Using Formal Concept Analysis for Corpus Visualisation and Relevance
Analysis
Fabrice Boissier
1 a
, Irina Rychkova
2
and B
´
en
´
edicte Le Grand
2
1
LRE, EPITA, 14-16 Rue Voltaire, Le Kremlin-Bic
ˆ
etre, France
2
CRI, Universit
´
e Paris 1 Panth
´
eon, Sorbonne, 90 Rue de Tolbiac, Paris, France
Keywords:
Corpus Visualization, Relevance Visualization, Topic Modeling, Formal Concept Analysis, Text Mining,
Natural Language Processing.
Abstract:
Corpora analysis is a common task in digital humanities that profits from the advances in topic modeling and
visualization from the computer science and information system fields. Topic modeling is often done using
methods from the Latent Dirichlet Allocation (LDA) family, and visualizations usually propose views based
on the input documents and topics found. In this paper, we first explore the use of Formal Concept Analysis
(FCA) as a replacement for LDA in order to visualize the most important keywords and then the relevance of
multiple documents concerning close topics. FCA offers another method for analyzing texts that is not based
on probabilities but on the analysis of a lattice and its formal concepts. The main processing pipeline is as
follows: first, documents are cleaned using TreeTagger and BabelFy; next, a lattice is built. Following this, the
mutual impact is calculated as part of the FCA process. Finally, a force-based graph is generated. The output
map is composed of a graph displaying keywords as rings of importance, and documents positioned based on
their relevance. Three experiments are presented to evaluate the keywords displayed and how well relevance
is evolving on the output map.
1 INTRODUCTION
In the age of data, knowledge is an essential fac-
tor that increases the capacity to make the best deci-
sions (North and Kumta, 2018). Data visualization re-
mains a difficult task for knowledge extraction as nu-
merous visualizations are available and each business
and/or application face their own challenges (An-
drienko et al., 2020)(Padilla et al., 2018)(Engebretsen
and Kennedy, 2020). It requires a perfect understand-
ing of the goal to achieve and the manipulated data.
Visualizing text corpora is a cross-domain appli-
cation of interest for information systems and digital
humanities (J
¨
anicke et al., 2015). Improvements in
visualization lead to better analysis of multiple types
of texts like historical newspapers (Menhour et al.,
2023) or even poetry (De Sisto et al., 2024). How-
ever, few work has been done to visualize the rele-
vance between documents based on their topics. Topic
modeling also contributes to texts’ analysis as a re-
search orientation of the information retrieval field
by ”uncovering latent text topics by modeling word
a
https://orcid.org/0000-0002-0067-6524
associations” (Hambarde and Proenca, 2023). Top-
ics are simply a list of words that share a similar
theme: either each word is strongly/directly express-
ing the theme, or the collection of words illustrates
an abstract theme by their semantic links (but alone,
they would not make sense). Multiple topic modeling
methods exist (Alghamdi and Alfalqi, 2015)(Kherwa
and Bansal, 2019). They usually consider documents
as a bag-of-words where the order of words is not im-
portant; only their occurrence in each document is im-
portant. The main trend in topic modeling consists of
probabilistic methods where topics are built based on
the probability of each term being part of a specific
topic. Recent advances in natural language process-
ing (NLP) also introduced neural networks combined
with traditional methods, allowing the capture of the
context of words within documents and reusing it to
analyze newer documents.
In this paper, we explore the use of Formal Con-
cept Analysis (FCA) (Ganter and Wille, 2012) instead
of more traditional topic modeling methods, and we
propose a visualization of the main keywords of a cor-
pus and documents’ relevance on a force-based graph.
The strength of FCA resides in the fact that it an-
120
Boissier, F., Rychkova, I. and Le Grand, B.
Using Formal Concept Analysis for Cor pus Visualisation and Relevance Analysis.
DOI: 10.5220/0013047800003838
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 3: KMIS, pages 120-129
ISBN: 978-989-758-716-0; ISSN: 2184-3228
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

alyzes the relationships within data and produces a
lattice that can be used for calculating useful mea-
sures like similarities. Instead of trying to find back
memberships of an object between multiple distribu-
tions or the best matching weight of a parameter in a
neural network, FCA highlights the existing relation-
ships between objects based on their attributes. FCA
is known as a viable text mining method (Carpineto
and Romano, 2004) and is a good candidate for mul-
tiple applications in the knowledge field (Poelmans
et al., 2013). FCA has been used in conjunction with
a topic modeling method (Akhtar et al., 2019) but not
instead of it.
