A Tool
Allowing Legal Scholars
to Experimentally Investigate EU
Case Law

Nicola Lettieri1,3,4(B) , Sebastiano Faro2 , Delfina Malandrino3 ,

Armando Faggiano3 , and Margherita Vestoso4
1
National Institute for Public Policies Analysis (INAPP), Rome, Italy
n.lettieri@inapp.org
2
Institute of Legal Information Theory and Techniques (ITTIG-CNR),
Florence, Italy
3
University of Salerno, Fisciano, SA, Italy
4
University of Sannio, Benevento, Italy

Abstract. Legal Informatics has recently witnessed a growing interest

towards the insights offered by the intersection among Network Analysis
(NA), visualization techniques and legal science research questions. Also
thanks to several seminal works, the field is ready to tackle new challenges
at a theoretical and application level. The first is to bring the network
approach into “genuinely legal” research questions. The second is to cre-
ate tools allowing legal scholars without technical skills to exploit NA
with two goals: (i) make experiments with NA and push new ideas both
in legal and NA science; (ii) use NA and visualization in their daily activi-
ties (e.g., legal analysis and information retrieval). Against this backdrop,
a truly interdisciplinary approach deeply involving legal experts/scholars
is needed. The paper presents an ongoing research project - EUCaseNet
- dealing with these challenges and aiming to explore the potentialities
of NA in supporting the study of EU case law.

Keywords: EU case law · Legal documents’ relevance

Network analysis · Visualization · Analytics · On-line laboratory

1 Introduction
As witnessed by a growing literature, Legal Informatics research area has been
developing a growing interest towards the insights offered by the intersection
among Network Analysis (NA) [1], visualization techniques and legal science
research questions (see, i.a., [2,3]). Thanks to its high level of abstraction and
ability to support the understanding of both structural and functional features of

c Springer Nature Switzerland AG 2018

U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 543–555, 2018.
https://doi.org/10.1007/978-3-030-00178-0_37
544 N. Lettieri et al.

real networks, NA is spreading in the legal domain focusing on different research

priorities:

– Analyze the structural and functional features of more or less wide areas
of legal systems (e.g., the level of complexity of legislation, case law, legal
literature);
– Determine the relevance of legal documents and sources, according to the
different meanings acquired by the concept of “relevance” itself in different
legal contexts;
– Study the relations between different expressions - or, in the words of [4]
different “legal formants” - of the legal phenomenon (relationship between
legislation, case law and legal literature, or between supranational case law
and domestic case law, etc.);
– Investigate the evolution of a legal order also in a predictive fashion;
– Design innovative visual analytics tools for better legal information commu-
nication and retrieval.

In the following sections we present a research project aiming to find new ways
to deal with the above mentioned priorities. The attention is focused, more in
detail, on the creation of EuCaseNet [5], an innovative tool enabling experimental
use of NA techniques in the study of EU case law. The first part of the paper
is devoted to a brief overview of the “NA and Law” research experiences that
are somehow inspiring our project, followed by a sketch of the challenges to be
faced in this field. In its second part the paper presents and discusses both the
domain-related and technical issues of the project, sketching on-going works and
future directions.

2 Network Analysis and Law: Current Experiences

and Future Challenges

The determination of the relevance of legal sources and documents with NA

techniques is one of the most discussed issues in recent times. The study of net-
works has been frequently used to analyze case law especially in common law
countries where precedents have a prominent role (see, i.a., [6–8]). An interesting
and seminal work is the one presented in [7], that exploits the HITS algorithm [9]
and the patterns in citations networks within and across cases to create impor-
tance scores identifying the most legally relevant precedents in the network of
US Supreme Court case law. By comparing different centrality measures, the
authors show how the HITS algorithm allows to identify, with an higher level of
precision than the other centrality measures, those cases in the network that are
“influential” (inwardly relevant) or “well founded in law” (outwardly relevant).
Another relevant research experience in this area is the one recently described
in [10] in which the PageRank algorithm is combined with InDegree and the
HITS algorithm to investigate the relevance of the decisions of the EU Court of
Justice.
 Network, Visualization, Analytics 545

As to the uses of NA to measure structural characteristics of legal systems,

such as the complexity of large bodies of law, a paradigmatic example is the
research proposed by [11] in a work focusing on the legal complexity of the
United States Code that borrows concepts and tools from a range of academic
disciplines, including computer science, linguistics, physics, and psychology. By
developing a knowledge acquisition protocol that considers structure, language
and interdependence of Code’s norms as elements contributing to its complexity,
authors employ NA methods to test the interdependence level of the norms of the
Code, concluding that the more norms are connected, the higher is the amount
of energy an individual should expend in order to acquire knowledge regard-
ing the content of the United States Code. The work presented in [12] deals
with a somehow similar challenge: the authors try to build the “legal neigh-
borhoods” of Dutch norms aiming to verify if, given the network of an article
in focus, other relevant documents could be automatically identified just using
“objective” meta-information deriving from the norm graph (in the same direc-
tion see [3]). Quantifying the complexity of a given legal system is also close
to the goal of a recent research project [13] which aims to explore the inter-
dependence between different kinds of legal documents, such as case law and
doctrine, or judgments and norms, presenting a multi-relational network model-
ing the hierarchy and showing the relationships between different legal sources of
the European Union law (Treaties, International agreements, Legislation, Com-
plementary legislation, Preparatory acts, Jurisprudence) in order to shed light
on properties and behaviours of its different elements. The different sub-graphs
built by the authors have been also employed to realize a temporal analysis of
the evolution of the European legislation network, as well as to test the resilience
of the entire system, measuring its vulnerability under specific cases that may
lead to possible breakdowns. In the same vein, i.e., predicting the future devel-
opments of a given legal system, in [7] the network centrality measures are used
to forecast which cases will be most cited in the future.
All the mentioned experiences gave a significant contribution in building a
bridge between law and NA and in creating the basis for further scientific and
application developments. In this scenario, another factor of paramount rele-
vance is represented by the involvement of legal scholars in NA experiments
in order to support both the design of research activities and the validation of
the obtained results. Legal experts knowledge is essential in identifying mistakes
or ambiguities emerging from the legal network analysis performed by non-legal
experts compared to the real features of the law. As a matter of fact, there is not
necessarily an overlap between the results that can be simply achieved through
a naive application of NA measures to legal data and what could be considered
relevant for the law. It could happen that some typical network properties, usu-
ally measured by computer scientists interested to the computational features
of a specific graph (e.g., the resilience), were not usefully applicable to networks
in legal context, especially without an expert explaining the legal rationale or
meaning of the raw outcomes. From this perspective, it is essential to create tools
allowing legal scholars without technical skills to exploit NA in the legal field
546 N. Lettieri et al.

with two goals: (i) make experiments with NA and, thus, push new ideas both
in legal and NA science; (ii) use NA and visualization in their daily activities
(e.g., legal analysis and information retrieval).

3 The EuCaseNet Project

EUCaseNet is an interdisciplinary research project aiming to apply data mining
and network analysis techniques to the study of EU case law made available in
open format by the EU Commission portal EUR-Lex. It aims not only to experi-
ment new ways to analyze legal corpora, but also to create an on-line laboratory
freely exploitable by legal scholars to conduct experiments by themselves. The
project so far has moved in a twofold direction. The first one consisted in the
design of a tool allowing “on the fly” application of NA techniques to the entire
corpus of EU case law (judgments, orders, opinions and conclusions of the Gen-
eral Advocate of the Court of Justice and of the General Court, from the date
of their creation) in XML format. The basic idea is to enable the exploration of
NA techniques by domain (law) experts stimulating an inductive process that
encompasses, in a cyclic form, two fundamental steps:

– on the basis of the knowledge of the legal domain and having a basic under-
standing of NA, it is crucial to choose the measures and the data to consider;
– apply measures, compare the results with those achieved by domain experts
through traditional legal analysis methods and establish which combination
of data and measures better map knowledge of legal reality.

The second direction aims to devise new measures and algorithms to be

validated according to the state of the art of legal literature.

3.1 System Functionalities and Application Workflow

EuCaseNet allows users to make experiments within the network of case law gen-
erated from the judgments of the EU Court of Justice. The tool allows to visu-
alize the connections among all the judgments given by the Court until today1
and make experiments to study relevant phenomena from the legal theory view-
point. More in detail, the judgments are considered as a network and mapped
on a graph: nodes correspond to judgments and edges correspond to citations
between them. Therefore, by leveraging NA and visualization techniques we are
able to analyze the network.
To perform experiments on the graph of judgments, EuCaseNet provides
filters and measures of the social network analysis. Specifically, to define the
color and the size of the nodes, users can apply several algorithms. Moreover
it is possible to filter the judgments on the available attributes, so as to focus

1
We currently consider judgments given until April 2014.
 Network, Visualization, Analytics 547

only on a subset of them. The process that allowed us to build the graph can be
illustrated as follows.

