A Global Supply Chain Risk Management Framework:

An Application of Text-mining to Identify Region-specific

Supply Chain Risks
Chih-Yuan Chua, Kijung Parkb and Gül E. Kremera*
a
Department of Industrial and Manufacturing Systems Engineering, Iowa State University, 2529
Union Drive, Ames, IA 50011, USA
b
Department of Industrial and Management Engineering, Incheon National University, 119
Academy-ro, Yeonsu-gu, Incheon 22012, Korea

Abstract

Nowadays global supply chains enable companies to enhance competitive advantages,

increase manufacturing flexibility and reduce costs through a broader selection of
suppliers. Despite these benefits, however, insufficient understanding of uncertain
regional differences and changes often increases risks in supply chain operations and
even leads to a complete disruption of a supply chain. This paper addresses this issue
by proposing a text-mining based global supply chain risk management framework
involving two phases. First, the extant literature about global supply chain risks was
collected and analyzed using a text-based approaches, including term frequency,
correlation, and bi-gram analysis. The results of these analyses revealed whether the
term-related content is important in the studied literature, and correlated topic model
clustering further assisted in defining potential supply chain risk factors. A risk
categorization (hierarchy) containing a total of seven global supply chain risk types and
underlying risk factors was developed based on the results. In the second phase,
utilizing these risk factors, sentiment analysis was conducted on online news articles,
selected according to the specific type of risk, to recognize the pattern of risk variation.
The risk hierarchy and sentiment analysis results can improve the understanding of
regional global supply chain risks and provide guidance in supplier selection.

Keywords: Global supply chain; Risk management; Data analytics; Text mining;
Sentiment analysis; Google News

1. Introduction

Global supply chains have been widely discussed to boost competitive advantages
[1]. From an economic perspective, global supply chain management can create a win-
win situation for associated stakeholders in that every role (or node) in the global supply
chain network can obtain operational benefits when all stages related to the
product/component production (e.g., design, manufacturing, assembly, testing, and
marketing) is coordinated worldwide [2]. Although operating global supply chains has
been considered as a common profit advantage in most industries, uncertain regional
factors regarding global business and logistics such as government stability, insufficient
infrastructure in a country, and trade barriers may make it difficult to have successful
supply chain management [3, 4].

A global supply chain network can be interpreted as an integration of both

internal and external stakeholders. This integration has positive impacts on corporate
operations such as innovative strategy, product quality and profitability [5]. However,
critical risks and challenges in global business operations may lead to the disruption of
the global supply chain although suppliers in multiple tiers and various nations are fully
coordinated to decrease the total supply chain cost. Christopher et al. [6] proposed a
classification for global sourcing risks that include supply risk, environmental and
sustainability risk, process and control risk, and demand risk. Among these risks, supply
risk, process and control risk, and demand risk have been widely discussed in practice
and academia because they are relatively easier to evaluate through quantitative
methods common to domestic supply chain cases. From a global business dynamics
vantage point, on the other hand, environmental and sustainability risks such as
fluctuations in environmental protection legislations, taxation differences, and political
stability are difficult to analyze and evaluate due to related uncertainties affected by
local political, social and economic realities. This insufficient understanding of regional
– often across national boundaries – differences and thus changes may significantly
influence and disrupt the entire supply chain performance [6]. One significant current
challenge in supply chain risk management is still about modeling and quantifying the
risks [7]. A comprehensive risk categorization is needed to better understand the scope
of global supply chain risk factors due to regional differences.

The above mentioned issues lead to the research questions: 1) What are the
significant global supply chain risks due to regional differences? 2) How can such risks
be identified and organized objectively? 3) How can the risk variation (factors and
levels) patterns be recognized? Although there are previous studies on global supply
chain risk management, most research applied case studies, operational-modelling or
probability-related methodologies to categorize risks and address supply chain design
problems [8, 9, 10]. Moreover, studies using qualitative literature review methods to
classify global supply chain risks have challenges in validating due to their limited
sample [11]. Therefore, a data-driven method is needed for researchers and practitioners
to identify global supply chain risks and understand the significant risk patterns [12].

Artificial intelligence has been increasingly applied in supply chain risk

management area; nevertheless, very few of these studies incorporated text analytics to
extract useful information on supply chain risks [13]. To identify regional risks of a
global supply chain by a data-driven approach, input data collection is the first and most
important step to undertake. However, there is a great deal of text information on
geographical differences and supply chain risks in published articles, business
magazines, and news articles. General public is also able to receive information about
natural disasters, government policies, and company operations from news media.
Therefore, this study attempts to discover how this information can be exploited to
reveal signals on global supply chain risks. Nowadays, such text-based information
could be easily accessed through open portals or databases for academic purposes.
Research databases such as Scopus, Engineering Village, and Google Scholar have a
wide variety of peer-reviewed research articles, which are organized for retrieval by
search algorithms. Published news articles are also stored on publisher’s website or
database. Platforms such as Google News or Yahoo News collect the latest news articles
from different mass communication media. These exemplify the abundance of
information and sharing mechanisms, and demonstrates that online news media have
already become a key element of social, economic, and cultural life worldwide [14].
This trend presents the possibility that such information can be sources of input to this
research; peer-reviewed journal articles can help define important risk factors with their
careful review processes, and online news articles aid in risk variation pattern
recognition because of their content and time-specific relevance.

To analyze these articles directly for discovering valuable insights about regional
supply chain risks, a text-oriented, unstructured data analysis method is required.
Therefore, text mining is adopted to extract information from the input documents. Text
mining is an approach that can analyze a large amount of documents (in a form of pre-
processed text format) and extract valuable information. Text mining techniques have
been applied to numerous areas such as political document analysis, energy and service
industries, information technology, and healthcare service sectors due to their
outstanding capability in analyzing unstructured data [15]. It is extensively applied in
knowledge discovery, retrieval, and management. Artificial intelligence is also
implemented in more advanced text-mining tools to analyze textual data to discover
patterns and insights in unstructured texts [10].

This study addresses the above stated research questions by developing a text-
mining-based risk management framework for global supply chains that consider up-
to-date text information. This framework could provide guidance for companies to
develop a resilient global supply chain considering the influences of regional
differences and changes.
 The remainder of this paper is organized as follows. Section 2 discusses previous
studies of global supply chain risk management and data analytics (including text
mining) applications. Section 3 demonstrates the proposed research framework. Section
4 presents the implementation and results of defining potential risk factors and risk
variation pattern recognition incorporating sentiment analysis. A discussion on
relevance to previous research and limitations of this study is also presented in this
section. Finally, the summary of the proposed framework, highlights of contributions,
and future research directions are included in Section 6.

2. Literature review

In this section, previous studies were investigated in two categories. First category
reviews the studies that discussed risk management in the context of global supply chain.
The second one illustrates the development of data analytics and its applications in
supply chain management. At the end of this section, a brief summary pointing to the
research gaps is presented to highlight the importance of the research question tackled
herein.

2.1. Global supply chain risk management

Due to the improvement of information sharing technologies and the trend of

globalization, domestic supply chains of most enterprises have evolved into global
supply chains by integrating multi-nation-resources to their current operations [16]. In
spite of the advantages of global supply chains, such as access to new markets,
capability to reduce cost, etc., there are numerous potential issues that remain to be
overcome by practitioners and scholars. In comparison to a relatively simple domestic
supply chain, the complexity of global sourcing frequently results from additional
regional risks such as taxes, export regulations, exchange rates, etc. [17]. Moreover,
diversity in culture, languages, and political environment are among the reasons why a
global supply chain is incredibly difficult to operate at a high performance level [16].
Therefore, understanding regional risks is crucial for region selection process in global
supply chain design.

Christopher et al. [6] classified global supply chain risks as supply risk,
environmental and sustainability risk, process and control risk, and demand risk. The
environmental and sustainability risks contain political instability, export regulations,
tariffs, etc., which would have negative impacts on economic, social, and sustainability
aspects of an enterprise. Ho et al. [18] suggested that global supply chain risks can be
categorized into macro and micro risks. Macro risks comprise factors such as natural
disasters, political environment, and regional regulation; while micro risks refer to
demand risk, supply risk, manufacturing risk, and infrastructure risk [19, 20, 21].

Risk assessment methods have also been discussed. Aloini et al. [9] proposed risk
identification and risk quantification. The risk identification involves factors that would
lead to potential disruption of the network (i.e., risk factors) and their possible effects
on performance should be determined. For their assessment, risk factors should also be
prioritized quantitatively according to their occurrence probabilities, severity, and
detection difficulties. For example, Tsai et al. [22] adopted the analytic hierarchy
process (AHP) to implement risk calibration on outsourcing retail chains in Taiwan.
Trkman and McCormack [23] built a risk identification conceptual model based on
supplier attributes, performance, and the specific environment of operation. Kayis and
Karningsih [24] proposed a supply chain risk identification system constructed under a
knowledge-based system tool technology.

