Examination of Document Clustering Based on Independent Topic

Analysis and Word Embeddings

Riku Yasutomi

, Seiji Yamada

and Takashi Onoda

Aoyama Gakuin University School of Science and Engineering, 5-10-1 Huchinobe, Chuo, Sagamihara, Kanagawa, Japan

National Institute of Informatics 2-1-2 Hitotsubashi, Chiyoda, Tokyo, Japan

Keywords: Topic Model, NLP, Independent Topic Analysis, ITA, Word Embedding, Word2vec, Machine Learning,

Document Clustering, News Paper.

Abstract: In recent years, research on text mining, which aims to extract useful information from textual data, has been

actively conducted. This paper focuses on document classification methods that extract topics from textual

data and assign documents to the extracted topics. Among these methods, the most representative is Latent

Dirichlet Allocation (LDA). However, it has been pointed out that LDA often extracts similar topics due to

the high amount of shared information between topics. Therefore, this paper proposes a document classifica-

tion method based on Independent Topic Analysis (ITA), which extracts topics based on the independence of

topics, and on Word Embedding, which learn word co-occurrence. This approach aims to avoid extracting

similar topics and to achieve information grouping that is closer to human intuition. As a comparative metric,

we used the agreement rate between the results of manually classifying documents into topics and those clas-

sified by each method. The results of the comparative experiment showed that the agreement rate for docu-

ment classification based on ITA and Word Embedding was the highest. From these results, it was suggested

that the proposed method could achieve document classification closer to human perception.

1 INTRODUCTION

In recent years, the widespread use of mobile devices

and laptops, as well as rapid advancements in infor-

mation technology, have been remarkable. As a result,

anyone can now easily disseminate large volumes of

textual data. These data are being accumulated daily,

and the amount of available textual data continues to

increase. In fact, the document information manage-

ment service market is on a steady path of growth.

Ministry of Internal Affairs and Communications of

Japan has projected an increase in the volume of dig-

ital data. The amount of digital data worldwide is ex-

pected to expand fourfold over the five-year period

since 2011. (MIC, 2012). However, on the other hand,

there is a limit to human information processing ca-

pacity. Therefore, research on text mining, which

aims to extract useful information from such data, has

been actively conducted (Shinobu, 2015).

In particular, various studies have been conducted

https://orcid.org/0009-0008-2675-4982

https://orcid.org/0000-0002-5907-7382

https://orcid.org/0000-0002-5432-0646

on methods for extracting topics from document data

and classifying documents, including hierarchical

clustering techniques applied to word embeddings

(Grootendorst, 2022) and methods utilizing word dis-

tributions. Grouping a large amount of information

into topics that align with human intuition is highly

sought after (Vivek, 2024).

Among these document classification methods,

the most representative approach is Latent Dirichlet

Allocation (LDA), proposed by Blei et al. (Blei,

2003). LDA is a distribution-based method, but many

studies have pointed out the problem of extracting

highly similar topics (Section 2.2).

Therefore, this paper focuses on Independent

Topic Analysis (ITA), proposed by Shinohara (Shino-

hara, 1999), which performs topic extraction with

high independence. However, since ITA is a topic ex-

traction method primarily used for topic tracking, it

must be adapted to document classification problems

by leveraging

word embeddings (Takahiro, 2020).

Yasutomi, R., Yamada, S. and Onoda, T.

Examination of Document Clustering Based on Independent Topic Analysis and Word Embeddings.

DOI: 10.5220/0013104100003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 185-192

ISBN: 978-989-758-737-5; ISSN: 2184-433X

185

As a method for obtaining word embeddings,

this paper employs Word2Vec, which is the simplest

in structure and facilitates analysis. The objective of

this study is to achieve document classification that

aligns closely with human intuition by focusing on

topic independence while utilizing word embed-

dings

The remainder of this paper is organized as fol-

lows: Section 2 explains LDA as the baseline method

and reviews related works. Section 3 presents the pro-

posed method for document classification. Section 4

describes the evaluation experiments of the proposed

method. Finally, Section 5 discusses comparisons

with methods other than LDA, highlights the limita-

tions of the proposed method, and outlines future

challenges, concluding the paper.

2 RELATED RESEARCH

This section discusses the most common document

classification method, LDA, its issues, and the objec-

tive of this study.