– Data retrieving
Judgments are freely available from the EUR-Lex portal2 . We downloaded
them in XML format.
– Judgment Parsing
The downloaded XML file includes, for each judgment, a large number of
information. We decided to analyze only a subset of them:
• Title and reference (Celex number) of the judgment;
• The name of the Advocate-General who delivered its conclusions;
• Classification according to the EUR-Lex classification schema;
• Date;
• Involved EU country;
• Relationships (expressed in terms of citations) among judgments;
• List of bibliographic references to comments on judgments published in
books and journals (expressed as a number).
– Graph creation, import and manipulation
Given the XML file of the judgments, we allowed users to visualize and manip-
ulate the generated graph through NA metrics (degree, betweenness, close-
ness, eccentricity), and other well-known algorithms (i.e., PageRank). In Fig. 1
we show a screenshot of EUCaseNet. In the center the generated graph of the
judgments is shown. On the left side, we show an Interactive Panel in which
it is possible to interact with the graph. Specifically, it is possible to:
• Define the size and the color (in red scale) of the node according to the
above-mentioned SNA metrics;
• Apply filters based on the attributes defined inside the judgments graph
and that we considered for our work;
• Select a data range in order to focus on a portion of the graph.
When a judgment is selected, all cited judgments and those citing it (graph
neighbours) will be highlighted. Moreover, on the right side, an Information
Panel will show the attributes of the selected judgment.

3.2 Architecture

Here we briefly describe the technical architecture of EuCaseNet, and how com-
ponents interact each other. As we can see in Fig. 2, EUCaseNet is a client-server
architecture. The Server component (EUCaseNet Server) is a Java Servlet and
manages both the database module (MySQL) accessible with JDBC, and the
XML Manager module; the Client component contains a Web Browser module
to make requests.
The EUCaseNet Server implements the third step of the aforementioned
process (Graph creation, import and manipulation), through the Servlet within
Apache Tomcat. A generic client can interact with EUCaseNet through a simple

2
http://eur-lex.europa.eu/homepage.html.
548 N. Lettieri et al.

Fig. 1. EuCaseNet: network of the judgments with details about the Bosman case.

Fig. 2. EuCaseNet architecture: overview.

Web Browser. When a user accesses to EUCaseNet an HTTP request is sent to

the Server that interacts with XML Manager in order to generate the graph.
Once the process has finished a JSON file is produced (with the information
about the graph) and is sent to the client through an HTTP response. The Web
Browser contains a JavaScript library to render the graphs. Users can interact
with them through the operations described in previous section.

3.3 Preliminary Experiments

After a short setup phase, we began testing the tool. First we confirmed that
even the EU case law network, like other legal networks, has a scale free topology.
As shown in Fig. 3, the degree distribution follows a power law.
 Network, Visualization, Analytics 549

Fig. 3. In-degree distribution of the judgments of the EU Court of Justice. Data are
plotted on doubly logarithmic axes via the cumulative distribution.

Hence, we decided to start dealing with the problem of the determination of

the relevance of judgments in the specific context of EU case law. Indeed, one
cannot ignore the difficulty, especially in legal information retrieval systems, to
trace the concept of relevance back to a unitary definition, due to the interplay
between different factors. As highlighted in [14], according to Saracevic’s theory
of multiple “manifestations” of relevance [15], a document cannot be conceived
relevant in an absolute sense. The importance of a legal document, in particular,
should be measured in relation to the role it takes with regard to the user,
the information retrieval system or the legal domain. Therefore, its relevance
can be conceived in terms of a match between the information object and a
range of elements, such as the query codification (i.e., algorithmic relevance), the
subject (i.e., topical relevance), users’ information need (i.e., cognitive relevance)
or motivations (i.e., situational relevance), as well as the opinion of the domain
experts (i.e., domain relevance).
Being aware of the “multi-dimensionality” of the concept of relevance [16]
(and of the fact that its different profiles partly overlap and are often indicated
by similar clues), we decided to focus on a profile of relevance - substantially
traceable to the concept of “domain relevance”- that is well-identified in a recent
550 N. Lettieri et al.

book edited by two scholars of EU law listing the “classics of the EU law” [17].
The concept is summed up as follows:

The “classic” judgment in the jurisprudence of a Court are those that

survive the passing of time but also those which, even when their concrete
legal answer has become obsolete or has even been overturned, have lived on
through their multiple effects in other areas of the law. They are judgments
of systemic impact, embodying a broader normative lessons about the legal
order in which that Court operates. To identify them we must look not only
at the particular judgment as such but also at how it has been interpreted
and developed by the Court in other cases. We also need to look at the
attention paid to it by lawyers, legal scholars and other social actors. The
law is also a function of how the judicial decisions are interpreted and
challenged by the broader legal and social communities in which a court
operates.

Taking cue from this definition, the volume identifies a set of 19 judgments
that we used as gold standard for our experiments. More in detail, keeping the
cases listed in the volume as a reference, we used EuCaseNet to measure the
values assumed, for each case, by the typical NA metrics:
– Degree centrality. This metric measures the number of links incident upon a
node. In graph theory, degree centrality weights the number of ties a given
node has, by allowing to identify which nodes are central in the network with
regard to information spreading and the ability to influence others in their
immediate neighbourhood. Moreover, in directed networks, we can distinguish
between in-degree and out-degree centrality, which respectively measure the
number of ties directed to the node and the number of ties that the node
directs to others.
– Closeness centrality. This measure expresses the mean length of all shortest
paths from a node to all other nodes in the network (i.e., how many hops on
average it takes to reach every other node), making possible to estimate how
fast a given node is reached by the other nodes of the network.
– Betweenness centrality. Betweenness centrality depends on the relative num-
ber of shortest paths that run through a node, by measuring how many times
a node works as a bridge along the shortest path between two other nodes. As
a result, the higher is the level of betweenness, the more the node is crucial
in allowing communication between all couples of nodes.
– Eccentricity. In the mathematical field of graph theory, the distance between
two nodes in a graph is the number of edges in a shortest path (also called a
graph geodesic) connecting them. This is also known as the geodesic distance.
Notice that there may be more than one shortest path between two nodes.
If there is no path connecting the two nodes, i.e., if they belong to different
connected components, then conventionally the distance is defined as infinite.
– Page-Rank. PageRank is a link analysis algorithm and it assigns a numerical
weighting to each element of a hyperlinked set of documents, such as the
World Wide Web, with the purpose of measuring its relative importance
 Network, Visualization, Analytics 551

within the set. The algorithm may be applied to any collection of entities
with reciprocal quotations and references.
– Eigenvector centrality. Grounded on an iterative process similar to Page-
Rank, eigenvector centrality allows to estimate the influence of a node in
a network, by weighting the connection level of nodes directly linked to it.
In other terms, the score assigned by the algorithm to each node depends on
scoring level of those connected to it, with high-scoring nodes having an higher
impact on target node’s eigenvector value then low-scoring nodes. Therefore,
the metric is useful in determining which are most “popular” nodes, which
are those linked to the most connected nodes.

Our goal was twofold: on the one hand, start testing the functionality of the
EUCaseNet tool; on the other hand, like other authors do (see, e.g. [10]) explore
even if in an initially “rough” way, the performance of each metric in mapping
the relevance of the cases belonging to our gold standard.
Table 1 shows the judgments and the values assumed by the above men-
tioned metrics. Table 2 provides an overview of the performance of each of the
parameters taken into account by comparing the relevance assigned to the 19
judgments by the gold standard (in our analysis these judgments can be consid-
ered as equally relevant) and the position that the same 19 judgments occupy in
a ranking based on NA measures values. In both tables the number of comments
published in books or journals that each case received from legal scholars can be
found (the values reported are contained in the EUR-Lex XML file).

Table 1. Measures and attributes of the case study judgments.

Judgment DoctrineIn-degreeOut-degreeEccentricityClosenessBetweennessPageRank
VAN GEND EN LOOS 46 8 0 0 0 0 0.00049
COSTA 14 23 0 0 0 0 0.00111
INT.HANDELSGESELLSCHAFT 2 19 0 0 0 0 0.00120
COMM./COUNCIL(case 22/70) 20 37 0 0 0 0 0.00121
NOLD 3 13 0 0 0 0 0.00072
DASSONVILLE 9 117 0 0 0 0 0.0020
DEFRENNE 17 48 0 0 0 0 0.0018
SIMMENTHAL 35 49 1 2 1.666 3315.4763 0.00075
REWE-ZENTRAL 23 70 0 0 0 0 0.00194
CILFIT 19 22 1 2 1.5 1506.7138 0.00031
LES VERTS 11 14 2 1 1.0 2921.9004 0.00066
FOTO-FROST 13 20 2 2 1.777 9078.7165 0.00046
WACHAUF 3 20 1 2 1.666 2554.2996 0.00040
ERT 8 42 10 8 3.352 54365.1861 0.00069
FRANCOVICH 74 55 14 7 3.4431 54442.0397 0.00093
BOSMAN 82 140 21 10 4.0945 272540.6123 0.00139
MARTiNEZ SALA 14 36 0 0 0 0 0.00026
COMM./UK(case 466/98) 16 1 12 9 4.8045 0 0.00006
BAUMBAST 10 22 11 12 5.7630 26859.8971 0.00018
552 N. Lettieri et al.