Although most of the above risk identification models demonstrate strengths on

specific cases or types of risk, they were mostly qualitative, with complicated processes
and not able to associate risk factors to cost implications. Therefore, a data-driven
approach that can systematically deal with unstructured data (e.g., text data) is required
to improve global supply chain risk management issues.

2.2. Data analytics applications in supply chain management

Data analytics has become increasingly crucial in numerous fields during the last
decade. With the improvement of information communication technologies, data can
be accessed and analyzed in a more transparent fashion. Data-driven decisions enable
enterprises to manage operations based on evidence instead of speculation or intuition
[25]. There are companies that started manipulating and extracting information from
data such as Google and Amazon, but the potential benefits of applying data analytics
in other industries might be even greater. One of the proofs is the phenomenon that
more and more organizations are training and recruiting data scientists to deal with data
related to transactions, operations, human resources, etc. [26, 27]. When decision
making is no longer about intuition but evidence-based solution, and with this concept
flourishing in most of the companies, the positive relation between data science and
supply chain management could easily be anticipated.

In data science for supply chain management, the key is to understand the data you
are collecting. The categories of data can be listed as sales, consumer, inventory, and
location and time [28]. This categorization considers not only the type of data, but also
the volume, update frequency, and source variety of data. Although there are different
characteristics of data in supply chain management, analytics applications can extract
useful information and create competitive advantages throughout supply chain: from
location-based marketing to distribution and logistics to supplier risk management [29].

Data analytics has been applied to supply chain management field considering
various operations, such as procurement, manufacturing, warehousing, logistics, and
demand management. Nguyen et al. [30] published a review article discussing those
applications and the adopted analytics methods to analyze the trends of big data analysis
in supply chain management. The article demonstrates how methods such as
optimization, classification, simulation, and clustering were implemented to help
improve supply chain management. For instances, Choi et al. [31] developed a novel
approach by using fuzzy cognitive map to solve supplier selection problems in IT
service procurement; Huang and Handfield [32] adopted supply maturity model and
information from ERP vendors’ public information (e.g., news articles, company
announcement, and white papers) to deal with sourcing risk problems; Wu et al. [33]
aggregated various data from Taiwanese light-emitting diode industry and used fuzzy
and grey Delphi methods to investigate supply chain risks and uncertainties. The above
summarized studies demonstrate the potential of data analytics in supply chain issues.

2.3. Text mining and its applications in supply chain management

From the review of the above mentioned previous studies, it is deduced that global
supply chain risk management has been discussed as an important issue in academia
and industry. The need for data-driven methods has also been emphasized by numerous
applications of data analytics in the supply chain field. However, while dealing with
risk management problems, a variety of sources related to specific risks involve mostly
text data from social media and online news, for example, on natural disasters, the
political environment, or tax regulations. This subcategory of data analytics is referred
to as text mining.

Text mining is a data analysis method to analyze unstructured text data and extract
useful information. Both descriptive analysis and predictive analysis are applicable in
text mining. The subtasks of text mining consist of information extraction, text
categorization and clustering, visualization, and association rule mining [15]. For
example, Chiu and Lin [34] used an information extraction technique to elicit key
concepts from product reviews, and integrated them with Kansei Engineering to build
a data-driven design application interface. The domain ontology is also a relatively new
text mining technique. The purpose of domain ontology is to build a semantic model
that captures the common knowledge in the field of interest [35]. One of the recent
applications is by Chang and Chen [36], who integrated text analytics and domain
ontology to assist in product concept evaluation and selection in crowdsourcing. Text
mining has also been applied to patent knowledge retrieval and extraction, known as
patent mining. Trappey et al. [37] incorporated ontology, patent analysis, and metadata
analysis in dental implant technology evaluation. As well for the patent mining focus,
Govindarajan et al. [38] developed the Excessive Topic Generation, a novel topic
modeling approach, which added the word distance relationship feature to the
traditional topic generation technique. These previous studies attest to the growing
importance and applicability of text mining over the last decade, due to its capability of
analyzing large datasets and thereby enhancing artificial intelligence in decision
making.

Supply chain management is another research area that text mining is utilized more
and more frequently. For example, Papadopoulos et al. [39] implemented content
analysis and confirmatory analysis on numerous text sources (i.e., tweets, news, and
other social media) to validate the theoretical framework they proposed about
explaining sustainability in supply chain. Kim and Kim [40] analyzed news articles and
sustainability reports in textile and apparel industry to explore the trends in sustainable
supply chain management. As for research addressing supply chain risk management
problems, Khan et al. [41] integrated literature review, text mining, and network
analysis to develop a framework of strategies and decision making to deal with
terrorism-related risks in supply chain management. Su and Chen [42] developed a
Twitter Enabled Supplier Status Assessment tool to improve supplier selection process.
They applied text mining on Twitter tweets to retrieve supplier related information and
then analyze potential risks and uncertainty. Song et al. [43] text-mined the commodity
safety regulations of cross-border e-commerce (CBEC), and convert them into risk
rules with fuzzy representation. The proposed method was proven to improve the
efficiency and accuracy of CBEC commodity risk evaluation.

Sentiment analysis, a subtask in text mining, has also been applied to supply chain
management. Sentiment analysis or opinion mining implements text mining and natural
language processing (NLP) to discover polarity and recognize emotions; in other words,
it helps us understand the tone and opinions of the authors of the textual information
[44]. In an application, Chae [45] developed a Twitter analytics framework combining
descriptive, content (text mining and sentiment analysis), and network analytics. This
study discovered that supply chain related tweets are often posted by supply chain
professionals, and some tweets have obvious sentiments about firms’ sales performance,
and even supply chain disruption. Sentiment analysis can also be used in finance.
Schumaker et al. [46] gathered a series of stock prices and the financial news articles
that were published at the same time intervals and used sentiments to assist in price
direction prediction.
 While data analytics and, specifically, text mining or text analytics, have been
applied to many supply chain management issues; to the best of our knowledge, the
specific research area of comprehensive risk categorization and management in global
supply chains has not much benefited from the opportunities these techniques present.
One of the reasons for this might be that the global supply chain risk factors have not
been well defined; that is, there is a variety of different viewpoints among researchers
on this subject. It is crucial to organize previously published concepts of supply chain
risk and further define them. Given the characteristics and capability of textual analysis,
a text-mining-based risk management framework can bridge the gap of undefined risk
factors, risk variation pattern identification, and further supplier selection process.
Defining risk factors can be done by using information extraction and text clustering to
identify potential risks from the extant literature. Then, risk variation pattern
recognition of a region or a company can be accomplished by analyzing sentiments
contained in online news articles. This way, the final goal of improving global supply
chain risk management could be fulfilled by thoroughly understanding and monitoring
the risks.

3. Methodology: global supply chain risk management framework

In this section, a global supply chain risk management framework is proposed and
separated into several subsections to illustrate the analysis mechanisms in sufficient
details. Figure 1 illustrates the proposed framework. The first step is text data retrieval
and preprocessing. The purpose of this study is to analyze texts from extant literature
and news articles to assist in global supply chain risk management operations. After
data collection and transformation, the second step is to extract and define current
potential supply chain risk factors and the risk hierarchy. The last step is risk variation
pattern recognition. Resultant risk factors are incorporated with supply chain related
online news feeds to discover aspect-wise sentiments and further calculate individual
risk levels for each type.
 Figure 1. The proposed framework

Throughout this study, R statistical programming language is used to process text

data and further analyze and visualize the results. In this research, tm and tidytext
packages [47, 48] are used for the analysis steps of the proposed framework. More
details are introduced in the following subsections.

3.1. Defining potential risk factors

3.1.1. Corpus setup: Peered-reviewed journal articles

First a collection of peer-reviewed journal articles related to supply chain risk

management was set up as the corpus. The peer-review process coming to journal
articles makes them a proper source to discover potential risk factors and their
definitions. In this study, Scopus was used to search for articles pertinent to supply
chain risk management. Titles, abstracts, and keywords of articles were set as the main
search fields. Terms such as supply chain, risk management, evaluation, assessment,
and identification were the search grid. In addition, we also limited the results to journal
articles that are in English and were published between years 2000 and 2020. As a result,
a total of 914 articles were downloaded as the input data of this research.

