Query Answering with Soft Entity Constraints

This is the code repository accompanying the paper Interactive Query Answering on Knowledge Graphs with Soft Entity Constraints.

Paper | Artifacts

Logical queries over knowledge graphs allow retrieving entities that meet constraints defined by logical formulas. In this project, we extend these with soft constraints that allow specifying that in addition to the logical constraints, an entity should be "like" or "unlike" specific exemplary entities:

If you find our work interesting, please use the following citation:

@article{
daza2026interactive,
title={Interactive Query Answering on Knowledge Graphs with Soft Entity Constraints},
author={Daniel Daza and Alberto Bernardi and Luca Costabello and Christophe Gueret and Masoud Mansoury and Michael Cochez and M.C. Schut},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=Qb6vIM7MxE}
}

Use the following to create an environment with conda and install the requirements:

conda create -n nqr python=3.12
conda activate nqr
pip install -r requirements.txt

In our work, we introduce query-answering datasets extended with soft entity constraints. We provide these datasets on Zenodo:

https://doi.org/10.5281/zenodo.20847352

Download and extract the dataset archives from the repository root:

• fb15k237-betae.tar.gz
• hetionet.tar.gz

mkdir -p data

tar xzf fb15k237-betae.tar.gz -C data
tar xzf hetionet.tar.gz -C data

The same Zenodo record also contains pretrained artifacts and raw experimental results, which can be used to reproduce the experiments and figures without retraining:

• fb15k237-betae_10_0.0002_nonforced.tar.gz
• hetionet_10_0.001_nonforced.tar.gz
• results.tar.gz

After extraction, the artifacts are organized as follows:

data/        Extended datasets with soft constraints
neural_adj/  Pretrained reranking artifacts
results/     Raw outputs used for the paper figures and tables

Extract them with:

mkdir -p neural_adj

tar xzf fb15k237-betae_10_0.0002_nonforced.tar.gz -C neural_adj
tar xzf hetionet_10_0.001_nonforced.tar.gz -C neural_adj

tar xzf results.tar.gz

The main entrypoint is the module nqr.qto.query. Given a dataset and a model, the module can train and test different methods for query answering with soft entity constraints. We provide configuration files for all our experiments in the configs directory. As an example, to test the Cosine method on the FB15k237 dataset, run

python -m nqr.qto.query --config configs/fb15k237/cosine.yaml

The generated results will be stored in the results directory.

Since we provide raw experimental results, it is also possible to replicate the figures and tables in our paper without the need to repeat the experiments. To replicate Fig. 1, run

python analysis/plots.py

To replicate Table 3, run

python analysis/tables.py