The paper is organized as follows: First, we ex-
plain the main methods currently used in topic model-
ing as well as topic and relevance visualizations. Sec-
ond, we present our processing pipeline from the in-
put documents to the output map. Then, we present
three experiments, the text materials used, their top-
ics, and their relevance. Finally, we discuss the results
and conclude.
2 RELATED WORK
2.1 Topic Modeling
Analyzing documents implies creating statistics on
the used terms. Term Frequency - Inverse Document
Frequency (TF-IDF) (Salton, 1983) calculates a ra-
tio from the frequency of each term to the total num-
ber of documents. Documents are therefore seen as
a ratio of words independent of their ordering, like a
bag-of-words. TF-IDF is not exactly a topic modeling
method, but it shows the importance and uniqueness
of terms within the corpus.
Latent Semantic Analysis (LSA) (Deerwester
et al., 1990) transforms documents into a latent se-
mantic space from which multiple outputs can be an-
alyzed. The main method behind LSA is the singular
value decomposition, which produces three matrices
based on a parameter K given by the user: a matrix
of terms per K features, a matrix of K features per K
features, and a matrix of K features per documents.
Multiple analyses can be done on those matrices, but
in our case, the terms per K features one is the most
important as it allows us to know how well each term
is linked to each feature (that in fact represents la-
tent topics). One problem induced by LSA is that it
can’t manage polysemy: each term is used as the same
entity in any document. Homonyms, like synonyms
and even various forms of the same word, can pro-
duce an inconsistency because of the missing context
of each document. Standardization of the input, like
stemming or even lemmatization, can partially lever-
age this problem.
Probabilistic Latent Semantic Analysis
(PLSA) (Hofmann, 1999) is an upgrade of LSA
that introduces a probabilistic point of view by
building a generative model for each text corpus.
Because topics are scattered within documents,
probabilities help find the terms that compose them.
PLSA relies on an aspect model (the probabilities
between terms, documents, and the latent topics) and
a mixture decomposition that obtains similar results
as the singular value decomposition, allowing PLSA
also to build the three matrices of LSA.
Latent Dirichlet Allocation (LDA) (Blei et al.,
2003) is also a generative probabilistic model that
aims to model text corpora. It aims to improve the
PLSA mixture decomposition by using a hierarchical
Bayesian model. LDA allows not only the finding of
topics of words within documents, in the case of text
corpora, but also its usage with more confidence than
PLSA as a probabilistic generative model in multiple
domains.
LDA can be combined with Bidirectional Encoder
Representations from Transformers (BERT) (Devlin
et al., 2018) in order to increase the quality of its
results (Peinelt et al., 2020)(George and Sumathy,
2023). BERTs are a collection of pre-trained neu-
ral networks designed to help researchers in NLP by
considering words and their neighbors before and af-
ter them. By their nature, BERTs do not produce a
list of topics but can be used to generate a summary.
Similarly to BERTs, it can be noted that Generative
Pre-trained Transformers (GPT) (Brown et al., 2020)
achieve numerous NLP tasks. Still, they also share
the drawback of not explaining the weights within the
neural network to build their answers.
Correlated Topic Model (CTM) (Blei and Laf-
ferty, 2006) is derived from LDA and uses another
distribution in order to better capture topics and their
relations within documents. Indeed, when a docu-
ment concerns a theme, it usually talks about some
neighbor topics: a text about travel might talk about
tourism, beaches, and airplanes, but probably not
about fighter jets. Topics are uncorrelated in LDA be-
cause of the Dirichlet distribution, whereas in CTM,
thanks to the logistic normal distribution, topics are
correlated and present links to one another. The pres-
ence of a topic triggers the possibility of finding one
or multiple other topics. CTM also proposes a visual-
ization of topics: each topic is represented in a bubble
of words, and each bubble is linked with the other cor-
related topics.
Using Formal Concept Analysis for Corpus Visualisation and Relevance Analysis
121

2.2 Visualization of Topics and
Relevance
The presentation of topics is important as reading a
list might confuse a human: a list forces an order of
reading, which is not always the best one depending
on the context. Visualization of topics is usually made
of tag clouds (Singh et al., 2017)(Lee et al., 2010),
which is convenient but shows terms as a whole block.
Some visualizations like (Singh et al., 2017) or (Gre-
tarsson et al., 2012) show the distinction between top-
ics and the terms composing them. However, if a
document used in the corpus is irrelevant or contains
too many irrelevant parts, the results might be altered
without the user being informed. In addition, each
document must be deeply reviewed to find the dis-
crepancy. Few work has been done to explicit the rel-
evance of documents one to another based on their
topics.