Although more data tends to have more objective results in the field of data mining,
the correlation calculation mechanism in text mining differs from the traditional
correlation analysis. In this research, correlation is calculated at document level. In
other words, the correlation represents the degree of two terms appearing in the same
document. Based on this mechanism, once the number of documents and the total
number of words increase, the correlation will be diluted. Therefore, the corpus with
911 documents was further reduced to a corpus with 118 documents, by limiting the
title with containing “supply”, “chain”, and “risk”. An even a smaller corpus with only
11 documents (well-cited by peer-reviewed journal articles) was also built in order to
obtain more significant correlation results. Note that this smallest corpus was also
human-investigated. This more condense corpus was read thoroughly and was
considered as the state-of-the-art relevant to supply chain risk management. A summary
of this condense corpus can be found in the appendix. We note that across the three
corpora the results were stable other than the correlation results, which became smaller
with the increasing corpus size.

Before further analysis, feature selection was conducted to reduce the

dimensionality of the corpus. In the text mining field, features or attributes represent
terms appearing in the corpus, and therefore feature selection can be done by data pre-
processing to remove less meaningful terms such as numbers and stop words. Several
pre-processing procedures were required to tune the corpus. Since the corpus is in
English, all letters were converted into lower case; all numbers and punctuation were
also removed. Stop words that are usually used to smooth the sentences were also
striped out. All words were then stemmed for simplification. The remaining terms are
more condense and meaningful after the pre-processing. Feature selection can also be
conducted by considering the sparsity or calculating the term-frequency-inverse-
document-frequency (TF-IDF). These two dimensionality reduction approaches serve
as alternatives to further improve the analysis results.

3.1.2. Term frequency and correlation analysis

After tuning the corpus, tm and tidytext packages were used to obtain summary
information of the text. First the corpus was tokenized into a per-term-per-document
(1-gram) data frame. This format is easy to further calculate term frequency. Bi-gram
analysis can provide interesting insights by analyzing the frequency of two consecutive
terms. Document term matrix is also a convenient format to calculate term frequency
and even correlations between terms. Table 1 illustrates an example of document term
matrix from the 11-article corpus. It is a form where rows represent each document and
columns represent each term appearing in the corpus. The entries are the number of
times that the corresponding terms appear in the corresponding documents. Package
tidytext provides a lot of useful functions to deal with tidy text format, and it also has a
function that is able to convert per-term-per document data frame into document term
matrix object, which is a required input format for topic modelling analysis using
package topicmodels [49].
 Article Stemmed term frequency
No. “abil” “abl” “abstract” “accept” “accomplish” “accord” “achiev” “across”
1 1 1 1 1 1 16 5 1
2 2 1 0 1 0 4 2 10
3 2 1 0 2 0 3 0 1
4 2 0 1 0 0 10 2 2
5 1 0 0 4 0 9 2 0
6 13 3 0 2 0 0 7 12
7 0 0 0 0 7 5 9 25
8 1 1 2 2 0 2 0 6
9 2 1 0 0 0 0 0 0
10 0 0 3 0 0 5 0 0
11 0 1 0 2 1 14 2 2
Table 1. Document-term matrix (partial)

After the tokenizing process and document term matrix conversion, term
frequencies were obtained by calculating column sums. From term frequency analysis,
information about how frequently a term appears in the corpus or a single document
can be obtained. Consequently, whether a term or its synonym is important can be
understood. In this research, global supply chain risk management journal articles were
analyzed, and therefore, from term frequencies, the important risk factors with their
attributes (related terms) can be derived.

Correlation analyses were also employed in this research. Both packages tm and
tidytext packages provide functions to calculate correlation. The correlation score
ranges from 0 to 1, meaning the degree of how terms are correlated. By interpreting
correlations, the relation between two words can be revealed. By investigating
correlations, terms having higher correlation with the term “risk” can be considered
more related to risk context. This is used to identify risk factors by emphasizing on
highly correlated terms. The n-gram analysis is also a useful method to discover
relations between terms. The concept of n-gram analysis is from the regular tokenizing
process. Initially the corpus was tokenized into a per-term-per-document (1-gram) data
frame. N-gram analysis is when not only one term but a fixed number of consecutive
terms are tokenized into the tidy data frame. This study employed bi-gram analysis to
further discover the connection between terms, especially between the term “risk” and
others. An illustration of 1-gram and bi-gram is exhibited in Figure 2.
 Figure 2. An example of 1-gram and bi-gram

3.1.3. LDA topic model

Topic model is a text mining unsupervised clustering technique frequently applied

to group unlabeled documents into representative topics. Blei et al. [50] first proposed
latent dirichlet allocation (LDA), a flexible generative probabilistic model for text
corpora and discrete data. It is a three-level hierarchical Bayesian model, and its
application on text mining is to cluster unlabeled documents. The basic concept of topic
model is that every document is assumed arising from a mixture of topics, and each
topic contains a mixture of terms from the corpora vocabulary. A package in R called
topicmodels [49] provide researchers a simple way to derive LDA topic models, and
package ldatuning [51] provides functions to tune the parameters, including finding the
optimal set of the number of topics. In this research, topic models provide an automatic
way to understand the main themes of the documents in the corpus. The use of topic
models along with term frequency and correlation analyses resulted in a more
comprehensive risk factor categorization dictionary.

3.2. Risk variation pattern recognition

After defining potential global supply chain risk factors, the next step is to
recognize risk variation pattern. A variety of global supply chain event information is
discussed in business magazines and news articles. Usually, online news articles are
updated in very short time intervals. Accordingly, this research focuses on online news
articles and conducts sentiment analysis to evaluate global supply chain risk variation
pattern.

3.2.1. Corpus setup: Online news articles

 Unlike collecting peer-reviewed academic journal articles in Section 3.1., here
online news articles were gathered. There are different social media platforms that
collect news from various sources. Google News application programming interface
(API) was employed to search for news of interest. Google news API was selected due
to its several advantageous features: 1) free to use; 2) collects news from diverse
sources, including different topics such as finance, business, technology, politics,
environment, etc.; 3) it has clear tutorial and instruction documentation. Thus, this
study selected Google News API as the source of online news articles.

The defined risk types and the underlying risk factors structured from the former
analysis served as a risk factor dictionary. This dictionary helps filter out news about
specific topics to further analyze supply chain risks for a particular company, industry,
or region. Specifically, the news articles in the hardware technology industry sector
published in 2018 were collected into the corpus for this research.

The feature selection process (data pre-processing) was implemented as described

above. Numbers, stop words, and punctuations were removed. The relationship and
trend between article sentiments over time are essential for capturing risk variation
pattern. Therefore, every article in the corpus was indexed by their publication date also.
This way sentiments can be seen over a time horizon and connections can be made to
critical decisions.

3.2.2. Sentiment analysis

Sentiment analysis is a method for calculating polarity or discover emotions from

words [44]. It can be implemented at either corpus level, document level, section level,
or even sentence level. This study incorporates lexicon-based sentiment analysis to
capture supply chain risk variation patterns, using predefined sentiment lexicons [52].

A sentiment lexicon comprising a list of words, assigned with different sentiments,

is required for sentiment analysis. In R package tidytext, there are several sentiment
lexicons, namely nrc, bing, AFINN, and loughran. The nrc lexicon categorizes its
sentiment word list as “positive”, “negative”, “anger”, “anticipation”, “disgust”,
“fear”, “joy”, “sadness”, “surprise”, or “trust”; the bing lexicon has binary values
of “positive” or “negative”; AFINN uses numerical scores to reflect sentiments,
ranging from -5 to 5, representing “positive” and “negative”; and loughran has
categorical values of “litigious”, “uncertainty”, “constraining”, and “superfluous”.
These lexicons were implemented individually on the news article corpus to generate
sentiment results. In the end, overall sentiments of the corpus and daily polarities over
time were investigated.
 3.2.3. Validation

The sentiment analysis results require further validation to ensure that sentiments
triggered by news articles are relevant to the proposed risk factors. However, for
different global supply chain risk types, researchers and practitioners may apply various
other performance metrics for evaluation. In order to highlight the synergy of the
proposed framework, the scope of the risk was narrowed down to financial and
operational aspects, and correspondingly, stock price was used for validation. Stock
price is volatile and can be affected by a variety of factors. Stock price can also be seen
as investors’ expectation of a company’s capability of operation and profitability. When
something risky to the corporation or negative events happen, it’s likely that it would
be reported in a news article. Indeed, given the plethora of investment companies and
news organizations, this reporting is expected. Therefore, by using the risk factor term
dictionary developed from our risk categorization hierarchy, news related to specific
risk type can be filtered out, and a particular company can be selected. By comparing
the sentiment polarity of those news articles and stock price of the company within the
same time interval, the risk variation pattern can be recognized and validated. Of
course, the underlying assumption is that investors are informed about the related
market events with implications on the company and that realized stock price is a
reflection of investor beliefs.