Table 2. The cases in which the judgment stands in one of the top 20 positions of the
ranking calculated using the measure considered are highlighted in grey

Judgment ranking
Doctrine In-degree Out-degree Eccentricity Closeness Betweenness PageRank Eigenvector

(max: 52) (max: 67) (max: 44) (max: 22) (max: 5301) Centrality

VAN GEND EN LOOS 16 59 44 22 5301 34 939 614

COSTA 39 44 44 22 5301 34 35 203

INT. HANDELSGE- 50 48 44 22 5301 34 33 515

SELLSCHAFT

COMM./COUNCIL (case 32 30 44 22 5301 34 31 256

22/70)

NOLD 49 54 44 22 5301 34 443 288

DASSONVILLE 45 2 44 22 5301 34 6 6

DEFRENNE 36 19 44 22 5301 34 12 33

SIMMENTHAL 18 18 43 20 5296 1557 386 44

REWE-ZENTRAL 32 6 44 22 5301 34 10 13

CILFIT 34 45 43 20 33 2265 1360 209

LES VERTS 41 53 42 21 20 1699 551 763

FOTO-FROST 39 47 42 20 18 81 1043 263

WACHAUF 49 47 43 20 5296 1832 1193 234

ERT 44 25 34 14 380 882 502 35

FRANCOVICH 2 12 30 15 415 884 95 15

BOSMAN 1 1 23 12 655 2 23 1

MARTiNEZ SALA 39 31 44 22 5301 34 1444 174

COMM./UK(case 466/98) 36 66 32 13 1330 183 614 2679

BAUMBAST 42 45 33 10 3621 1791 1599 395

4 Discussion and Future Work

A brief analysis of the experiments and of the overall experience so far con-
ducted allows us to sketch some considerations. From a legal standpoint, we can
say that we have just scratched the surface: there is a long series of challenges to
deal with the essential contribution of legal experts in order to better understand
how the NA paradigm with its methodological and technical apparatus can shed
new light in our understanding of legal phenomena.
As our preliminary results clearly show, when applied to a “simple” citation
network, traditional NA measures have poor performances. Incongruences arise,
just to give an example, if you consider the rankings assigned by the measures
and the parameters listed in Table 2 to Van Gend en Loos: the most relevant
precedent in EU case law - according to the Court itself3 - shows very low value
of In-Degree centrality and 0 in all the other measures, although it collects a
relevant number of citations from legal literature. A first explanation of this
result could be related, among the other reasons, to the problem of “obliter-
ation by incorporation”, or OBI [18]. Borrowed from sociology of science and
well-known in the literature of citation analysis and sociology of the law, OBI
describes the phenomenon that typically occurs when an idea, such as a scientific
development, is considered so influential to make unnecessary to explicitly cite

3
http://curia.europa.eu/jcms/jcms/P 95693/.
 Network, Visualization, Analytics 553

its contributors. As highlighted by Garfield [18] while explaining the impact of

OBI on the usefulness of the explicit citations analysis,

In the course of this hypothesized process [obliteration by incorpora-

tion], the number of explicit references to the original work declines in the
papers and books making use of it. Users and consequently transmitters of
that knowledge are so thoroughly familiar with its origins that they assume
this to be true of their readers as well.

The phenomenon is not surprising in the legal field. The “authority” cit-
ing [19], i.e., the use of citations to provide an authoritative basis for a statement
in the citing work, is common in such field, especially in case-law system, where
judges typically cite a case to extend its “authority” to their own decisions.
However, it is not unusual that an important legal principle, affirmed by a given
judgment, became so commonly accepted that later the precedent expressing
it is not cited anymore. Therefore OBI can represent a major issue when the
citation analysis is the only method used for identifying the relevance of specific
documents over long run, since citations allow to detect only the explicit links
(i.e., an express reference), without any consideration for those just implicitly
contained in a document (e.g., the case reinterpreted as general principles of law
within an argumentative path). It is clear, against this backdrop, the importance
to better reflect on the ontology of the phenomenon investigated (e.g., relevance)
that is on its concrete expressions (data, relations) and on the algorithms more
suitable to identify it. Many aspects have to be deepened and taken into account.
Different studies, for example, suggest to combine the research on the citation
network of a given set of legal precedents with the network of scholarly opinions,
in order to obtain a more detailed picture of which decisions and verdicts should
be reckoned as relevant in a given legal system [20,21]. Another study [7] uses
legal journals as benchmark of the forecasts made starting from the analysis
of the decisions network. The main idea is that the analysis of cross citations
between judicial decisions could be inadequate to identify the relevance of a case,
so it should be integrated by the evaluations arising from the legal literature.
It is to be noted that there is a huge difference between citations by scholars
and by judges, which depends, first of all, on the reasons why both cite a case.
As highlighted by [21], judges typically cite other decisions only if they are
relevant for the cases that are accidentally brought to their attention. Therefore,
it may happen that a landmark case can be undervalued vis-a-vis less-relevant
decisions which are habitually cited in routine cases. Scholars indeed have a view
on legal relevance which is not confined to those decisions relevant for a specific
case under scrutiny. To define the relevance of a case they consider many other
elements, e.g., the political impact of a given decision or its popularity among
the people, usually ignored by the judges in deciding what cases they have to
cite. Therefore, the number of citations that a given document has received from
legal scholars could be considered just as a clue of its relevance, since it provides
no information about the reasons why the document was cited.
554 N. Lettieri et al.

Other approaches can be exploited in this respect. Some researchers [22,

23], have proposed to use semantics-based legal citation network for having a
better awareness of the arguments surrounding the citation that appears to be
fit to deal with the OBI phenomenon. Another approach worthy of attention is
represented by the combination of NLP and argument extraction (see, i.a. [24]),
a methodology that is opening up new and promising possibilities to support
both legal information retrieval and the identification of authoritative sources.
These considerations suggest to pay attention to the definition of the research
question namely the nature and the features of information to be extracted from
the network, in order to achieve a better understanding of the way in which the
network itself should be built.
From a technical standpoint, we have realized a freely available online tool4
that, even with room for improvements, is efficient enough to allow to domain
experts to make basic experiments with EU case law data. Regarding the future
developments, we will invite domain experts to use the system involving them
in the validation of both the tool and the methodology; to this aim, EuCaseNet
features will be extended in order also to allow the analysis of different legal
sources. Moreover, we will work on more refined operational definitions of legal
research questions (starting from the concept of relevance) to be addressed by
means of NA techniques. On the one hand, we will try to identify other data
sources and suitable metrics (e.g., drawing inspiration by the techniques which
are being developed in multimodal network analysis area [25,26]); on the other
we will implement more advanced techniques for the visualization of complex
structures that begin to be used in more other research areas [27,28].

References
1. Easley, D., Kleinberg, J.: Networks, Crowds, and Markets: Reasoning About a
Highly Connected World. Cambridge University Press, New York (2010)
2. Lettieri, N., Winkels, R., Faro, S.: Network Analysis in Law. ESI, Naples (2013)
3. Lettieri, N., Altamura, A., Malandrino, D.: The legal macroscope: experimenting
with visual legal analytics. Inf. Vis. 16(4), 332–345 (2017)
4. Sacco, R.: Legal formants: a dynamic approach to comparative law (installment I
of II). Am. J. Comp. Law 39(1), 1–31 (1991)
5. Lettieri, N., Altamura, A., Faggiano, A., Malandrino, D.: A computational app-
roach for the experimental study of EU case law: analysis and implementation.
Soc. Netw. Anal. Min. 6(1), 56:1–56:17 (2016)
6. Post, D.G., Eisen, M.B.: How long is the coastline of law? Thoughts on the fractal
nature of legal systems. J. Leg. Stud. 29, 545 (2000)
7. Fowler, J.H., Spriggs, J.F., Jeon, S., Wahlbeck, P.J.: Network analysis and the law:
measuring the legal importance of precedents at the US Supreme Court. Polit.
Anal. 15(3), 324–346 (2007)
8. Smith, T.A.: The web of law. San Diego Law Rev. 44, 309 (2007)
9. Kleinberg, J.M.: Hubs, authorities, and communities. ACM Comput. Surv. (CSUR)
31(4es), 5 (1999)
10. Malmgren, S.: Towards a theory of jurisprudential relevance ranking. Using link
analysis on EU case law, Graduate thesis, Stockholm University (2011)
4
http://www.isislab.it:20080/snam/index.php#one.
 Network, Visualization, Analytics 555