(Assa et al., 1997) presents a relevance map be-
tween topics and keywords in the case of web search
queries. Their proposal uses dimension reduction to
create a 2D map and gravitational forces for placing
nodes. Similarly, VIBE (Olsen et al., 1992)(Olsen
et al., 1993) and its variants (Ahn and Brusilovsky,
2009) create Points of Interest (POI) that act precisely
like topics around which documents are placed based
on relevance. The main drawback is that it does not
highlight the most important common words and top-
ics but only places resources based on their relevance
to all the topics found.
(Fortuna et al., 2005) proposed a map of terms
and documents based on MultiDimensional Scaling
(MDS) methods. Terms and documents are placed
on a two-dimensional map, and the background color
varies based on the density. The main drawback of
this contribution is the difficulty of getting a clear
overview of the documents and terms.
In (Newman et al., 2010), the authors proposed
a topic map after studying topic modeling and 2D
projection. First, they compare three topic model-
ing methods and end by using LDA. Next, they com-
pare three methods of projection, namely Principal
Component Analysis (PCA), Local Linear Embedding
(LLE), and Force Directed Layout (FDL) which is the
best. The topic map presents documents as nodes col-
ored by their most important topic. Their position de-
pends on their relevance to one another. However, the
authors concluded that the evaluation of visualization
is complex and could be made only by human judg-
ment; in addition, they also stated that maps with a
dozen documents are probably the most accurate and
valuable in understandability and navigation for a hu-
man.
TopicViz (Eisenstein et al., 2012) proposes a vi-
sualization of documents and topics by nodes with
a force-directed layout and, more importantly, inter-
action with the user. The topic modeling method
used is LDA. The user can pin topic nodes in the
workspace, making the documents float based on their
relevance to each topic. Such a map allows one to dis-
tinguish which document is more relevant to some or
more topics based on its position. The user can also
pin document nodes, making the topics float between
them. This visualization is particularly interesting be-
cause document and topic nodes can be pinned, allow-
ing it to show relevance. However, the user must pin
the nodes himself in order to see the relevance. Based
on the number of detected topics, deciding where to
pin each topic to see the documents’ relevance better
might be difficult.
PaperViz (di Sciascio et al., 2017) is a dedicated
tool for researchers during the paper-gathering step.
It offers multiple views for multiple contexts: tree
hierarchy for search queries, a tag cloud of the 20
most frequent terms, the strength of the relationship
between documents and a search query or a collec-
tion, and references management. The main strength
of PaperViz is its completeness in the user interface.
3 VISUALIZATION PIPELINE
WITH FORMAL CONCEPT
ANALYSIS
The visualization pipeline comprises two main phases
(see Figure 1): semantic pre-processing, where docu-
ments are analyzed in order to produce an occurrence
matrix of terms per documents, and structural anal-
ysis, where the matrix is analyzed in order to create
a graphical representation of the relevance. Glob-
ally, the pipeline relies on natural language processing
(NLP) methods in the first phase and formal concept
analysis (FCA) methods in the second phase.
3.1 Semantic Pre-Processing
The semantic pre-processing phase aims to extract the
most important terms and concepts from each doc-
ument and gather them within a matrix representing
the whole text corpus. This phase is composed of 5
steps as follows:
(PI.0) Selection of Documents by the User: the
user selects the documents for analysis and compari-
son. Two requirements must be fulfilled for the best
results: each document must have enough content,
and the content must be mainly textual.
KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems
122

Figure 1: The main steps of the pipeline.
(PI.1) Extraction of Texts: each document is
transformed into a regular flat text file. This step re-
lies on optical character recognition (OCR) methods.
In our experiments, we used PDFtoText
1
as the OCR.
(PI.2) Cleaning of Extracted Texts: each text is
cleaned in order to increase its quality and reduce its
size, typically by removing the useless spaces and ar-
tifacts that the OCR would have created and, eventu-
ally, some of the stop words. In our experiments, we
used TreeTagger (Schmid, 1994)(Schmid, 1995) with
a custom list of words to keep.
1
https://www.xpdfreader.com/pdftotext-man.html
Figure 2: The formal concept analysis (PII.1) sub-steps.
(PI.3) Disambiguation: each cleaned text is
transformed into a list of named entities by resolv-
ing polysemy and synonymy problems. Advanced
NLP methods are required for this task. In our ex-
periment, we used BabelFy (Moro et al., 2014) as it
understands multiple languages and calculates three
scores for each recognized named entity. The named
entities are also transformed into unique references
from BabelNet (Navigli and Ponzetto, 2012), allow-
ing us to manipulate the exact same named entities in
all documents, whatever the input languages are.