4. Results and discussion

4.1. Potential risk factors: Global supply chain risk hierarchy

In this section, results of the Defining Potential Risk Factors phase were illustrated
and discussed. Term frequency analysis, correlation analysis, and LDA topic modelling
were conducted on three corpora, the 11-article-human-investigated one (corpus A), the
118-article-machine-selected one (corpus B), and the 911-article-initially-machine-
selected one (Corpus C). The reduction of the corpus size from Corpus C to B and then
to A was implemented in order to gauge the validity of the results. Based on the
mechanism of document level correlation, the more documents and words in a corpus,
the lower the correlation would be. However, if the corpora were appropriately chosen,
the identified risk factors, by and large, would remain the same. All three corpora were
used for the entire analysis. A summary of the subject corpora is shown in Table 2.

Dimension Corpus A Corpus B Corpus C

Number of documents 11 118 911
Number of terms 154,429 1,341,628 10,368,960
Number of terms (after pre-processing) 57,434 531,310 5,008,339
 Unique stemmed terms 4,927 21,205 128,782
Unique stemmed terms (after TF-IDF reduction) 2,205 10,351 53,655
Table 2. Summary of the corpora

All corpora were pre-processed by removing numbers, punctuations, stop words,

and terms that appear regularly in journal articles such as “introduction” and
“references”. Regular expressions were used to allow the corpora only contain the main
body of the articles, filtering out the reference list and other irrelevant terms. Only
meaningful terms were retained. Document term matrices were generated for further
analysis. As shown in Table 2, numbers of terms in all corpora reduced significantly
after pre-processing. Unique stemmed terms were used for term frequency and
correlation analysis, and TF-IDF reduction was conducted on these stemmed terms for
topic modelling.

Term frequency were conducted to all corpora. Figure 3, 4, and 5 demonstrate the
distribution of term frequency and word cloud visualization. An implication from these
figures is that although they are independent (i.e., there are no identical articles within
the corpora), there are many frequent terms that are similar. This implies that the
condensed corpus A is a valid representative of the supply chain risk management
research field, even though it contains very few documents. Therefore, the analysis
results from corpus A was considered as the primary structure for defining potential risk
factors complemented by the results from corpus B and C.

Table 3 presents the identical and different terms of high frequency (partial) to
further explain Figures 3, 4, and 5. The identical terms are reasonable and include “risk”,
“suppli”, etc. They result from the scope of the corpus development focusing on supply
chain risk management. Different high frequency terms explain that the corpus A
mainly contains articles that applied literature review methods to categorize risk factors,
and the corpus B set their focus more on risk assessment and mitigation, especially in
specific business sectors.

Identical “risk”, “suppli”, “chain”, “manag”, “product”, “supplier”, “model”, “process”, “cost”,
terms “intern”, “strategi”, “factor”,“decis”, “oper”, “analysi”, “approach”,
Different “global”, “sourc”, “company”, “review”, “logist”, “inform”, “industri”, “disrupt”,
terms “level”, “system”, “mitig”
Table 3. Comparison of terms of high frequency among corpus A, B, and C (partial)
Figure 3. Term frequency of corpus A

Figure 4. Term frequency of corpus B

 Figure 5. Term frequency of corpus C

In order to discover potential risk factors, correlations between terms (specifically

with the term “risk”) were analyzed. The terms having correlations greater than 0.6
were examined and are listed in Table 4. All 1,030 entries were investigated and 84
terms related to supply chain management were screened out.

Stemmed / original words Correlation Stemmed / original words Correlation

“terror” / terrorism and terrorist 0.93 “micro” / micro 0.83
“demand” / demand 0.92 “nuclear” / nuclear 0.83
“weather” / weather 0.92 “obsolesc” / obsolescence 0.83
“extern” / external 0.90 “oper” / operation 0.83
“breakdown” / breakdown 0.88 “organisa” / organization 0.83
“replenish” / replenishment 0.88 “overstock” / overstock 0.83
“labor”/ labor 0.87 “profit” / profit and profitable 0.83
“utilis” / utilization 0.87 “schedul” / schedule 0.83
“macro”/ macro 0.86 “technolo” / technology 0.83
“account”/ account and “technic” / technical
0.85 0.83
accounting
“disast” / disaster “telecommun” /
0.85 0.83
telecommunication
“disrupt”/ disruption 0.85 “turbul” / turbulence 0.83
“retail”/ retailing and retailer 0.85 “budget” / budget 0.82
“social”/ social 0.85 “yield” / yield 0.81
“variat” / variation 0.85 “insur” / insurance 0.81
“war” / war and warfare 0.85 “competitor” / competitor 0.79
“chain” / chain 0.84 “earthquake” / earthquake 0.78
“maritim” / maritime 0.84 “bankruptcy” / bankruptcy 0.77
“strike”/ strike 0.84 “natur” / nature and natural 0.77
“suppli” / supply and supplier 0.83 “efficienc” / efficiency 0.76
“financ” / finance and financial 0.83 “hurrican” / hurricane 0.76
“flood” / flood 0.83 “instrument” / instrument 0.76
“fire” / fire 0.83 “qualit” / quality 0.76
“flow” flow 0.83 “scarc” / scarce 0.,76
“fluctuat” / fluctuate and “short” / short and shortage
0.83 0.76
fluctuation
 “abnorm” / abnormal and “substitut” / substitute
0.83 0.75
abnormality
“bottleneck” / bottleneck 0.83 “credit” / credit 0.75
“bullwhip” / bullwhip 0.83 “intern” / internal 0.74
“capac” / capacity 0.83 “assest” / asset 0.73
“catastroph” / catastrophe “manag” / manage and
0.83 0.73
management
“deficient” / deficient 0.83 “econom” / economic 0.72
“disassembl” /disassembly 0.83 “network” / network 0.72
“disclosur” / disclosure 0.83 “uncertainti” / uncertainty 0.72
“downturn” / downturn 0.83 “capit” / capital 0.71
“ecosystem” / ecosystem 0.83 “price” / price 0.70
“efficaci” / efficacy 0.83 “transpar” / transparent 0.67
“employe” / employee and “transport” / transportation
0.83 0.67
employer
“horizont” / horizontal 0.83 “safeti” / safety 0.66
“inflat” / inflation 0.83 “transact” / transaction 0.65
“inform” / information 0.83 “dissatisfact” / dissatisfaction 0.64
“infrastruc” / infrastructure 0.83 “environment” / environment 0.64
“loan” / loan 0.83 “tsunami” / tsunami 0.61
Table 4. Terms highly correlated with “risk” (correlation threshold = 0.6)

Take “terror” and “war” (having 0.9-ish correlation with “risk”) as examples, a
global supply chain has stakeholders throughout the world, and devastating events such
as terrorist attacks and wars would significantly impact the local suppliers at the event
regions. Consequently, this will have tremendous influence on supply chain network,
from raw material supply, economic breakdown, to destruction of firms’ infrastructure.

Environmental changes and climate variations are also crucial to global supply
chain networks. Terms such as “weather”, “flood”, “earthquake”, and “tsunami” can be
categorized as environmental risks, and such natural disasters would, for instance,
damage suppliers’ inventory, raw material and manufacturing facilities, causing
operational challenges and greatly affecting the supply chain. The correlation results
above can be seen as inputs to yield a comprehensive dictionary for us to refer to while
defining supply chain related risk factor attributes.

Although correlation analysis provided meaningful results for constructing and

defining potential supply chain risk factors, the stemming feature selection process of
the corpora and multiple meanings of words might lead to different interpretations.
Most of the time when introducing risk factors, phrases with two words such as “supply
risk” or “inventory risk” are used. In order to take this issue into consideration, a bi-
gram analysis was applied to discover how often a word appears consecutively with
“risk”.

Figure 6 demonstrates the frequent terms that appear right after “risk”, and Figure
7 displays the frequent terms appearing right in front of “risk”. As shown in Figure 6,
the phrases “risk type” and “risk mitig” appear more than 50 times and 20 times,
respectively, indicating the corpus highlights discussion of risk types and risk
mitigation. Other terms that come after risk also provided information on the subjects
discussed in the articles while talking about “risk”. Moreover, terms that come directly
before risk give us more intuitive information on the potential risk factors. In Figure 7,
there are several frequent terms that come before “risk” that demonstrate interesting
connections with risk factors, such as “sourc” (e.g., source), “suppli” (e.g., supply),
“chain”, “financi” (e.g., financial), “demand”, “manufactur” (e.g., manufacture),
“environment”, “infrastruct” (e.g., infrastructure) and so on. Figure 8 visualizes the bi-
gram network (only shows the bi-grams that occur more than 20 times) for easy
detection of relationships between terms.