2.1 LDA (Latent Dirichlet Allocation)

In LDA, documents are represented as random mix-

tures of latent topics. The graphical model is shown

in Fig 1. In this model, 𝑀 represents the number of

documents. 𝑁 denotes the number of words. 𝐾 stands

for the number of topics. Each plate represents re-

peated structures. Furthermore, 𝛼 is the parameter of

the Dirichlet distribution, which serves as the prior

distribution of the topic distribution 𝜃



for each doc-

ument. And 𝛽 is the parameter of the Dirichlet distri-

bution, which serves as the prior distribution of the

word distribution 𝜙



for each topic. 𝑧

,

denotes the

topic assignment corresponding to the 𝑛-th word in

document 𝑚, indicating from which topic each word

Figure 1: Graphical Model of LDA.

was generated. 𝑤

,

represents the actual observed

word in the 𝑛-th position of document 𝑚. Fig 1 illus-

trates the LDA process. The topic distribution per

document and the word distribution per topic are con-

trolled by the values of the Dirichlet parameters 𝛼

and 𝛽. First, the latent topic 𝑧 of a word is determined

from the document-specific topic distribution 𝜃

through a multinomial distribution. Then, using the

latent topic of the word and the word distribution 𝜙

for each topic, the observed word 𝑤 (bag of words) is

generated. In other words, LDA performs topic ex-

traction and classification. The probabilistic genera-

tive model of words and documents forms the basis

of LDA, allowing it to extract topics and classify doc-

uments effectively.

2.2 LDA Challenges

Many studies have pointed out the problem of extract-

ing highly similar topics with LDA. Mu et al. (2024)

discussed the limitations of traditional topic modeling

methods like LDA, highlighting issues such as the ex-

traction of similar topics and the difficulty in inter-

preting topics. As a result, it can become challenging

to interpret the topics (Aletras and Stevenson, 2014),

often leading to inaccurate topic labels during label

assignment (Gillings and Hardie, 2023). To make the

results interpretable by humans, post-processing of

the model's output may sometimes be required

(Vayansky and Kumar, 2020).

Due to these issues, there is a possibility that the

classification results deviate from human intuition.

Section 3 will provide an explanation of the proposed

method, which focuses on ensuring independence be-

tween topics.

3 PROPOSED METHOD

In this section, we propose a document classification

method that addresses the objective stated in Section

2.1 by utilizing ITA and Word2Vec. Section 3.1 ex-

plains ITA, Section 3.2 describes Word2Vec, and

Section 3.3 provides an explanation of the document

classification method that combines these approaches.

3.1 ITA

We introduce Independent Topic Analysis (ITA),

proposed by Shinohara (Shinohara, 1999). As illus-

trated in Fig 2, ITA is a method for extracting topics

with high independence. Highly independent topics

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

186

are those that share a small amount of mutual infor-

mation with each other. In other words, this method

Figure 2: Image Diagram of Independent Topic Analysis

(ITA).

aims to avoid the extraction of similar topics and an-

alyse the overall topic structure of documents. Let 𝑡∈

1,...,𝑘 denote the topic, 𝑑 ∈1,...,𝑛 the docu-

ment, and 𝑤∈



1,...,𝑚



the word, which are the

common variables used in this method. ITA performs

Singular Value Decomposition (SVD). On the word

frequency matrix 𝐴 of the documents to determine

the significance of word 𝑤 in topic 𝑡 and the signifi-

cance of document 𝑑 in topic 𝑡. To evaluate the inde-

pendence between topics, we use the kurtosis of

higher-order statistics as a metric. Kurtosis is calcu-

lated as the difference between the fourth moment

and a normal distribution with the same mean and

variance. From this value, we calculate the concentra-

tion of words for the topics being used.

ITA seeks to extract topics from document data

such that the word concentration of topics is maxim-

ized and the independence between topics is maxim-

ized. Moreover, since ITA is a topic extraction

method, it requires a document classification method

to utilize these extracted topics. In Section 3.2, we ex-

plain Word2Vec, which is used for document classi-

fication.

3.2 Word2Vec

We classify documents based on the topics extracted

using ITA by employing Word2Vec, proposed by To-

mas et al. (Tomas, 2013). Word2Vec is a model used

to learn vector space representations of vocabulary in

natural language processing. In this paper, we utilize

the Continuous Bag of Words (CBOW) model, as il-

lustrated in Fig 3. CBOW is a neural network com-

prising an input layer, hidden layer, and output layer.

For each of the C words in document 𝑘 , a one-hot

vector is formed and used as the input. The error be-

tween the predicted output and the correct labels

(one-hot vectors) for the actual target words are com-

puted using the cross-entropy loss function. The

weights 𝒘′ of the output layer are obtained through

weight updates via backpropagation. Each column of

these weights represents the distributed representa-

tion of each word. In other words, CBOW is a model

that predicts the central word from its surrounding

context. Consequently, this model captures the se-

mantic similarity between words through co-occur-

rences, representing it as a low-dimensional continu-

ous vector.

3.3 ITA+Word2Vec

As mentioned in Section 3.1, ITA uses the frequency

of word occurrences as input. From this, ITA seeks to

identify topics that maximize the concentration of

words and the independence of each topic from the

document data. Due to this characteristic, there is a

possibility that rare words, which appear only in spe-

cific articles, are preferentially extracted. Addition-

ally, the co-occurrence of words is not taken into ac-

count during the topic extraction process. Therefore,

we want to classify rare words that are assigned high

importance within each topic while considering co-

occurring words. To achieve this, we will introduce

word embeddings. This classification method will

henceforth be referred to as the "Vector Classifica-

tion." By adopting this approach, it becomes possible

Figure 3: Image Diagram of ITA (Independent Topic Anal-

ysis).