11. Katz, D.M., Bommarito II, M.J.: Measuring the complexity of the law: the United
States Code. Artif. Intell. Law 22(4), 337–374 (2014)
12. Winkels, R., Boer, A.: Finding and visualizing context in Dutch legislation. In:
Proceedings of NAiL 2013 (2013)
13. Koniaris, M., Anagnostopoulos, I., Vassiliou, Y.: Network Analysis in the Legal
Domain: A complex model for European Union Legal Sources. arXiv preprint
arXiv:1501.05237 (2015)
14. van Opijnen, M., Cristiana, S.: On the concept of relevance in legal information
retrieval. Artif. Intell. Law 25(1), 65–87 (2017)
15. Saracevic, T.: Relevance reconsidered. In: Information science: Integration in Per-
spectives. Proceedings of the Second Conference on Conceptions of Library and
Information Science, Copenhagen, Denmark, pp. 201–218 (1996)
16. Cosijn, E., Ingwersen, P.: Dimensions of relevance. Inf. Process. Manag. 36(4),
533–550 (2000)
17. Maduro, M., Azoulai, L.: The Past and Future of EU Law: The Classics of EU Law
Revisited on the 50th Anniversary of the Rome Treaty. Bloomsbury Publishing,
London (2010)
18. Garfield, E.: Citation Indexing: Its Theory and Application in Science, Technology,
and Humanities. ISI Press (1979)
19. Posner, R.A.: The Theory and Practice of Citations Analysis, with Special Refer-
ence to Law and Economics. Special Reference to Law and Economics (1999)
20. Agnoloni, T., Pagallo, U.: The power laws of the Italian constitutional court, and
their relevance for legal scholars. In: 28th International Conference on Legal Knowl-
edge and Information Systems. JURIX (2015)
21. van Opijnen, M.: Citation analysis and beyond: in search of indicators measuring
case law importance. In: JURIX (2012)
22. Zhang, P., Koppaka, L.: Semantics-based legal citation network. In: Proceedings of
the 11th International Conference on Artificial Intelligence and Law, ICAIL 2007,
pp. 123–130 (2007)
23. Panagis, Y., Sadl, U.: The force of EU case law: a multi-dimensional study of case
citations. In: 28th International Conference on Legal Knowledge and Information
Systems (JURIX) (2015)
24. Ashley, K.: Applying argument extraction to improve legal information retrieval.
In: ArgNLP (2014)
25. Du, N., Wang, H., Faloutsos, C.: Analysis of large multi-modal social networks:
patterns and a generator. In: Balcázar, J.L., Bonchi, F., Gionis, A., Sebag, M. (eds.)
ECML PKDD 2010. LNCS (LNAI), vol. 6321, pp. 393–408. Springer, Heidelberg
(2010). https://doi.org/10.1007/978-3-642-15880-3 31
26. Ghani, S., Kwon, B.C., Lee, S., Yi, J.S., Elmqvist, N.: Visual analytics for multi-
modal social network analysis: a design study with social scientists. IEEE Trans.
Vis. Comput. Graph. 19(12), 2032–2041 (2013)
27. Malandrino, D., Pirozzi, D., Zaccagnino, G., Zaccagnino, R.: A color-based visual-
ization approach to understand harmonic structures of musical compositions. In:
19th International Conference on Information Visualisation, IV 2015, Barcelona,
Spain, 22–24 July 2015, pp. 56–61 (2015)
28. De Prisco, R., Lettieri, N., Malandrino, D., Pirozzi, D., Zaccagnino, G., Zaccagnino,
R.: Visualization of music plagiarism: analysis and evaluation. In: 20th Interna-
tional Conference Information Visualisation, IV 2016, Lisbon, Portugal, 19–22 July
2016, pp. 177–182 (2016)
 Electronic Evidence Semantic Structure:
Exchanging Evidence Across Europe
in a Coherent and Consistent Way

Maria Angela Biasiotti(B) , Sara Conti, and Fabrizio Turchi

Institute of Legal Information Theory and Technique, 50127 Florence, Italy

mariangela.biasiotti@ittig.cnr.it
http://www.ittig.cnr.it/en/persone/ricerca/maria-angela-biasiotti

Abstract. In a cross-border dimension considering the specific collabo-

ration among European Union Member States related to criminal inves-
tigations and criminal trials, it becomes crucial to have a common and
shared understanding of what Electronic Evidence is and how it should
be treated in the EU context and in the EU MS. In this context the
EVIDENCE project developed a tailor-made categorization of relevant
concepts that allow to rely on a common and shared knowledge in this
domain. The categorization provides a starting analysis for the exchange
of Electronic Evidence and data between judicial actors and LEAs, with
a specific focus on issues of the criminal field and criminal procedures.
This semantic Structure might represent a good starting point for the
alignment of electronic evidence concepts all over Europe in across bor-
der dimension. This categorisation is significant as it is one of the few
initiatives to identify and classify relevant concepts in a domain, which
currently lacks of clear boundaries and touches upon different disciplines.

Keywords: Electronic evidence categorization

Electronic evidence exchange · Judicial cooperation · European Union

1 Introduction

Crime has become global, and almost all crimes involve electronic evidence.
A significant problem has become the exchange of data, across jurisdictions
and between the domestic participants in the criminal judicial process. Taking
this development and the problems into account, the EVIDENCE Project1 was
conceived. The project concluded that the European Union needs to develop a
better means to exchange information and evidence relating to crimes quickly
1
European Informatics Data Exchange Framework for Court and Evidence, (fund-
ing scheme: CSA (Supporting Action), Call ID FP7-SEC-2013-1; grant agreement
no: 608185; duration: 32 months (March 2014 – October 2016); coordinator: Con-
siglio Nazionale delle Ricerche (CNR-ITTIG), Italy; EU funding: Euro 1,924,589.00);
http://www.evidenceproject.eu.
c Springer Nature Switzerland AG 2018
U. Pagallo et al. (Eds.): AICOL VI-X 2015–2017, LNAI 10791, pp. 556–573, 2018.
https://doi.org/10.1007/978-3-030-00178-0_38
 Electronic Evidence Semantic Structure 557

from one country to another for the purpose of investigating crime in a timely
manner. The exchange becomes crucial in counterterrorism operations and when
dealing with global crimes. At the same time, a secure and trusted exchange of
information and of electronic evidence relating to crimes is an important element
in order to promote judicial cooperation in criminal matters, as well to contribute
to an effective and coherent application of EU Mutual Legal Assistance2 (MLA)
and European Investigation Order3 (EIO) procedures.
In a cross-border dimension considering the specific collaboration among
European Union Member States related to criminal investigations and crimi-
nal trials, it becomes crucial to have a common and shared understanding of
what Electronic Evidence is and how it should be treated in the EU context and
in the EU MS.
In this scenario, the categorization has been carried out taking into consid-
eration the “status quo” governing the collection, preservation and exchange of
electronic evidence at International and European Union levels.
In fact, Member States in Europe have basically different criminal legal sys-
tems and different tradition in the sources of criminal law.
That is, there is a lack of uniformity within Member States criminal leg-
islations, thus a common understanding in the electronic evidence domain is
required. The categorization of electronic evidence domain did not rely on a
simple comparative analysis and overview of national laws; but it just fostered
the building up of a common and shared “language” related to the handling
of electronic evidence, with the final aim of highlighting common requirements
being able to guarantee a uniform regulation of the use of electronic evidence
itself.
The Council of Europe and the European Union are, in this respect, two
supranational entities whose actions aim to develop a common legal substrate
and implement legislative harmonization between the various EU countries and
also with third States. Therefore, the Evidence project team took in due con-
sideration the most important documents and laws related to the criminal field,
issued by those entities.
With regard to electronic evidence, the Council of Europe Convention on
Cybercrime is highly relevant: although electronic evidence may not necessarily
2
European Convention on Mutual Assistance in Criminal Matters, Strasbourg,
20/04/1959, ETS No. 030; Council of Europe Convention on Laundering, Search,
Seizure and Confiscation of the Proceeds from Crime, Strasbourg, 08/11/1990, ETS
No. 141; Council of Europe Convention on the Transfer of Sentenced Persons, Stras-
bourg, 21 March 1983, ETS No. 112; Mutual assistance in criminal matters between
Member States, Council Act of 29 May 2000 establishing in accordance with Article
34 of the Treaty on European Union the Convention on Mutual Assistance in Crim-
inal Matters between the Member States of the European Union, 2000/C 197/01,
OJ C 197, 12.7.2000; Second Additional Protocol to the European Convention on
Mutual Assistance in Criminal Matters, Strasbourg, 8 November 2001, CETS No.
182; Council of Europe Convention on Cybercrime, Budapest, 23 November 2001,
ETS 185.
3
Directive 2014/41/EU.
558 M. A. Biasiotti et al.