Figure 6. Frequent terms that come after “risk”

 Figure 7. Frequent terms that come before “risk”

Figure 8. Bi-gram network

Since there is a great difference in number of documents among different corpus,

an unsupervised learning algorithm, can assist in understanding the main themes of
each document and grouping them into several topics. Corpus A has only 11 articles,
and is a condensed collection; it therefore provides less information about topics when
topic modelling is used. Corpus C has 911 articles, but applying topic modelling on
large corpus would lead to low interpretability. When the optimal number of topic is
too large, decision makers would have a hard time interpreting the topic clustering
results. Accordingly, LDA topic modelling was conducted on corpus B to generate topic
clustering to have a more inclusive understanding of each individual document.

LDA topic modelling requires a fixed prior-set number of topics, similar to other
clustering approaches. There isn’t a definite way to select the number of topics for LDA
model. However, there are several metrics such as KL-divergence, density-based
inherent connection, and Bayesian statistics, that can help determine the optimal
number of topics, and they are built in package ldatuning [53, 54, 55, 56]. Package
ldatuning [51] was used to determine the optimal number of topics. The result of the
tuning process shows that clustering performance is consistently optimal when the
number of topics is from 20 to 80. In this research, topic modelling served as a
complimentary tool for discovering the mixture of topics of the corpus, and assuring
there was no important risk-factor-related terms being neglected from the former
analysis. Thus, the number of topics was set to 20. This setting also lead to better
readability compared to a large number of topics.

Table 5 presents 20 of the topics generated from corpus B, and five representative
terms are listed under each topic. TF-IDF was used to reduce dimensionality. The
original document term matrix of corpus B contained 21,205 unique terms, and after
TF-IDF screening process the matrix was reduced to 10,351 terms. Terms such as “risk”,
“supply” and “chain” appear in nearly all documents, leading to having low TF-IDF,
and therefore were removed. These words are relatively less important for topic
modelling since they appear in most of the documents in the corpora.

The topic model revealed the main themes of the documents in corpus B. These
topics indicate that the current supply chain risk management field has been focusing
on specific industry sectors’ risk management; this can be explained by the terms such
as “steel”, “food”, “maritime” and “hardwar”. From the results, it is also clear that
sustainability and green supply chain are among the high interest areas. It can also be
observed that corporate management, strategy, and relationships with stakeholders in
their supply chain are also essential given the terms “tactic”, “organis”, “stakehold”,
“trust”, and “partnership”. A variety of decision support methodologies (Delphi-method,
analytic hierarchy process, failure mode and effects analysis, fuzzy theory and Bayesian
method) appear in the topics. This implies that researchers and practitioners have
focused on aggregation methods for risk quantification and uncertainty inherent in risk
factors.
 LDA Topic model
Topic 1 Topic 2 Topic 3 Topic 4 Topic 5
“stakehold” “steel” “pig” “green” “delphi”
“sustain” “flow” “barrier” “fuzzi” “attack”
“trust” “china” “crime” “fmea” “metric”
“scm” “pharmaceut” “humanitarian” “partnership” “stakehold”
“firm” “turbul” “organis” “tactic” “hardwar”
Topic 6 Topic 7 Topic 8 Topic 9 Topic 10
“fuzzi” “countri” “food” “contract” “water”
“food” “farmer” “resili” “retail” “financial”
“oil” “wheat” “agri” “expertis” “motor”
“maritim” “corpor” “entropi” “payment” “crisi”
“ahp” “bin” “perish” “profit” “food”
Topic 11 Topic 12 Topic 13 Topic 14 Topic 15
“scrm” “rfid” “trigger” “food” “resili”
“fuzzi” “retail” “resili” “node” “legitimaci”
“firm” “contract” “oem” “retail” “plant”
“influenc” “wholesal” “profit” “tier” “wind”
“matrix” “profit” “accuraci” “bayesian” “sustain”
Topic 16 Topic 17 Topic 18 Topic 19 Topic 20
“firm” “ahp” “food” “terror” “food”
“flexible” “criteri” “hazard” “resili” “dairi”
“rerout” “hydroge” “climat” “biomass” “hazard”
“vehicl” “cluster” “bioenergi” “algorithm” “green”
“electr” “pairwis” “energi” “bender” “ahp”
Table 5. Topic model (k = 20)
Figure 9. Proposed risk type and risk factors
 Based on the results from term frequency and correlation analysis, bi-gram analysis,
and LDA topic modelling, global supply chain risk was categorized into seven types
and each type contains several potential risk factors derived from critical stemmed
terms, as shown in Figure 9.

In the proposed hierarchical categorization, global supply chain risk is decomposed

into Political, Environmental, Financial, Supply and demand, Logistic, System, and
Operational aspects. Each aspect has its representative terms as its risk factors. For
example, while events related to terrorism or tariff happen in a specific time and region,
they could potentially impact the whole supply chain network and further affect the
performance of relevant stakeholders.

One other example from the categorization is that environmental risk would
increase when a region is struck by a natural disaster like an earthquake or a hurricane.
The frequency of a particular natural disaster in a region would also affect the potential
environmental risk. If a region is used to have earthquakes or hurricanes, the local
manufacturing companies would have high environmental risks, and therefore they
would probably take actions to mitigate the possible damage that might be imposed by
the disasters. This implies that even though the global supply chain risk types and
factors are categorized by a hierarchical structure in this study, these risks types and
factors may not be independent. In other words, several risk factors or risky events
would have chain reactions impacting the entire or part of a global supply chain, and
the fundamental information and composition of a supply chain (industry types, product
types, involved regions, etc.) would be consequently crucial in order to discover the
connection among those risk factors.

Although the proposed global supply chain risk hierarchical categorization did not
reveal inner-connections among the risk elements, it is fundamentally important for
enterprises to understand different risk types and their underlying factors. The proposed
risk management framework reorganized the state-of-the-art research articles of supply
chain risk management by text mining techniques, and developed a comprehensive
categorization for potential risk types and factors. This risk categorization can therefore
become guidance to companies for their supply chain risk management operations.

In the following section, this categorization dictionary was used as a classifier for
filtering out specific online news articles from Google News related to particular supply
chain risks, and then sentiment analysis was conducted for capturing risk variation.

4.2. Risk variation pattern recognition: Sentiment analysis on Google News

 This section demonstrates how risk variation pattern recognition phase in the
proposed framework can be implemented to improve global supply chain management.

First, Google News API was used on the collected news articles. News articles
related to hardware technology industry sector were filtered out, and the search
parameters were set to obtain financial and operational risk related documents. Risk
factor terms such as finance, profit, revenue, technology, capacity, and cost were the
keywords for the search. Moreover, the subject of articles was narrowed down to focus
on only one company (Yageo® Corp.) and limited to the publication year of 2018. This
setting filtered out 69 news articles in total. Through a brief scanning of articles
(human-investigation), a total of 10 identical (Google News API have access to
different sources of news so it might have repeated articles) or irrelevant articles were
eliminated. A total of 59 news article about financial and operational aspects were then
set as the study corpus. An example article is shown below. This example provides a
brief background introduction of the company:

“A Taiwanese manufacturer of components used in everything from smartphones to

computers has turned into the world hottest stock amid surging demand for its products.
Yageo Corp. has jumped more than 800 percent since the start of last year, the most
among stocks on MSCI Inc. global gauge, to add $8.1 billion of value. The company,
which makes capacitors and resistors that are essential for regulating electrical flows
within electronic products, boosted profit more than five-fold from a year earlier in the
first quarter as competitors cut output. Chinese production of such components has
been reduced as the government moved to curb pollution, while leading Japanese
manufacturers have shifted capacity into other products, creating a shortage that has
benefited Yageo, according to Allan Lin, assistant vice president at Concord Securities
Co. The result is an “astonishing” increase in earnings per share, Lin said. Analysts
continue to be bullish on the company, which supplies companies including Apple Inc.,
Intel Corp., and Sony Corp. There are 12 buys, zero holds and just one sell on the stock.
The firm reported a 99 percent increase in sales in April from a year earlier. “The gain
in our stock is determined by the market based on our earnings results,” Yageo
spokesperson Sandy Chang said by phone. Chairman Pierre Chen has a 7.7 percent
stake, according to the latest data compiled by Bloomberg. --- With assistance by Sofia
Horta E Costa, and Young-Sam Cho

After collection of articles, the corpus was tokenized into a tidy data frame, and all
numbers, punctuations, and stop words were removed. It initially had 31,273 words;
after data preprocessing, only 15,954 words (which 3,345 of them are unique words)
were left for sentiment analysis.
 Package tidytext exists 4 built-in lexicons, nrc, bing, AFINN, and loughran. Table
6 provides the basic information of these lexicons. AFINN uses numeric values from -
5 to 5 to reflect sentiments, bing only uses “positive” and negative, loughran specify
sentiments in “constraining”, “litigious”, superfluous, and “uncertainty”, and nrc is
categorized in not only “positive” and “negative” but also emotions like “anger” and
“disgust”. Besides categories, nrc contains more words than other three lexicons;
followed by bing; and the AFINN has the least.