(PI.4) Filtering of Terms: the most irrelevant
named entities are removed based on the coherence
score attributed by BabelFy in the previous step. In
our experiments, we require a coherence score of at
least 0.05 to keep a named entity. This score was em-
pirically chosen because it removes way more irrele-
vant named entities than relevant ones.
3.2 Structural Analysis
The structural analysis phase comprises two steps
that calculate metrics in order to produce the mutual
impact graph showing the relevance of documents.
(PII.1) The formal concept analysis is the first
step, divided into five sub-steps (see Figure 2). Its
objective is to produce the mutual impact matrix be-
tween terms and documents from the occurrence ma-
trix in order to evaluate the relevance of documents.
• Normalisation: occurrences of terms per docu-
ments are transformed into proportions in order
to reduce the length disproportions between doc-
uments. Absolute values are converted into per-
centages per line. Thus, all documents are treated
equally.
• Transposition: the matrix is transposed in order to
change the point of view from documents charac-
terized by occurrences of terms into terms charac-
terized by their appearances within documents.
• Binarisation Strategy: building the lattice re-
quires a formal context, or, in other words, a
binary matrix. Multiple strategies of binariza-
tion exist (also called refinement strategies) (Jaffal
et al., 2015). In our case, we use the most simple
Using Formal Concept Analysis for Corpus Visualisation and Relevance Analysis
123

one: the direct strategy that transforms all values
> 0 as 1 and keeps 0 as 0.
• Lattice Construction: the formal context repre-
senting terms within documents is used for build-
ing a lattice and its formal concepts (Belohlavek,
2008). A formal concept is a node containing ob-
jects and attributes (at least one of them). In our
case, the objects are terms, and the attributes are
the documents.
• Metric Calculus: the lattice is analyzed, and the
mutual impact (Jaffal et al., 2015) metric is cal-
culated by comparing the appearances of couples
of terms and documents within each formal con-
cept. The mutual impact shows how strong the
relationship is between each term and each docu-
ment. This bond is calculated for each term and
each document with the formula:
MI(O
i
, A
j
) =
formal concepts containing O
i
and A
j
formal concepts containing O
i
or A
j
where O
i
represents a term and A
j
represents a
document. The output is a mutual impact matrix
with a value representing the bond between each
term and each document.
(PII.2) The construction of the mutual impact
graph is the final step. Its objective is to visualize
on a map the terms and their importance within the
corpus, as well as the documents and their relevance.
The visualization uses the mutual impact matrix as an
adjacency matrix and produces a graph of terms and
documents. We used Gephi (Bastian et al., 2009) with
the ForceAtlas (first version) spatialisation algorithm.
Nodes are moved until they find their optimal position
thanks to attractive and repulsive forces. Nodes are
repulsing each other, and edges attract nodes based
on the values of the edges. Because of the input for-
mat, the visualization is a bipartite graph: a set of
nodes represents documents, and another set repre-
sents terms. As presented in Figure 3, the nodes of
documents are colored in grey and are linked to nu-
merous nodes of terms. Unlike, these term nodes
are colored based on the number of neighbors (the
warmer the color, the more the node of term is linked
to different nodes of documents). When focusing on
the nodes of terms, we can find terms from every doc-
ument in the center of the map and terms from fewer
documents scattered around (terms, present in only
one document, are forming specific groups for each
document far from the central set). The central set of
terms is, in fact, a global view of the text corpus with
its keywords. When focusing on the nodes of docu-
ments, we can visually see the relevance of documents
within the corpus by checking how close documents
are to the central set of terms.
Figure 3: The output map of terms and documents.
4 EXPERIMENT
4.1 Scenarios of Evaluation
A proof of concept (Elliott, 2021) has been realized
with a specific scenario to check the pipeline’s valid-
ity and properties. A prototype has been developed
2
and used in three proof of concept demonstrations. A
first case expects to visualize the content of 9 PHP
courses. The mutual impact graph is visualized to
check the corpus’ main keywords and the documents’
relevance. A second case introduces a Java course in
the corpus. The validity of the mutual impact graph
is checked by comparing the first and second cases:
the Java course should be the most irrelevant because
it is not specialized in PHP. A third case is presented
in order to check how correcting a document impacts
the results.