origin only from cybercrime, this is the main framework for the categorization
domain as it offers many provisions to improve investigations where electronic
evidence is involved.
Section 1, Chapter II (substantive law issues) of the Convention has been very
relevant to the electronic evidence categorization point of view4 . In particular,
this Section deals with the definition of offences, grouped in 4 different categories
(offences against the confidentiality, integrity and availability of computer data
and system; computer-related offences; content-related offences, finally offences
related to infringements of copyright and related rights) related to the use of
computer networks and Internet.
For each of the mentioned categories, these offences are defined and described:
illegal access, illegal interception, data interference, system interference, misuse
of devices, computer-related forgery, computer-related fraud, offences related to
child pornography and offences related to copyright and related rights.
The categorization of the Electronic evidence domain based on these distinc-
tions: the different offences have represented a starting point to build up the
categorization itself, as the basis for the development of the “legal” classes (in
particular, the class Crime and the class Rule).
The categorization classes also based on several initiatives, at EU level,
directed to create a common framework to combat crime in general, and also
crimes in which electronic evidence is involved, by establishing a cooperation
between Member States.
Such as the Convention on mutual assistance in criminal matters5 , which aims
at improving the speed and efficiency of judicial cooperation between Member
States.
It is worth mentioning the European evidence warrant (EEW), which may be
used to obtain any objects, documents and data for use in criminal proceedings
for which it may be issued6 , aiming at facilitating speedier cooperation between
Member States in criminal proceedings, specifically in the transfer of evidence
(and, obviously electronic evidence) [3].
A significant role for the Electronic Evidence categorization has been also
played by the Directive 2013/40/EU of the European Parliament and of the
Council of 12 August 2013, on “Attacks against information systems”7 , and by
the Directive 2014/41 of the Parliament and of the Council of 3 April 2014, on
the “European Investigation Order” 4 (EIO)8 .
The Directive 2013/40/EU gives common definitions in the area of attacks
against information systems: offences of illegal access to an information system,
illegal system interference, illegal data interference and illegal interception. There
is the attempt to establish a common framework “of minimum rules concerning

4
Explanatory Report to the Convention on Cybercrime.
5
Council Act of 29 May 2000.
6
Council Framework Decision 2008/978/EU.
7
Directive 2013/40/EU.
8
Directive 2014/41/EU.
 Electronic Evidence Semantic Structure 559

the definition of criminal offences in that area”, in which electronic evidence

becomes of the utmost importance.
As for the EIO, this order should apply to all investigative measures aimed
at gathering evidence (and also, electronic evidence) and it will promote the
growth of a uniform and comprehensive system for obtaining evidence, in cases
with a cross-border dimension.
Finally, developing a common and shared language on the Electronic Evi-
dence domain has properly been the aim of the categorization team, as a starting
point for creating a unanimous framework for a coherent, consistent and uniform
application of new technologies in the collection, use and exchange of electronic
evidence in the criminal field.
According to this point of view, to make it possible, has been necessary to
build the “legal” classes of the categorization complying with the definitions of
the Convention on Cybercrime and in general with a specific view to all the EU
criminal legislation, as to have a common minimum standard of offences and
terms within the Electronic Evidence domain.
In this context the EVIDENCE project developed a tailor-made categoriza-
tion of relevant concepts that allow to rely on a common and shared knowledge
in this domain. The categorization provides a starting analysis for the exchange
of Electronic Evidence and data between judicial actors and LEAs, with a spe-
cific focus on issues of the criminal field and criminal procedures. Moreover, due
consideration is devoted to the impact of harmonized procedures on how police,
prosecutors and lawyers in criminal cases, and parties in civil cases, handle Elec-
tronic Evidence so as to preserve its integrity.
The proposed categorization offers a common terminology for the electronic
evidence domain bearing into account the twofold challenges to be considered
according to the Electronic Evidence lifecycle:

– the treatment of evidence/data/information by means of ICT in the context

of criminal law, considering the entire evidence lifecycle from the incident on,
passing through the investigative phase;
– the exchange of evidence/data/information by means of ICT complying with
the issues related to specific requirements and shared procedures.
The process of handling electronic evidence can be broken down under dif-
ferent phases (see Fig. 2), and the exchange process can occur during any of
these phases.

When electronic evidence is acquired and exchanged, information about its

provenance must be maintained to help establish authenticity and trustwor-
thiness. This information includes the origin of the evidence, its condition, and
any unique or distinctive characteristics. Any irregularities in the handling of
electronic evidence might spoil the probative value of the evidence, making it
not admissible.
All the relevant concepts related to these requirements have been included in
the Categorization.
560 M. A. Biasiotti et al.

2 Some Electronic Evidence Issues

The very nature of data and information held in electronic form makes it easier
to manipulate than traditional forms of data. When acquired and exchanged, the
integrity of the data must be maintained and proved, i.e. demonstrated that the
electronic evidence has not been altered since the time it was created, stored or
transmitted. Legislation on criminal procedures in many European countries was
enacted before the use of electronic evidence, although many Member States have
amended their legislation to accommodate the new form of evidence. However,
some issues remain, and include, but are not limited to, the following:
(i) In certain countries there are defined rules as to admissibility of evidence
in legal proceedings, while in other countries admissibility is flexible.
(ii) Legislation and policies may affect an investigation. For example, privacy
and data protection laws in some Member States may prevent the collec-
tion of evidence, and varied data retention periods across jurisdictions may
hinder investigations.
(iii) Legislation may furthermore not sufficiently address the realities of modern
investigations, especially when it comes to evolving new technologies9 .
The EVIDENCE Project concluded with the following results:
(i) The categorization of electronic evidence.
(ii) A survey on the legal position in handling and exchanging electronic evi-
dence in Europe.
(iii) A survey on the technical position in handling and exchanging electronic
evidence in Europe, along with a proposal for the representation of data
and meta data involved in the exchange process.
(iv) A plan for realising a Common European Framework for the Exchange
of Electronic Evidence. The plan provides a brief overview of the legal
issues; considerations relating to standards; technical aspects; law enforce-
ment requirements; the nature of the market, data protection issues, and
the challenges these topics pose. The aim of the Common European Frame-
work Framework for the Exchange of Electronic Evidence is to improve the
efficiency of investigations and judicial procedures while maintaining ade-
quate safeguards aimed at protecting relevant fundamental human rights
and respecting clear standards of conduct.

3 Preliminary Remarks on the Concept of “Electronic

Evidence”
There is such a wide variety of possible evidences that potentially anything may
become the evidence of a crime. Moreover, the introduction and the extensive
9
See EVIDENCE Deliverable 3.1 - Overview of existing legal framework in the
EU Member States: http://s.evidenceproject.eu/p/e/v/evidence-ga-608185-d3-1-
411.pdf and EVIDENCE Deliverable 3.2 - Status quo assessment and analysis of
primary challenges and shortcomings: http://s.evidenceproject.eu/p/e/v/evidence-
ga-608185-d3-2-412.pdf.
 Electronic Evidence Semantic Structure 561

use of ICT has additionally generated new forms of crimes or new ways of per-
petrating them, as well as a new type of evidences. This different setting implies
that, although all kinds of evidences have to be handled according to criminal
law and procedures, the ‘new’type of evidences need additional and specific ways
of handling. For instance, to give an initial suggestion of some of the new prob-
lems that arise, an increasing number of crimes (not only cybercrime) involve
geo-distributed Electronic Evidence and therefore an evidence ‘location’ needs to
be re-conceptualized, including issues concerning direct access to extraterritorial
data by LEAs (see [1,6,7,11]).
Electronic Evidences, as the traditional ones, have to be acquired and handled
following specific procedures that demonstrate their authenticity and integrity.
The fulfilment of these principles specifically applied to Electronic Evidence
requires the adoption of ad hoc procedures that demonstrate that the Electronic
Evidence has not been altered since the time it was created, stored or trans-
mitted. Such procedure has been developed in the field of digital forensics (see
[4,5]).
The idea to conceptualize the source of evidence and to outline the transition
from generic evidence to digital/Electronic Evidence establishing standards and
methods to assure authenticity over time also descends from diplomatics. This
field is providing an important contribution in this domain, by stating the prin-
ciple called “record trustworthiness”. This principle has two qualitative dimen-
sions: reliability and authenticity. Reliability means that the record is capable
of standing for the facts to which it attests, while authenticity means that the
record is what it claims to be. The trustworthiness of records as evidence is of
particular interest to EVIDENCE project, where Electronic Evidence is trust-
worthy being the result of the process of assessing its reliability and authenticity.
One of the main aims of the project was to develop a common and shared
understanding on what electronic evidence is, together with the relevant con-
cepts (digital forensics, criminal law, criminal procedure, criminal international
cooperation) as well as to draft a proposed ‘standard process’ occurring when a
crime occurs.
A definition of electronic evidence needs to be is broad enough to include all
kinds of evidence regardless of their origin. This was a particularly important for
the aim of the EVIDENCE Project, which focused on the exchange, as well as
on the harmonized handling of electronic evidence within a common European
framework. Based on these premises, the following definition is proposed:

Electronic Evidence is any data resulting from the output of an analogue

device and/or a digital device of potential probative value that are generated
by, processed by, stored on or transmitted by any electronic device. Digital
evidence is that Electronic Evidence which is generated or converted to a
numerical format.