Table 7 illustrates the match ratio between the individual lexicons and the news
article corpus. Generally, the match ratio is low. It’s because not every word has
sentiment and there is no perfect general-purpose lexicon with suitable sentiment words
for a specific domain. To further improve match ratio, a specific domain lexicon can be
generated by sentiment classification. Since nrc and bing have better matching
performance, they were used for sentiment analysis.

Words in
Lexicon “anger” “anticipation” “constraining” “disgust”
lexicon
AFINN 2,476 NA NA NA NA
bing 6,785 NA NA NA NA
loughran 3,916 NA NA 184 NA
nrc 6,468 1,247 839 NA 1,058
“fear” “joy” “litigious” “negative” “positive”
AFINN NA NA NA 1,597 879
bing NA NA NA 4,782 2,006
loughran NA NA 903 2,355 354
nrc 1,476 689 NA 3,324 2,312
“sadness” “superfluous” “surprise” “trust” “uncertainty”
AFINN NA NA NA NA NA
bing NA NA NA NA NA
loughran NA 56 NA NA 297
nrc 1,191 NA 534 1,231 NA
Table 6. Word counts and categories of lexicons

Lexicon Matched words Words in articles Match ratio

AFINN 267 0.08
bing 370 0.11
3345
loughran 327 0.10
nrc 589 0.18
Table7. Match ratio of lexicons and the corpus

By joining the lexicons (nrc and bing) to the tidy data frame of the news article
corpus, words matched with the sentiment word list were counted, and sentiments were
calculated for each category by summing them up. The corpus level sentiments
applying nrc were calculated as shown in Figure 10. The corpus level sentiments seem
to be very positive with over 1,500 positive words, twice as many as the negative words.
Moreover, presence of emotions such as “trust” and “anticipation” show that a great
proportion of the news articles convey a promising perception of Yageo® Corp. on
financial and operational aspects as these were chosen as the scope of the corpus. This
result implies that Yageo® Corp. had less stress on these two aspects in terms of
sentiments in 2018, and this could be interpreted as the company had less risk in terms
of finance and operations.

Figure 10. Corpus level sentiments (nrc)

Although in nrc lexicon, the corpus level sentiment appears to be extremely

positive, the positive word count is only slightly greater than the negative words
applying bing lexicon (Figure 11). A possible reason is that lexicon bing actually has
more negative words than nrc does. Both lexicons have similar total numbers of words,
but bing has 4,782 negative words and nrc only has 3,324. On the other hand, bing also
has less positive words (2,006), than nrc does (2,312). This causes the differences
between these two corpus level sentiment results.

Figure 11. Corpus level sentiments (bing)

Other than overall sentiments, polarity and percent positive over time can also
provide insightful information. In this analysis, lexicon bing was used because polarity
is only related to positive and negative sentiments. The objective of this analysis was
to evaluate sentiments on the temporal horizon. Therefore, the articles were ungrouped
and regrouped in terms of tokens by date. From the 59 articles corpus, there were only
50 unique dates. The sum of the sentiment word count was then obtained. Polarity for
each date was calculated using date-wise positive count minus date-wise negative count.
Each date’s percent positive was obtained by taking the percentage of date-wise
positive count over the sum of date-wise positive and negative counts. Figure 12
illustrates the polarity and percent positive on the date level.

Figure 12. Polarity and percent positive over time using bing (daily)

This figure shows that throughout the 50 dates in 2018, the polarity was gently
varying. There were only several severe differences in January (negative), May
(positive), October (negative), and November (positive). In May, the polarity had the
largest variation moving toward positive. However, because the number of articles
published and collected weren’t the same for each date, there was a slight bias in the
figures. For example, in percent positive figure, 5/22 and 5/23 had high percent positive
value, but they had relatively low polarity (still positive). In these two dates, they are
lower number of articles, and even though the percentage of positive is high, the
difference between positive and negative could still be relatively low compared with
other dates having more articles published. Despite this issue, this result still implies
that authors of these articles were writing more positive than negative news in 2018.

In order to validate sentiment results and their relationship with the corresponding
risk factors, historical stock prices of Yageo® Corp. in 2018 were plotted for comparison.
Since the corpus is mainly about financial and operational aspects, collected by using
the proposed risk types and factors, stock prices may be an appropriate indicator. As
shown in Figure 13, Yageo® Corp. had a significant jump in the middle of 2018, and
this was actually its highest point throughout the past few years. The variation and trend
in this figure also match the sentiment results. High points and low points of stock
prices are clearly similar to the pattern of polarity and percent positive results. In
addition, the sentiment polarity and stock price in this time interval are correlated; a
0.30 correlation score with p-value of 0.04 demonstrates a significant moderate positive
correlation. This significant correlation indicates that sentiments could reflect Yageo®
Corp. stock price trend. Although the stock prices are influenced by numerous factors,
negative events or news about potential risks published on news articles may dominate
investors’ confidence and expectation level of the target company’s performance.
Therefore, stock price could become one of the evaluation approaches to capture its
financial and operational risk variation patterns.

Figure 13. Yageo® Corp. stock price in 2018

4.3. Discussion

This study demonstrates the application of the proposed global supply chain risk
categorization and the implementation of sentiment analysis on risk variation pattern
recognition. After implementing frequency and correlation analysis and topic modeling
on a total of 911 journal articles related to global supply chain risk, a holistic risk
categorization of seven risk types and 81 risk factor terms was developed. This
categorization informs industry practitioners and academic researchers about the wide
spectrum of risk factors relevant to global supply chains. Under the globalized
economic environment, some of these risk factors are extremely region-sensitive, such
as natural disasters, government policies, tax regulations, currency exchange rates, etc.
Therefore, it is essential that firms manage their operations and develop risk mitigation
strategies to account for these uncertainties. The proposed framework can assist firms
in automatically and agilely developing a customized and comprehensive risk hierarchy;
further, they can target the critical ones (can be judged by industry specialists or
decision makers) to manage risk.

The risk factor terms are also a dictionary of global supply chain risk, which can
be used to filter news articles that relate to different risk types. This study applied the
risk hierarchy to collect 59 news articles from Google News API that discussed Yageo®
Corp.’s financial and operational aspects of risk in 2018. The result of the sentiment
analysis shows that the sentiment polarity is positively correlated with the company’s
stock price throughout the year. Stock price is an indicator of the investors’ expectation
of the company’s performance, and when risk events having negative performance
consequences occur, it would be affected accordingly. The correlation between
sentiment polarity and stock price implies that sentiments deduced from the news
articles could signal supply chain risk. Practitioners such as supply chain analysts and
managers could be benefit from such signals as they make short and long term decisions.
A more accurate risk prediction model could also be developed based on the results of
the sentiment analysis.

The presented framework fills several research gaps. For example, Rangel et al.
[57] proposed a supply chain risk classification and sorted the risks into five
management processes in the supply chain (plan, source, make, deliver and return);
however, they applied qualitative literature review on 16 different risk classifications
developed in previous research. By contrast, the proposed systematic artificial-
intelligent-based framework can generate results more efficiently since it has the power
to analyze a larger dataset in a less-time-consuming manner. The results from a data-
driven method would also have less human bias. In cases, where data is mostly human
judgement-based (e.g., survey data such as presented by [58]), concerns in relation to
potential bias may arise. In addition, several studies applied data-driven approaches or
analyzing big data in supply chain management or risk analysis field, but they either
focus only on intrinsic processes of supply chain, specific industry, or evaluation based
on internal numeric data collected from case companies [29, 59]. In contrast, the
proposed framework developed a holistic risk categorization supported by a relatively
large dataset covering extant literature from year 2000 to 2019, which include multiple
 viewpoints on global supply chain risks and a variety of industry sectors. Finally,
although there are a few supply chain risk studies that incorporated text mining (e.g.,
Khan et al. [41] focused on terrorism for its effect on supply chain risk, and Wu et al.
[33] tackled sustainability), to the best of our knowledge, there has never been an
attempt to develop a data-driven comprehensive supply chain risk framework.
Moreover, there is no previous research incorporating sentiment analysis of news
articles to investigate risk variation patterns. Global supply chain risk management has
become an important research focus, and by filling the identified research gaps, this
study provides a holistic supply chain risk framework for companies to develop
strategies to manage risk.