In the regular case, 9 PHP courses in French are
processed through the pipeline. These courses present
web development with HTML, PHP, and MySQL for
beginners. They are denoted as C1-C9 in the follow-
ing figures. 6 of these courses are made of slides (C1,
C2, C4, C7, C8, C9), and 3 are made of regular texts
(C3, C5, C6). A Java course (CJA) in text format
is later introduced to check how the mutual impact
graph behaves when some errors appear. This course
is also in French and in a text format. The third ex-
periment only works on courses in text format. There-
fore, we introduced 4 more PHP courses in text for-
mat (C11, C15, C17, C19) to avoid disproportion by
keeping a close number of input documents as in the
other cases.
2
https://github.com/metalbobinou/CREA-phd
KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems
124

4.2 General View of the Corpus’
Content
The documents are processed through phase I, and the
corpus is transformed into a matrix of occurrences.
The mutual impact graph is generated using the direct
strategy in step PII.1.
Figure 4 shows terms as nodes colored following a
cold-warm schema. Red nodes are terms occurring in
all the documents, orange nodes are terms occurring
in all the documents minus one, etc. In the regular
case, the terms in red nodes are: post, nombre (num-
ber), m
´
ethode post (post method), fichier (file), code,
donn
´
ee (data), langage (language), site, php, naviga-
teur (browser). These terms are typical of a course
on web development with PHP. They are extended
with the terms in orange nodes like base de donn
´
ee
(database) or even foreach, which are also typical for
a website in PHP that uses a database. Terms that
are present in fewer documents but still in more than
half of the corpus (the yellow and green nodes) are
also typical of web development for nearly all of them
(session, mysql, utilisateur (user), ...).
Figure 4: Central set terms of the mutual impact graph in
regular case.
Figure 5: Central set terms of the mutual impact graph in
Java case.
In Figure 5, the Java course is added to the corpus.
The terms in red nodes are the same as in Figure 4,
except php which becomes an orange node. This be-
havior is expected as the Java course does not dis-
cuss PHP; therefore, one document does not include
it. The terms in other colors are still relevant as they
mainly concern client-server programming, OOP vo-
cabulary, and similar topics. It can be noted that an-
other color is introduced in order to show that an ad-
ditional document is present. Nearly all of the terms
in green in the first case are now light green. In the
entire graph, some of the terms in green are repulsed
into dark green nodes (meaning they are missing from
one more document). However, a majority of terms
from the first case are still present in the second case
with the same number of document edges.
4.3 Relevance of Documents
Figure 6: Relevance of documents in regular case.
Figure 7: Relevance of documents in Java case.
Using Formal Concept Analysis for Corpus Visualisation and Relevance Analysis
125

Figures 6 and 7 represent the relevance of documents
in two cases: the regular case with only PHP courses,
and the Java case with the additional Java course. The
relevance’s visualization is also produced from the
mutual impact graph, except the focus is mainly on
the grey nodes representing documents.
In the regular case of PHP courses (Figure 6), the
documents in the slide format are closer to the central
set than the text ones. In the Java case (figure 7), the
Java document (CJA) is the most distant of the central
set. It can be noted that C6 is as distant from the cen-
tral set as CJA. This discrepancy is explained by the
fact that the C6 document contains not only a PHP
course but also reports of students’ projects in more
than half of the document. These reports discuss var-
ious business problems that require a website (online
shoe store, online music store, etc). Therefore, the
document is not exactly a pure PHP course like the
others.
In order to test how the mutual impact graph re-
acts when a document is corrected, we compared mul-
tiple cases of courses while correcting one of them.
Document C6 is written in three parts with nearly the
same amount of pages: the regular PHP course, the re-
ports of the students’ projects, and an advanced PHP
course. For the experiment, we used PHP courses in
text format only (in order to avoid the effect of mix-
ing slide and text formats), and we corrected docu-
ment C6 by removing the students’ projects first and,
later, the advanced chapters. The experiment was also
done twice, with and without the Java course, in order
to have a better view of the effect of correction on a
corpus with and without an irrelevant course.
In the regular case without any modification (Fig-
ure 8), C6 is the most outlier in both regular and
Java cases. In the Java case, CJA is the sole docu-
ment nearly as far as C6. After removing the stu-
dents’ projects from C6 (Figure 9), it becomes one
of the closest documents from the central set in both
cases, indicating that it became way more relevant to
the corpus than previously. In the Java case, CJA be-
comes the most outlier, keeping its irrelevance. Fi-
nally, when the advanced chapters are also removed
from C6 (Figure 10), it becomes the closest to the cen-
tral set in the Java case, and the second closest in the
regular case. C11 and C15 become the most outliers
in the regular case. In the Java case, CJA and C11 are
the most outliers.
In each figure from 8 to 10, document C6 becomes
closer to the central set thanks to the corrections. It
can be noted that CJA stays the most outlier in ev-
ery case, and the other documents move away a bit.