The term data includes any analogical or digital item, because these items
may be the output of analogue devices or data in digital form.
562 M. A. Biasiotti et al.

In particular, it has been chosen to use the term data to include any analogical
and/or digital item specifying that these items may be the output of analogue
devices or data in digital form. With respect to this notion it must be considered
that Electronic data cover every method by which data are made available: over
the internet, stored on a computer, a smartphone, a separate hard disk, a CD-
ROM, a USB stick. There is no distinction to be made between data created
in analogue or digitized forms and the ones born digitally. Therefore scanned
images are also to be included in this term.
More specifically a print-out is merely a secondary evidence of the original
(or primary) digital version. A scanned version of a paper document is a sec-
ondary item of evidence because it is a copy. And it remains a copy, even if the
original is destroyed. A distinction relevant to the project categorization relates
to primary and secondary evidence (see [8,10]), also in relationship with physical
and Electronic Evidence. “In physical word the distinction between primary and
secondary evidence lies on the difference between the production of an original
document to prove a content and the submission of inferior evidence, such as
a copy of document termed, secondary evidence”. The primary evidence of a
document in electronic form is different from the primary evidence of a physical
object and this difference is significant for the EVIDENCE categorization. The
concept of primary or original in the electronic dimension is related to the hard-
ware or storage media where the file is kept. Then the print out of the document
on paper represents the secondary evidence in a human readable format. Mason
underlines that “even if the hard drive or storage device is correctly identified
as primary evidence, the physical items is of no value unless a person testifies to
its relevance and qualities that make it pertinent”. For this reason courts rely
on the production of the output of electronic data in a human-readable format,
printed on paper, which can be considered as a secondary evidence of Electronic
Evidence.
The term potential indicates that it has been considered the entire Electronic
Evidence lifecycle from the very start of the investigation process (the occurrence
of a criminal incident) to the handling and presentation in Courts. This allows us
to take in due consideration the exchange/transmission of Electronic Evidence
that may occur in any phase after the incident.
The term probative value highlights the aspect of relevance of the Electronic
Evidence (not its admissibility in courts that depends on the different jurisdic-
tions as well as on the adjudicator evaluation) and also indicates the fact-finding
process that transforms potential Electronic Evidence into an Electronic Evi-
dence.
Moreover, coherently with the inclusion of all forms of Evidence, the defini-
tion stresses on information that is manipulated, generated through, stored on
or communicated by any electronic device, thus including any type of computers
and/or information systems. Therefore, in the current definition, any physical
and analogue evidence, which has once been digitized (e.g. image sent by fax and
then printed again) is considered Electronic Evidence. These three dimensions of
the Electronic Evidence as conceived by the EVIDENCE Categorisation struc-
 Electronic Evidence Semantic Structure 563

ture takes in due consideration the basic distinction between evidence which was
born digital (e.g. data on a hard drive) and evidence which has been digitized
only later (see Fig. 1).
Categorization perspective considers:
– Digital evidence: the evidence that is originally born digital as created by any
digital device (computer or computer like-device);
– Analogical Evidence: the evidence originally in analogue format that enters
into the digitization process acquiring the digital status;
– Digitalized Physical evidences: the evidence that is born physical and entered
into the digitization process acquiring the electronic status.

Fig. 1. From source of evidence to electronic evidence

4 The Electronic Evidence Life Cycle

The process of handling electronic evidence can be divided in several phases. The
first phase includes the identification, collection and anti-contamination precau-
tions (searching the scene, collecting the evidence, packaging and labelling and
creating documents reporting the activities performed at every step) of elec-
tronic evidence. In the second phase, the acquisition of the source of evidence
takes place, determining which items are most likely to serve the purposes of
the investigation, which are the most time sensitive, which are most at risk of
being lost or corrupted, including the identification of similar issues. During
the third phase, the findings are evaluated and interpreted. The fourth phase
includes the presentation of the results in a report, which should include fac-
tual findings, interpretation, and expert opinion. The report and presentation
are essential steps in the electronic evidence lifecycle, because the court will
examine the report that should contain all relevant findings as well as technical
and non-technical explanations of the case and its issues. During each of the
phases involved in handling electronic evidence, it is essential to guarantee the
preservation of the evidence: every precaution must be taken when collecting
564 M. A. Biasiotti et al.

evidence, because any break in the process or improper handling of a procedure

could spoil the probative value of the evidence, potentially making it inadmis-
sible. The same principle applies to documenting procedures. Everything must
be documented: how the evidence was found, its condition, model and serial
numbers, markings, etc. The exchange may happen in different phases of the
electronic evidence lifecycle. Figure 2 illustrates the different phases of electronic
evidence.

Fig. 2. Electronic evidence life cycle

On the basis of the life-cycle outlined above, eight different concepts have
been identified by the EVIDENCE Project. The concepts have been organized
and classified as follows:

(i) Crime is an act, default or conduct prejudicial to the community, for which
the person responsible may, by law, be punished by a fine or imprisonment.
(ii) Sources of electronic evidence: comprise any physical, analogical and dig-
ital device (computer or computer like device) capable of creating infor-
mation that may have a probative value in legal proceedings.
(iii) A process is a series of actions or steps taken in order to achieve a particular
end within the electronic evidence lifecycle.
(iv) Electronic evidence is any information (comprising the output of analogue
devices or data in digital form) of potential probative value that is manip-
ulated, generated through, stored on or communicated by any electronic
device.
(v) A requirement represents principles or rules related both to legal rules and
handling procedures that are necessary, indispensable, or unavoidable to
make potential electronic evidence admissible in legal proceedings.
(vi) A ‘stakeholder’ (interested party) includes people or organizations having
a concern in or playing a specific role in the electronic evidence lifecycle.
(vii) A rule contains a set of explicit or understood regulations or princi-
ples governing conduct or procedures for the identification, collection,
 Electronic Evidence Semantic Structure 565

preservation, analysis, exchange and presentation of electronic evidence

in a cross border and national dimension.
(viii) Digital forensics is the application of forensic science to electronic evidence
in a legal matter.

These main classes have been hierarchically structured in sub-classes that

may be easily updated and maintained. The concepts are directly linked to
the class they refer to according to the EVIDENCE Project conceptual model
developed from the perspective of the life-cycle of electronic evidence.

5 EVIDENCE Semantic Structure: An Overview

Figure 3 shows the classes of the categorization reporting their main relationships
at a high level of detail. The core of the categorization is the class Electronic
Evidence that is connected with all the other classes. The Electronic Evidence
class is related with the class Source of Evidence (see also Fig. 4), through the
relation “is contained in”, is managed applying specific processes and is validated
according to legal and technical Requirements, generally studied in the different
disciplines of Digital Forensics; it concerns different types of crimes. Different
types of Stakeholders are concerned with the Electronic Evidence, who apply
specific rules, such as standards, soft and hard laws to examine it.

Fig. 3. EVIDENCE project categorization - main relationships at high level

6 EVIDENCE Semantic Structure Development:

Methodology

The following activities were carried out to develop the categorization:

1. Identification of information sources;

566 M. A. Biasiotti et al.

Fig. 4. Class source of electronic evidence, hierarchical view of main items

2. Extraction of relevant concepts, using bottom-up and top-down strategies;

3. Choice of a standard way to represent and support the categorization activi-
ties;
4. Structuring relevant concepts;
5. Review of the conceptual model/categorization internally.

The first step consisted in building the relevant bibliography on Electronic

Evidence domain, identifying sources of information relevant to the domain.
The sources were also collected in full text, mainly in PDF format, and then
transformed in text format for the automatic treatment foreseen later on by the
adopted methodology10 .
The sources are divided in:

– academic papers and monographs (45)

– guidelines (30 sources)
– report (13 sources)
10
Copyright note: documents included in the corpus have been downloaded using (a)
biblioproxy.cnr.it, the internal web platform of Italian National Research Coun-
cil to access commercial bibliographic resources, (b) downloaded from the Web if
freely accessible. Please note that the full text collected during the categorization
building-up activities had only research purposes and excluded any commercial use
of resources.
 Electronic Evidence Semantic Structure 567

– project reports (16 sources)

– legislative code of practice and regulation commentaries (16 sources)
– recommendation and policy document (8 sources).

6.1 Extraction of Relevant Concepts

In this phase relevant concepts of the Electronic Evidence semantic structure

were identified combining a top-down approach with a bottom-up approach.
This double strategy allowed to build up the EVIDENCE Categorisation on two
different layers: an upper layer made of relevant and abstract concepts identified
as “classes” and a lower layer populated by concepts of the Electronic Evidence
concrete reality, identified as “instance”. The connection between the two layers
is made through sub-classes partly extracted from the top-down strategy and
partly coming from the bottom-up extraction, as shown in Fig. 5.