There are several limitations in the study which could be addressed in the future
research. First, the proposed risk categorization didn’t discover the dependency
between risk factor types. Risk factors of one type could also have influences on other
types. Second, in risk variation pattern recognition phase, the subject was narrowed
down to one specific company in hardware technology sector for analysis. This is for
the purpose of demonstrating the implementation of the proposed framework. Global
supply chain is a system with numerous stakeholders, and hence the risks should also
be interconnected. The company’s upstream and downstream stakeholders should also
be taken into consideration in terms of risk variation capturing. Second, there were only
59 news articles in 2018 discussing about Yageo® Corp. financial and operational
aspects. Additional sources and more risk aspects could be added to the analysis to
generate more comprehensive results. Finally, stock price was used to validate the
efficacy of sentiment analysis. However, for different aspects of risks and different type
of focal companies, additional metrics should be considered to evaluate the analysis
performance.

5. Conclusion

This research aimed at utilizing text data from extant literature and online news
articles to address global supply chain risk management. Significantly, a global supply
chain risk management framework was proposed. The framework has three essential
phases: First, two main corpora were set up of peer-reviewed journal articles and online
news articles, respectively. Feature selection (data pre-processing) was conducted to
remove noise terms and reduce dimensionality. Second, research journal articles related
to global supply chain risks were collected and analyzed by term frequency, correlation,
and bi-gram analysis. Topic modelling was also used to assist the analysis.
Implementation of these text mining methods enabled the generation of a global supply
chain risk categorization with a total of seven risk types and a dictionary of risk factor
terms. The risk types contain political, environmental, financial, supply and demand,
logistic, system, and operational aspects. This categorization and the associated
underlying risk factors can improve decision maker knowledge of global supply chain
risks. Third, after categorizing risk types and defining potential risk factors, sentiment
analysis was implemented on online news articles to recognize the variation of supply
chain risks. Google News API was utilized to collect articles related to Yageo® Corp.,
a Taiwanese world-leading passive component provider. The underlying risk factor
dictionary was used for article collection, where the financial and operational aspect of
risks were targeted. The overall corpus level sentiments, polarity, and percent positive
were calculated; these metrics indicated that the firm had a relatively positive pattern
throughout 2018. Moreover, the sentiment polarity score and the stock price of Yageo®
Corp. are moderately (positive) correlated. This result implies that the overall sentiment
of the articles related to financial and operational supply chain risk is able to capture
the risk variation pattern, after comparing the sentiments with stock prices.

This research is one of the first to utilize texts from extant literature to develop a
risk categorization hierarchy and conducting sentiment analysis on news articles to
reveal risk variation pattern. The academic contribution of this study is that it enhances
the related extant works through developing a global supply chain risk categorization
and risk pattern recognition framework. The positive impact on society and industry is
that thanks to the power of text mining, decision makers of enterprises may update
global supply chain risk types automatically, enriching their understanding of potential
risks. They can also be aware of the dynamically changing relative importance of risks
from variation patterns, obtained by analyzing real-time online news articles.
Accordingly, further risk mitigation actions can be made. Overall, the proposed
framework can improve enterprises’ and researchers’ understanding about global
supply chain risks and further provide guidance in risk management conundrums.

The future extensions of this research will focus on analyzing the dependencies
among risk types. Understanding the relationship among different risk types would
significantly improve the global supply chain risk categorization and management. A
large number of companies should be analyzed in the risk variation pattern recognition
phase, preferably including co-dependent stakeholders in a supply chain. Their
connections, responses and impacts to different risk events would also affect the risk
management process.
Appendix

Article Authors
Title Keywords Journal
No. (year)
1 Supply chain Aloini et al. Supply chain Business Process
management: a review (2012) management; Management
of implementation Construction industry; Journal
risks in the Risk management
construction industry

2 Approaches to Christopher et Globalization; Risk Supply Chain

managing global al. (2011) management; Risk Management: An
sourcing risk analysis International
Journal

3 Supply chain risk Ghadge et al. Supply chain The International

management: present (2012) management; Risk Journal of
and future scope management; Supply Logistics
chain risk management; Management
Systematic literature
review; Text mining

4 Supply chain risk Ho et al. Supply chain risk International

management: a (2015) management; Risk types; Journal of
literature review Risk factors; Risk Production
management methods; Research
Literature review

5 Integrated fuzzy multi Kannan et al. Green supply chain Journal of

criteria decision (2013) management (GSCM); Cleaner
making method and Supplier selection; Multi- Production
multi-objective objective linear
programming programming (MOLP);
approach for supplier Maxi-min method; Order
selection and order allocation
allocation in a green
supply chain

6 Global supply chain Manuj and Risk management; International

risk management Mentzer Supply chain Journal of
strategies (2008) management; Physical
International business Distribution &
Logistics
Management

7 Defining supply chain Mentzer et al. N/A Journal of

management (2001) Business
Logistics

8 Product architecture Pashaei and Product design; Supply Chain

and supply chain Olhager Systematic literature Management: An
design: a systematic (2015) review; Platform; International
review and research Integral, Modular, Supply Journal
agenda chain design
 9 A decision framework Soni and Risk management; Journal of
for assessment of risk Kodali (2013) Supply chain Modelling in
associated with global management; Decision Management
supply chain making, Assessment;
PROMETHEE; Goal
programming; Global;
Facility location

10 Identifying risk issues Tang and Supply Chain; Risk International

and research Musa (2011) Management; Journal of
advancements in Citation/Co-citation Production
supply chain risk Analysis Economics
management

11 Strategic production- Vidal and Distribution; Strategic European Journal

distribution models: A Goetschalckx production-distribution of Operational
critical review with (1997) models; Global supply Research
emphasis on global chain modeling; Global
supply chain models logistics systems

References

[1] Kannan, D., Khodaverdi, R., Olfat, L., Jafarian, A., & Diabat, A. (2013). Integrated
fuzzy multi criteria decision making method and multi-objective programming
approach for supplier selection and order allocation in a green supply chain. Journal of
Cleaner production, 47, 355-367.

[2] Mentzer, J. T., DeWitt, W., Keebler, J. S., Min, S., Nix, N. W., Smith, C. D., & Zacharia,
Z. G. (2001). Defining supply chain management. Journal of Business logistics, 22(2),
1-25.

[3] Barry, J. (2004). Supply chain risk in an uncertain global supply chain environment.
International Journal of Physical Distribution & Logistics Management, 34(9), 695-
697.

[4] Park, K., Kremer, G. E. O., & Ma, J. (2018). A regional information-based multi-
attribute and multi-objective decision-making approach for sustainable supplier
selection and order allocation. Journal of Cleaner Production, 187, 590-604.

[5] Koufteros, X., Vonderembse, M., & Jayaram, J. (2005). Internal and external integration
for product development: the contingency effects of uncertainty, equivocality, and
platform strategy. Decision Sciences, 36(1), 97-133.

[6] Christopher, M., Mena, C., Khan, O., & Yurt, O. (2011). Approaches to managing
global sourcing risk. Supply Chain Management: An International Journal, 16(2), 67-
81.
[7] Heckmann, I., Comes, T., & Nickel, S. (2015). A critical review on supply chain risk–
Definition, measure and modeling. Omega, 52, 119-132.

[8] Braithwaite, A. (2003). The supply chain risks of global sourcing. LCP consulting.

[9] Aloini, D., Dulmin, R., Mininno, V., & Ponticelli, S. (2012). Supply chain management:
a review of implementation risks in the construction industry. Business Process
Management Journal, 18(5), 735-761.

[10] Ghadge, A., Dani, S., & Kalawsky, R. (2012). Supply chain risk management: present
and future scope. The International Journal of Logistics Management, 23(3), 313-339.

[11] Ponis, S. T., & Ntalla, A. C. (2016). Supply chain risk management frameworks and
models: a review. International Journal of Supply Chain Management, 5(4), 1-11.

[12] Yan, W., He, J., & Trappey, A. J. (2019). Risk-aware supply chain intelligence: AI-
enabled supply chain and logistics management considering risk mitigation.

[13] Baryannis, G., Validi, S., Dani, S., & Antoniou, G. (2019). Supply chain risk
management and artificial intelligence: state of the art and future research
directions. International Journal of Production Research, 57(7), 2179-2202.

[14] Mitchelstein, E., & Boczkowski, P. J. (2009). Between tradition and change: A review
of recent research on online news production. Journalism, 10(5), 562-586.