Therefore, the corrections show improvements in the
positioning of C6 while keeping the irrelevance of
CJA.
Figure 8: Relevance of text documents (C6 intact).
Figure 9: Relevance of text documents (no student
projects).
KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems
126

Figure 10: Relevance of text documents in the regular case
(no student projects, nor advanced chapters).
5 DISCUSSION
In the mutual impact graph of the regular case (Fig-
ure 4), the set of terms present in all of the documents
at the center of the graph, shown as red nodes, are rel-
evant to the content of the text corpus, and they also
produce a clear summary of the main keywords. Each
ring of colors around the central set adds more terms
relevant to the corpus. When the Java course is in-
troduced (Figure 5), the terms in all the documents
are nearly the same: only php is a bit repulsed. As
the Java course also mainly talks about programming
and partially about web development with dedicated
frameworks, the results are nearly unchanged, which
is the expected behavior. The mutual impact graph
with the terms lets a user quickly get an idea of the
subject and keywords composing the corpus. It can
be used to quickly discover a new academic field and
find the keywords that best describe it. Another usage
for this graph would be to help build a book’s index:
the keywords are highlighted, and the author selects
which words to keep or remove.
Concerning the relevance of documents in Fig-
ure 7, the Java course is the most distant one with
C6 (which contains a lot of unwanted content) in the
first experiment. This behavior is perfect for the cur-
rent case: the teacher who would like to use exist-
ing documents is warned that C6 and CJA should be
checked more precisely in order to detect if their con-
tent is relevant. Even if the text documents are distant
from the slide documents, the most irrelevant ones
are far away, allowing the user to measure the rele-
vance visually. In the third experiment, the relevance
of document C6 is greatly improved because of the
corrections applied while keeping the irrelevant doc-
ument CJA far away. It must be noted that the shape
of the graph changes because it reflects the relevance
of each document relating to the general relevancy of
the whole corpus. The mutual impact graph is a pic-
ture of the corpus as a whole: it is not a graph about
the relevance of one or some documents against one
or some other documents. Multiple use cases can be
derived from the mutual impact graph. The graph
would help users select the best documents about one
or more topics or remove the most irrelevant ones.
Another usage would be for a teacher to compare its
own course with the existing ones or even with re-
search articles to check how close it is to the state of
the art.
The global results show that FCA, with the mu-
tual impact measure, can highlight a corpus’ main
terms and even show its documents’ relevance. It
does not create a list of topics nor calculate the prob-
ability of each term being included in a topic like
LDA, but it does reflect the importance of each term
for the whole corpus. This behavior is expected by
the nature of FCA: it ”differs from statistical data
analysis in that the emphasis is on recognizing and
generalizing structural similarities, such as set inclu-
sion relation from the data description, and not on
mathematical manipulations of probability distribu-
tions” (Carpineto and Romano, 2004). However, we
acknowledge that BabelFy does participate actively
in the process of topic modeling by recognizing the
named entities and evaluating their relevance to each
portion of the text. It must be noted that the con-
struction of the formal context also filters some terms.
FCA (formal context’s construction + metrics calcu-
lus) and BabelFy must be used together, or at least,
FCA with an entity linking tool. Another limitation
concerns the fact that the documents used in the ex-
periments were written in French and with the goal
of teaching. In our opinion, language shouldn’t be a
problem as BabelFy is well trained for recognizing a
lot of languages, but mixing different documents writ-
ten with distinct goals or distinct formats might affect
the quality of the resulting map in a similar way that
can be seen in Figure 6 where the documents in the
slide format are closer to the central set than the text
ones.
Using Formal Concept Analysis for Corpus Visualisation and Relevance Analysis
127

6 CONCLUSION
In this paper, we proposed a visualization pipeline for
textual corpora analysis based on FCA instead of the
usual LDA for the topic modeling step. Mutual im-
pact was used within FCA in order to produce a ma-
trix for the force-based graph algorithm. The pipeline
produces a map that can be used in two ways:
• The main keywords are placed by order of impor-
tance, allowing the reader to quickly get an idea
of the topics contained in the corpus
• Documents are placed based on their relevance to
the keywords found, allowing the reader to see an
eventual discrepancy in the chosen texts.
The map presents a visualization of the corpus as
a whole. Removing a document impacts the visual-
ization because of the absence of a node and because
the topic modeling step does not work on the same
texts. To evaluate these claims, we presented a case
study on multiple PHP courses and an intruder Java
course (also about programming). First, a map dis-
played the most important keywords and the varia-
tions with and without the Java course. Then, one of
the PHP courses containing more than half of its text
about out-of-scope topics was corrected, showing a
significant upgrade in the output.