Fig. 5. Top-down method, building up of classes, sub classes and instances

The Top-down approach allowed to extract relevant concepts from a subset of

the collected sources using a manual method. The selection of the most relevant
documents was carried out on the basis of following criteria:

– the relevance of the author;

– the relevance of the organization responsible for the publication: for example
the Association of Chief Police Officers (ACPO), the Council of Europe, the
National Institute for Standard and Technology, the National Institute of
Justice (US), the International Standard Organization;
– a direct evaluation of the document relevance within the scientific literature
of the domain.

The top-down method consisted in reading the selected set of documents

and identifying the most relevant concepts for each source and for the domain
568 M. A. Biasiotti et al.

of interest. This resulted in a first list of concepts. The different lists and results
were compared and this cross evaluation phase allowed to create a shared and
common list of relevant concepts as well as to assign to each concept a specific
weight.
The most common concepts, identified in this phase, have been included in
the Categorization (http://www.evidenceproject.eu/categorization).
Then some main classes have been identified and many concepts, already
selected, were assigned to those classes as subclasses or instances. In order to
verify whether the terms selected in the top down approach were those gener-
ally used in the specialist domain a bottom-up strategy was implemented. This
allowed also to verify the set of concepts manually collected and increase the
number of terms to be taken into consideration.
The bottom-up phase adopted a semi-automatic concept extraction using a
Natural Language Processing basic technique. This parallel activity was per-
formed on a subset of documents gathered by all project partners, specifically
on 91 documents out of the previously collected 128 for the following reasons:

– Documents 122 to 128 were too technical;

– Documents 35, 36, 37, 45, 86, 118 were not in English;
– Documents 22, 26, 39 had no pdf full version available.

The automatic extraction relied on a lemmatization process determining the

lemma for each word found in the documents corpus. In particular the extrac-
tion process has exploited the WordNet Lemmatizer that uses the WordNet
Database11 to lookup lemmas12 . Also a stop word list has been developed in
order to eliminate articles, prepositions, etc. The result of this step was the
identification of a list containing the main concepts of the Electronic Evidence
domain based on the relevance of the concepts identified as well as the context
where the concepts were extracted. In this phase a collection of definitions of
the relevant concepts were also gathered and analysed.
Automatic systems are based on statistical methods that generally require
a high number of documents to perform a meaningful linguistic analysis. Even
if our corpora was not so large, it was sufficient to determine the most relevant
terms in the collection, using an information retrieval technique. For this aim
it has been used the Term Frequency-Inverse Document Frequency (TF-IDF )
formula to discover which terms were more relevant for the Electronic Evidence
domain. TF-IDF weight is a method often used in information retrieval and text
mining. This weight is a statistical measure used to evaluate how important a
word is within a document or in a collection or corpus. The importance increases
proportionally to the number of times a word occurs in the document but it is
offset by the frequency of the word in the corpus.
11
http://wordnet.princeton.edu/.
12
A lemma is a canonical form of the word that may appear in several inflected forms.
For example, in English, the verb ‘to investigate’ may appear as ‘investigate’, ‘inves-
tigated’, ‘investigates’, ‘investigating’. The base form, ‘investigate’, that one might
look up in a dictionary, is called the lemma for the word.
 Electronic Evidence Semantic Structure 569

On the basis of the previous considerations, concerning the reliability of

the automatic methods, all lemmas automatically extracted (more than 21.000
terms) were reduced to about 16.000 lemmas. They were then manually exam-
ined in order to detect the terms not included in the top-down methods and also
to distinguish and select concepts belonging to the specific language domain.
Moreover terms belonging to common language were also eliminated. The result-
ing list of terms was refined and then merged with the previous concepts, obtain-
ing a list of about 220 terms, setting A few of the new terms selected during the
bottom-up process are shown in Table 1.

Table 1. Terms selected during the bottom-up process

Lemma Occurrences
Cybercrime 3.341
Requirement 905
Seize 854
Integrity 636
Preservation 630
Fraud (online fraud) 450
Judge 448
Prosecutor 435
Label (labeling) 293
Authenticity 268
Police officer 134
Phishing 127
Encase (encase forensics) 122

Finally, to exploit the statistical data gathered by the bottom-up method, it

has been extracted the occurrences of the main syntagms identified during the
top-down strategy.
In other words the syntagms have been extracted in a manual way and later
each item of the list has been used for an automatic processing in order to
establish the related frequency.
Table 2 shows the result of this processing.

6.2 Choice of a Standard Way to Represent and Support The

categorization Activities
In order to represent the categorization, it has been chosen SKOS (Simple
Knowledge Organization System)13 that “provides a model for expressing the
13
SKOS Simple Knowledge Organization System Primer, W3C Working Group Note
18 August 2009, http://www.w3.org/TR/skos-primer.
570 M. A. Biasiotti et al.

Table 2. Occurrences of syntagms from EVIDENCE project corpus

Syntagm Occurrences
Digital evidence 1.933
Electronic Evidence 1.076
Digital forensic(s) 926
Child pornography 806
Computer forensic(s) 613
Chain of custody 196
Incident response 176
Forensic(s) examiner 165
Law enforcement agency(ies) 161
Best practice(s) 146
Hash value 138
Police officer(s) 134
Personal computer 109
Cyber crime(s) 61
Expert witness 54
Digital evidence specialist(s) 53

basic structure and content of concept schemes such as thesauri, classification

schemes, subject heading lists, taxonomies, ‘folksonomies’, other types of con-
trolled vocabulary, and also concept schemes embedded in glossaries and termi-
nologies”. SKOS has the advantage of expressing knowledge organization systems
in a machine-understandable way within the framework of the Semantic Web.
It was introduced by the W3C and has become a W3C recommendation, based
on other Semantic Web standards such as RDF and OWL. The main element of
the SKOS vocabulary is the concept. “Concepts are the units of thought—ideas,
meanings, or (categories of) objects and events—which underlie many knowledge
organization systems. As such, concepts exist in the mind as abstract entities
which are independent of the terms used to label them14 ”. Moreover, “concepts
can be identified using URIs, labelled with lexical strings in one or more natu-
ral languages, assigned notations (lexical codes), documented with various types
of note, linked to other concepts and organized into informal hierarchies and
association networks, aggregated into concepts schemes, grouped not labelled
and/or ordered collections, and mapped to concepts in other schemes” 15 . To
implement the categorization with SKOS, it has been used, iQvoc a vocabulary
management tool16 that combines easy-to-use human interfaces with Semantic
Web interoperability.
14
http://semanticweb.com/introduction-to-skos b33086.
15
http://www.w3.org/TR/skos-primer.
16
http://iqvoc.net/.
 Electronic Evidence Semantic Structure 571

According to the knowledge management system chosen, for each identified

term the following features were provided:

– A Label: SKOS allows the use of multiple labels, one single preferable and
multiple alternatives, nevertheless this feature has been used only for a limited
set of concepts;
– A singular Definition, even though SKOS allows the use of multiple definitions
for the same concept;
– Structuring concepts in classes and sub classes using hierarchical relationships
– An Editorial Note where all reliable sources, are specified to determine the
single definition of the concept. In particular it has been used:
• “team definition” to specify definition set by the team that worked on
the categorization task;
• “Definition based on...” to indicate that an original source was adopted
with minor changes and/or adapted to the domain;
• “E. Casey (see [2]) etc.” to indicate that it has been cited exactly the
definition provided in the source.
– Scope Note that explains and clarifies the meaning of the term and describes
its intended use as a subject heading.
– a Relation term that explains the type of association between the defined term
and other terms belonging to other classes and subclasses of the categorisa-
tion. The type of relation is specified for each association. When relationships
are transitive, associations are reported accordingly both in related and relat-
ing class (i.e. Forensic examiner uses Digital forensic tool; Digital forensic tool
is used by Forensic examiner).
– an Historic Note in order to handle the changes of the concepts definitions and
to keep track of all changes affecting the definition or scope of that specific
concept during the entire Project lifecycle.

The EVIDENCE project does not take into consideration a multilingual ver-
sion of the proposed categorization due time scheduling allocation. Moreover a
linguistic alignment of concepts in different languages would require not only
a translation of terms but also the compliance with specific legal and judicial
order as well as with the specific context in which those concepts are placed
(legislation, case law, European, domestic, etc.). Every concept identified has
been represented using, at minimum, the following data model:
Description/Preferred label (obligatory, unique): the name of the concept;
Definition of the concept (one is obligatory, more are optional);
Scope Notes an extended explanation of the concept and its context of use
(optional, one or more).
An example of the data model is given in Table 3:
572 M. A. Biasiotti et al.