[15] Gupta, V., & Lehal, G.S. (2009). A survey of text mining techniques and applications.
Journal of Emerging Technologies in Web Intelligence, 1(1), 60-76.

[16] Meixell, M. J., & Gargeya, V.B. (2005). Global supply chain design: A literature review
and critique. Transportation Research Part E: Logistics and Transportation Review,
41(6), 531-550.

[17] Tang, O., & Musa, S.N. (2011). Identifying risk issues and research advancements in
supply chain risk management. International Journal of Production Economics, 133(1),
25-34.

[18] Ho, W., Zheng, T., Yildiz, H., & Talluri, S. (2015). Supply chain risk management: a
literature review. International Journal of Production Research, 53(16), 5031-5069.

[19] Chopra, S., Sodhi, M.S. (2004). Managing risk to avoid supply-chain breakdown. MIT
Sloan Management Review, 46, 53–62.

[20] Cucchiella, F., Gastaldi, M. (2006). Risk management in supply chain: A real option
 approach. Journal of Manufacturing Technology Management, 17, 700–720.

[21] Tummala, R., Schoenherr, T. (2011). Assessing and managing risks using the supply
chain risk management process (SCRMP). Supply Chain Management: An
International Journal, 16, 474–483.

[22] Tsai, M. C., Liao, C. H., & Han, C. S. (2008). Risk perception on logistics outsourcing
of retail chains: model development and empirical verification in Taiwan. Supply Chain
Management: An International Journal, 13(6), 415-424.

[23] Trkman, P., & McCormack, K. (2009). Supply chain risk in turbulent environments—
A conceptual model for managing supply chain network risk. International Journal of
Production Economics, 119(2), 247-258.

[24] Kayis, B., & Dana Karningsih, P. (2012). SCRIS: A knowledge-based system tool for
assisting manufacturing organizations in identifying supply chain risks. Journal of
Manufacturing Technology Management, 23(7), 834-852.

[25] McAfee, A., Brynjolfsson, E., Davenport, T. H., Patil, D. J., & Barton, D. (2012). Big
data: the management revolution. Harvard business review, 90(10), 60-68.

[26] Power, D.J. (2016). Data science: supporting decision-making. Journal of Decision
systems, 25(4), 345-356.

[27] Van der Aalst, W.M. (2014). Data scientist: The engineer of the future. In Enterprise
interoperability VI (pp. 13-26). Springer, Cham.

[28] Waller, M.A., & Fawcett, S.E. (2013). Data science, predictive analytics, and big data:
a revolution that will transform supply chain design and management. Journal of
Business Logistics, 34(2), 77-84.

[29] Sanders, N.R. (2016). How to use big data to drive your supply chain. California
Management Review, 58(3), 26-48.

[30] Nguyen, T., Li, Z. H.O.U., Spiegler, V., Ieromonachou, P., & Lin, Y. (2018). Big data
analytics in supply chain management: A state-of-the-art literature review. Computers
& Operations Research, 98, 254-264.

[31] Choi, Y., Lee, H., & Irani, Z. (2018). Big data-driven fuzzy cognitive map for
prioritising IT service procurement in the public sector. Annals of Operations
Research, 270(1-2), 75-104.
[32] Huang, Y. Y., & Handfield, R.B. (2015). Measuring the benefits of ERP on supply
management maturity model: a “big data” method. International Journal of Operations
& Production Management, 35(1), 2-25.

[33] Wu, K. J., Liao, C. J., Tseng, M. L., Lim, M. K., Hu, J., & Tan, K. (2017). Toward
sustainability: using big data to explore the decisive attributes of supply chain risks and
uncertainties. Journal of Cleaner Production, 142, 663-676.

[34] Chiu, M. C., & Lin, K. Z. (2018). Utilizing text mining and Kansei Engineering to
support data-driven design automation at conceptual design stage. Advanced
Engineering Informatics, 38, 826-839.

[35] Formica, A. (2006). Ontology-based concept similarity in formal concept

analysis. Information sciences, 176(18), 2624-2641.

[36] Chang, D., & Chen, C. H. (2015). Product concept evaluation and selection using data
mining and domain ontology in a crowdsourcing environment. Advanced Engineering
Informatics, 29(4), 759-774.

[37] Trappey, C. V., Trappey, A. J., Peng, H. Y., Lin, L. D., & Wang, T. M. (2014). A
knowledge centric methodology for dental implant technology assessment using
ontology based patent analysis and clinical meta-analysis. Advanced Engineering
Informatics, 28(2), 153-165.

[38] Govindarajan, U. H., Trappey, A. J., & Trappey, C. V. (2019). Intelligent collaborative
patent mining using excessive topic generation. Advanced Engineering Informatics, 42,
100955.

[39] Papadopoulos, T., Gunasekaran, A., Dubey, R., Altay, N., Childe, S. J., & Fosso-Wamba,
S. (2017). The role of Big Data in explaining disaster resilience in supply chains for
sustainability. Journal of Cleaner Production, 142, 1108-1118.

[40] Kim, D., & Kim, S. (2017). Sustainable supply chain based on news articles and
sustainability reports: Text mining with Leximancer and DICTION. Sustainability, 9(6),
1008.

[41] Khan, M.N., Akhtar, P., & Merali, Y. (2018). Strategies and effective decision-making
against terrorism affecting supply chain risk management and security: A novel
combination of triangulated methods. Industrial Management & Data Systems, 118(7),
1528-1546.

[42] Su, C.J., & Chen, Y.A. (2018). Risk assessment for global supplier selection using text
 mining. Computers & Electrical Engineering, 68, 140-155.

[43] Song, B., Yan, W., & Zhang, T. (2019). Cross-border e-commerce commodity risk
assessment using text mining and fuzzy rule-based reasoning. Advanced Engineering
Informatics, 40, 69-80.

[44] Cambria, E., Schuller, B., Xia, Y., & Havasi, C. (2013). New avenues in opinion mining
and sentiment analysis. IEEE Intelligent systems, 28(2), 15-21.

[45] Chae, B. K. (2015). Insights from hashtag# supplychain and Twitter Analytics:
Considering Twitter and Twitter data for supply chain practice and
research. International Journal of Production Economics, 165, 247-259.

[46] Schumaker, R. P., Zhang, Y., Huang, C. N., & Chen, H. (2012). Evaluating sentiment
in financial news articles. Decision Support Systems, 53(3), 458-464.

[47] Feinerer, I., & Hornik, K. (2015). tm: Text Mining Package. R package version 0.6-2.

[48] Silge, J., & Robinson, D. (2016). tidytext: Text mining and analysis using tidy data
principles in r. The Journal of Open Source Software, 1(3), 37.

[49] Hornik, K., & Grün, B. (2011). topicmodels: An R package for fitting topic
models. Journal of Statistical Software, 40(13), 1-30.

[50] Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent dirichlet allocation. Journal of
machine Learning research, 3(Jan), 993-1022.

[51] Murzintcev Nikita (2019). ldatuning: Tuning of the Latent Dirichlet Allocation Models
Parameters. R package version 1.0.0. https://CRAN.R-project.org/package=ldatuning

[52] Medhat, W., Hassan, A., & Korashy, H. (2014). Sentiment analysis algorithms and
applications: A survey. Ain Shams engineering journal, 5(4), 1093-1113.

[53] Arun, R., Suresh, V., Madhavan, C. V., & Murthy, M. N. (2010, June). On finding the
natural number of topics with latent dirichlet allocation: Some observations. In Pacific-
Asia conference on knowledge discovery and data mining (pp. 391-402). Springer,
Berlin, Heidelberg.

[54] Cao, J., Xia, T., Li, J., Zhang, Y., & Tang, S. (2009). A density-based method for
adaptive LDA model selection. Neurocomputing, 72(7-9), 1775-1781.

[55] Deveaud, R., SanJuan, E., & Bellot, P. (2014). Accurate and effective latent concept
modeling for ad hoc information retrieval. Document numérique, 17(1), 61-84.
[56] Griffiths, T. L., & Steyvers, M. (2004). Finding scientific topics. Proceedings of the
National academy of Sciences, 101(suppl 1), 5228-5235.

[57] Rangel, D. A., de Oliveira, T. K., & Leite, M. S. A. (2015). Supply chain risk
classification: discussion and proposal. International Journal of Production
Research, 53(22), 6868-6887.

[58] Giannakis, M., & Papadopoulos, T. (2016). Supply chain sustainability: A risk
management approach. International Journal of Production Economics, 171, 455-470.

[59] Xu, M., Cui, Y., Hu, M., Xu, X., Zhang, Z., Liang, S., & Qu, S. (2019). Supply chain
sustainability risk and assessment. Journal of Cleaner Production, 225, 857-867.