We consider FCA an exciting method for topic
modeling and expect to try other metrics on the lat-
tice in order to find more possible usages. Multiple
usages and combinations have been already proposed
in (Poelmans et al., 2013), but we expect to use the
conceptual similarity metric (Jaffal et al., 2015) for
an even more precise combination of terms. Also, a
deeper comparison with LDA and other newer meth-
ods like neural networks might be interesting as the
construction of the results does not rely on probabili-
ties and is perfectly transparent thanks to the set the-
ory behind FCA. Numerous applications in digital hu-
manities can be considered with FCA because trans-
parency, and therefore explainability, are already part
of its design by highlighting relationships between
objects.
REFERENCES
Ahn, J.-w. and Brusilovsky, P. (2009). Adaptive visualiza-
tion of search results: Bringing user models to visual
analytics. Information Visualization, 8(3):167–179.
Akhtar, N., Javed, H., and Ahmad, T. (2019). Hierarchical
summarization of text documents using topic model-
ing and formal concept analysis. In Data Manage-
ment, Analytics and Innovation: Proceedings of ICD-
MAI 2018, Volume 2, pages 21–33. Springer.
Alghamdi, R. and Alfalqi, K. (2015). A survey of topic
modeling in text mining. Int. J. Adv. Comput. Sci.
Appl.(IJACSA), 6(1).
Andrienko, G., Andrienko, N., Drucker, S. M., Fekete, J.-
D., Fisher, D., Idreos, S., Kraska, T., Li, G., Ma, K.-
L., Mackinlay, J. D., et al. (2020). Big data visual-
ization and analytics: Future research challenges and
emerging applications. BigVis 2020: Big data visual
exploration and analytics.
Assa, J., Cohen-Or, D., and Milo, T. (1997). Displaying
data in multidimensional relevance space with 2d vi-
sualization maps. In Proceedings. Visualization’97
(Cat. No. 97CB36155), pages 127–134. IEEE.
Bastian, M., Heymann, S., and Jacomy, M. (2009). Gephi:
an open source software for exploring and manipulat-
ing networks. In Third international AAAI conference
on weblogs and social media.
Belohlavek, R. (2008). Introduction to formal concept anal-
ysis. Palacky University, Department of Computer
Science, Olomouc, 47.
Blei, D. and Lafferty, J. (2006). Correlated topic models.
Advances in neural information processing systems,
18:147.
Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent
dirichlet allocation. Journal of machine Learning re-
search, 3(Jan):993–1022.
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,
Askell, A., et al. (2020). Language models are few-
shot learners. Advances in neural information pro-
cessing systems, 33:1877–1901.
Carpineto, C. and Romano, G. (2004). Concept data anal-
ysis: Theory and applications. John Wiley & Sons.
De Sisto, M., Hern
´
andez-Lorenzo, L., De la Rosa, J., Ros,
S., and Gonz
´
alez-Blanco, E. (2024). Understand-
ing poetry using natural language processing tools: a
survey. Digital Scholarship in the Humanities, page
fqae001.
Deerwester, S., Dumais, S. T., Furnas, G. W., Landauer,
T. K., and Harshman, R. (1990). Indexing by latent
semantic analysis. Journal of the American society
for information science, 41(6):391–407.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.
(2018). Bert: Pre-training of deep bidirectional trans-
formers for language understanding. arXiv preprint
arXiv:1810.04805.
di Sciascio, C., Mayr, L., and Veas, E. (2017). Exploring
and summarizing document colletions with multiple
coordinated views. In Proceedings of the 2017 ACM
Workshop on Exploratory Search and Interactive Data
Analytics, pages 41–48.
Eisenstein, J., Chau, D. H., Kittur, A., and Xing, E. (2012).
Topicviz: Interactive topic exploration in document
collections. In CHI’12 Extended Abstracts on Human
Factors in Computing Systems, pages 2177–2182.
Elliott, S. (2021). Proof of concept research. Philosophy of
Science, 88(2):258–280.
Engebretsen, M. and Kennedy, H. (2020). Data visualiza-
tion in society. Amsterdam university press.
KMIS 2024 - 16th International Conference on Knowledge Management and Information Systems
128

Fortuna, B., Grobelnik, M., and Mladenic, D. (2005). Visu-
alization of text document corpus. Informatica, 29(4).
Ganter, B. and Wille, R. (2012). Formal concept analysis:
mathematical foundations. Springer Science & Busi-
ness Media.