Table 3. Example of concept acquisition in SKOS

Description Acquisition
Definition The process of creating a copy of data within a defined set
Scope note The Digital Evidence First Responder should adopt a suitable
acquisition method based on the situation, cost and time, and
document the decision for using a particular method or tool
appropriately. Digital evidence, by its very nature, is fragile
and can be altered, damaged or destroyed by improper
handling or examination. Examination is best conducted on a
copy of the original evidence. The original evidence should be
acquired in a manner that protects and preserves the integrity
of the evidence. (ISO 27037)
Editorial note Definition based on the ISO/IEC 27037:2012

7 Conclusions
This semantic Structure might represent a good starting point for the alignment
of electronic evidence concepts all over Europe in a cross border dimension. This
categorisation is significant as it is one of the few initiatives to identify and
classify relevant concepts in a domain, which currently lacks of clear boundaries
and touches upon different disciplines. When the activities of the EVIDENCE
Project started, the knowledge on this domain and the awareness about it were
very limited and a few persons were able to speak about it in a comprehensive
way. Even actors directly involved in the treatment of electronic evidence by
default (public prosecutors, LEAs and judges) demonstrated real important gaps
and challenges in their knowledge and training.
The Categorization has been mainly exploited for the Evidence project activ-
ities in order to determine a common terminology over the different work pack-
ages. The future steps would consist of turning this first result into an ontology
fruitful to the digital forensic community. This entails to compare to other Euro-
pean projects related to the digital forensics domain and the electronic evidence
exchange among the involved/interested stakeholders.
Furthermore it will be essential to take into account the differences in national
legislation within the EU member States, and the evidence rules amongst coun-
tries with similar legal tradition, an issue that would primarily find a solution
through legislative and/or policy action. Policy makers should define an efficient
regulation for the treatment and exchange of electronic evidence.
Another essential feature is to meet the need to develop definitions that are
future-proof in order not to be affected by technological developments.
The status quo at the beginning of the EVIDENCE project was “I know
electronic evidence exists, I know I cannot make it without but I don’t know how
to deal with it and treat and handle it without compromising it...”.
Now that the cornerstone has been put in place the future work to be done
will be to migrate this simple categorization into a more sophisticated way of
 Electronic Evidence Semantic Structure 573

managing concepts and their semantic relationship such as the ontological ones
and to match this semantic structure with others already existing in comple-
mentary domain such as digital forensics (see [9,12]).

References
1. Association of Chief Police Officers UK. Good Practice Guide for Digital
Evidence (2012). http://www.digital-detective.net/digital-forensics-documents/
ACPO Good Practice Guide for Digital Evidence v5.pdf
2. Casey, E.: Digital Evidence and Computer Crime. Forensic Science, Computers,
and the Internet, vol. XXVII, 3rd edn, p. 807. Elsevier, Amsterdam (2011). ISBN
9780123742681
3. Murphy, C.C.: The European evidence warrant: mutual recognition. In: Eckes, C.,
Konstadinides, T. (eds.) Crime Within the Area of Freedom Security and Justice:
A European Public Order. Cambridge University Press, Cambridge (2011)
4. ĆosiĆ, J., ĆosiĆ, Z.: An ontological approach to study and manage digital chain
of custody of digital evidence. J. Inf. Organ. Sci. 35(1), 1–13 (2011)
5. ĆosiĆ, J., ĆosiĆ, Z.: The necessity of developing a digital evidence ontology. In:
Proceedings of the 23rd Central European Conference on Information and Intelli-
gent Systems, pp. 325–230. University of Zagreb (2012)
6. Council of Europe: Electronic Evidence Guide. A basic guide for police officers,
prosecutors and judges (2014)
7. ISO/IEC 27037: Guidelines for identification, collection, acquisition, and preserva-
tion of digital evidence (2012)
8. Mason, S.: Electronic Evidence, p. 934. LexisNexis-Butterworths, London (2012).
ISBN 9781405779876
9. Park, H., Cho, S.H., Kwon, H.-C.: Cyber forensics ontology for cyber criminal
investigation. In: Sorell, M. (ed.) e-Forensics 2009. LNICST, vol. 8, pp. 160–165.
Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-642-02312-5 18
10. Schafer, B., Mason, S.: The characteristics of digital evidence. In: Mason, S. (ed.)
Electronic Evidence, p. 25. LexisNexis Butterworths, London (2012)
11. SWGDE: Digital evidence: standards and principles, forensic science communica-
tions, vol. 2, no. 2, p. 2 (2000). www.swgde.org
12. Talib, A.M., Alomary, F.O.: Toward a comprehensive ontology based-investigation
for digital forensics cybercrime. Int. J. Commun. Antenna Propag. 5(5), 263–268
(2015)
 Author Index

Abdulrab, Habib 394 Francesconi, Enrico 257

Adebayo, Kolawole John 337 Futia, Giuseppe 380
Al Khalil, Firas 105
Andrade, Francisco C. P. 233 Gomes, Marco 528
Araszkiewicz, Michał 187 Gómez-Pérez, Asunción 410
Arisaka, Ryuta 427 Górski, Łukasz 325
Aryani, Amir 245 Governatori, Guido 287
Gracia, Jorge 410
Baehrens, David 205 Grossmann, Georg 124
Bainbridge, Bridget 124
Batouche, Brahim 257 Hajlaoui, Najeh 257
Bianchi, Ilaria 301 Hausstein, Brigitte 245
Biasiotti, Maria Angela 556 Hong, Kihyun 365
Boella, Guido 337
Boer, Alexander 39, 89 Jouvelot, Pierre 220
Branting, L. Karl 465 Julian, Vicente 233
Brown, Bradford 465
Butler, Tom 105 Khalil, Mohamad 394
Klas, Claus-Peter 245
Canova, Lorenzo 380 Kłodawski, Maciej 187
Carrabina, Jordi 140 Klügl, Peter 205
Casanovas, Pompeu 1, 124, 140, 245, 350, Kollingbaum, Martin 317
410 Konstantinov, Hristo 272
Ceci, Marcello 105
Cervone, Luca 301 Lagioia, Francesca 513
Conti, Sara 556 Lettieri, Nicola 543
Contissa, Giuseppe 513 Li, Wenhao 124
Costa, Angelo 233 Liepina, Ruta 172
Lim, Seungtak 365
Dallmeier-Tiessen, Sunje 245 Lippi, Marco 513
De Koker, Louis 124 Liu, Jixue 124
Di Caro, Luigi 337 Lorenzetti, Tiziano 257
Dragoni, Mauro 287
Draicchio, Francesco 301 Malandrino, Delfina 543
Manghi, Paolo 245
Matthes, Florian 157
El Ghosh, Mirna 394 Mayer, Wolfgang 124
Melandri, Alessio 380
Faggiano, Armando 543 Mendelson, Danuta 124
Faro, Sebastiano 543 Merkhofer, Elizabeth 465
Fawei, Biralatei 317 Mimouni, Nada 451
Fiorelli, Manuel 257 Moguillansky, Martín O. 495
576 Author Index

Morando, Federico 380 Simari, Guillermo R. 495

Moreira, Teresa C. 233 Simon, Christian 205
Simon, Kai 205
Naja, Hala 394 Simov, Kiril 272
Nazarenko, Adeline 451 Simova, Iliana 272
Novais, Paulo 233, 528 Soh, Chiseung 365
Stellato, Armando 257
O’Brien, Leona 105 Stumptner, Markus 124
Osenova, Petya 272
Tomanek, Katrin 205
Pagallo, Ugo 1, 23 Torroni, Paolo 513
Palmirani, Monica 1, 72, 301 Turbati, Andrea 257
Pan, Jeff Z. 317 Turchi, Fabrizio 556
Paschke, Adrian 325 Tyankov, Tenyo 272
Pin, Benoit 220
Poblet, Marta 245 Unsworth, Kathryn 245
Poudyal, Prakash 443
van der Torre, Leendert 350, 427
Ramakrishna, Shashishekar 325 van Doesburg, Robert 478
Reis, Andre 140 van Engers, Tom 39, 89, 478
Reis, Simone R. N. 140 Vestoso, Margherita 543
Reschenhofer, Thomas 157 Vibert, Hughes-Jehan 220
Rhim, Young-Yik 365 Villata, Serena 1, 287
Rizzi, Williams 287
Rodríguez-Doncel, Víctor 245, 350, 410 Waltl, Bernhard 157
Rossi, Arianna 72 Wang, Jingbo 245
Ruggiero, Francesco 380 Watts, David 124
Weiss, Brandy 465
Salotti, Sylvie 451 Wyner, Adam 172, 317
Santos, Cristiana 350, 410
Sartor, Giovanni 1, 172, 513 Yeh, Alexander 465
Satoh, Ken 427 Yelshyna, Aliaksandra 233
Schmedding, Florian 205 Yoshino, Hajime 56
Schmitz, Peter 257
Sileno, Giovanni 39, 89 Zeleznikow, John 528