George, L. and Sumathy, P. (2023). An integrated clus-
tering and bert framework for improved topic model-
ing. International Journal of Information Technology,
15(4):2187–2195.
Gretarsson, B., O’donovan, J., Bostandjiev, S., H
¨
ollerer, T.,
Asuncion, A., Newman, D., and Smyth, P. (2012).
Topicnets: Visual analysis of large text corpora with
topic modeling. ACM Transactions on Intelligent Sys-
tems and Technology (TIST), 3(2):1–26.
Hambarde, K. A. and Proenca, H. (2023). Information re-
trieval: recent advances and beyond. IEEE Access.
Hofmann, T. (1999). Probabilistic latent semantic index-
ing. In Proceedings of the 22nd annual international
ACM SIGIR conference on Research and development
in information retrieval, pages 50–57.
Jaffal, A., Le Grand, B., and Kirsch-Pinheiro, M. (2015).
Refinement strategies for correlating context and user
behavior in pervasive information systems. Procedia
Computer Science, 52:1040–1046.
J
¨
anicke, S., Franzini, G., Cheema, M. F., and Scheuermann,
G. (2015). On close and distant reading in digital hu-
manities: A survey and future challenges. EuroVis
(STARs), 2015:83–103.
Kherwa, P. and Bansal, P. (2019). Topic modeling: a
comprehensive review. EAI Endorsed transactions on
scalable information systems, 7(24).
Lee, B., Riche, N. H., Karlson, A. K., and Carpendale,
S. (2010). Sparkclouds: Visualizing trends in tag
clouds. IEEE transactions on visualization and com-
puter graphics, 16(6):1182–1189.
Menhour, H., S¸ahin, H. B., Sarıkaya, R. N., Aktas¸,
M., Sa
˘
glam, R., Ekinci, E., and Eken, S. (2023).
Searchable turkish ocred historical newspaper collec-
tion 1928–1942. Journal of Information Science,
49(2):335–347.
Moro, A., Raganato, A., and Navigli, R. (2014). Entity
linking meets word sense disambiguation: a unified
approach. Transactions of the Association for Com-
putational Linguistics, 2:231–244.
Navigli, R. and Ponzetto, S. P. (2012). Babelnet: The au-
tomatic construction, evaluation and application of a
wide-coverage multilingual semantic network. Artifi-
cial Intelligence, 193:217–250.
Newman, D., Baldwin, T., Cavedon, L., Huang, E., Karimi,
S., Martinez, D., Scholer, F., and Zobel, J. (2010). Vi-
sualizing search results and document collections us-
ing topic maps. Web Semantics: Science, Services and
Agents on the World Wide Web, 8(2-3):169–175.
North, K. and Kumta, G. (2018). Knowledge manage-
ment: Value creation through organizational learning.
Springer.
Olsen, K. A., Korfhage, R. R., Sochats, K. M., Spring,
M. B., and Williams, J. G. (1993). Visualization of
a document collection: The vibe system. Information
Processing & Management, 29(1):69–81.
Olsen, K. A., Williams, J. G., Sochats, K. M., and Hirtle,
S. C. (1992). Ideation through visualization: the vibe
system. Multimedia Review, 3:48–48.
Padilla, L. M., Creem-Regehr, S. H., Hegarty, M., and Ste-
fanucci, J. K. (2018). Decision making with visualiza-
tions: a cognitive framework across disciplines. Cog-
nitive research: principles and implications, 3:1–25.
Peinelt, N., Nguyen, D., and Liakata, M. (2020). tbert:
Topic models and bert joining forces for semantic sim-
ilarity detection. In Proceedings of the 58th annual
meeting of the association for computational linguis-
tics, pages 7047–7055.
Poelmans, J., Kuznetsov, S. O., Ignatov, D. I., and Dedene,
G. (2013). Formal concept analysis in knowledge pro-
cessing: A survey on models and techniques. Expert
systems with applications, 40(16):6601–6623.
Salton, G. (1983). Introduction to modern information re-
trieval. McGraw-Hill.
Schmid, H. (1994). Probabilistic part-of-speech tagging us-
ing decision trees. In New methods in language pro-
cessing, page 154.
Schmid, H. (1995). Improvements in part-of-speech tagging
with an application to german. In In Proceedings of
the ACL SIGDAT-Workshop, pages 47–50.
Singh, J., Zerr, S., and Siersdorfer, S. (2017). Structure-
aware visualization of text corpora. In Proceedings of
the 2017 Conference on Conference Human Informa-
tion Interaction and Retrieval, pages 107–116.
Using Formal Concept Analysis for Corpus Visualisation and Relevance Analysis
129