Examination of Document Clustering Based on Independent Topic Analysis and Word Embeddings

187

to discover documents that include semantically sim-

ilar words. Even if they don’t contain rare, highly im-

portant words within the topics. The process of docu-

ment classification using Vector Classification is il-

lustrated below (Figure 4).

I. Perform the following process for topic 𝑡∈



1,...,𝑘



(a) Convert all constituent words 𝑤∈



1,...,𝑚



of each topic into word embed-

ings.

(b) Convert the magnitude of each word embed-

ing into the importance of word 𝑤 in topic 𝑡.

to generate the word embedding 𝑉

,

for

topic 𝑡.

II. Perform the following process for document

𝑑∈



1,...,𝑛



(a) Apply weighting to each word using TF-IDF.

(b) Similarly to step I. convert the constituent

words into word embeddings, adjust their

magnitudes, and sum them to generate the

word embedding 𝑉

,

for document 𝑑.

III. Calculate the cosine similarity cos(𝑉

,

𝑉

,

) between the 𝑉

,

of each document

and the 𝑉

,

of all topics.

IV. Assign the document to the topic with the

highest similarity.

Through the aforementioned process, we classify

documents into topics using the

word embeddings

obtained by Word2Vec. Word embeddings are learned

based on word co-occurrences. ITA is a method that

utilizes word occurrence frequencies and leverages

relationships among topics. But it does not take word

co-occurrences into consideration. Therefore, we in-

corporate word embeddings in the process of classi-

fying articles into topics.

4 EVALUATION EXPERIMENT

In the evaluation experiments, we will compare the

classification results obtained using LDA with the re-

sults from the proposed method, Vector Classification.

We will also assess these results against human clas-

sification results. Based on these results, we will ex-

amine methods that are closer to human intuition.

Figure 4: The method for converting documents and topics

into word embeddings in vector classification.

4.1 Used Data and Evaluation Method

In the experiments, we utilize weekly article data

from the Mainichi Shimbun in Japan for the year 2017

(The Mainichi, 2017). The data comprises articles

from weeks 1, 2, 3, 4, 5, 10, 20, 30, 40, and 50. Ini-

tially, we extract and classify 13 topics for each week

using LDA. Subsequently, we extract the same num-

ber of topics from the same data using Independent

Topic Analysis. For these topics, we perform classifi-

cations using both a topic-article classification based

on word importance (hereafter referred to as TF Clas-

sification) and the proposed method, Vector Classifi-

cation.

TF Classification calculates the cosine similarity

between the word importance derived from the topics

extracted through ITA. And the word importance of

each article, weighted using TF-IDF. The cosine sim-

ilarity between vectors 𝑥





⃗

and 𝑥





⃗

is computed using

the following formula.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

188

𝑐𝑜𝑠(𝒙





⃗

,𝒙





⃗

𝒙





⃗

∙ 𝒙





⃗

|𝒙





⃗

| | 𝒙





⃗

This represents the cosine of the angle between the

two vectors, where a larger value indicates a greater

similarity between the topic and the article. We cal-

culate the cosine similarity for all topics and articles,

assigning each article to the topic with the highest

similarity. The number of extracted topics is set to 13.

It is consistent with the number of article categories

in the Mainichi Shimbun. Additionally, the parame-

ters for LDA are set to 𝛼 = 50/𝑘 and 𝛽 = 0.01, fol-

lowing the research by Steyvers et al. (Steyvers,

2007) They are implemented using the Gensim li-

brary in Python. Parameter exploration is conducted

based on coherence values to confirm the differences

in results due to parameter settings. The parameters

for Word2Vec were set to size=600 and window=10,

based on the study by Oren et al. (Oren, 2016).

By comparing the classification results of ITA,

LDA and human classification, we aim to investigate

which method aligns more closely with human per-

ception. The human classification involves the author

manually assigning articles to topics extracted by

Table 1: The overlap of the top five words by probability for the 13 topics extracted using LDA.

topic

𝑡

probability

𝑤=1 𝑤=2 𝑤=3 𝑤=4 𝑤=5

1 scorers Tokyo Hyogo Osaka Chiba

2 cooperation Saito funeral art museum newspaper

3 cooperation

voluntary evacua-

tion

Saito funeral newspaper

4 Japan America picture world Tokyo

5 cooperation Saito funeral mother newspaper

6 defense Tanaka opponent champion judgement

7 party chairman constituency incumbent case party

8 Kawasaki Kashima match Urawa Yokohama ma

9 cooperation Saito funeral newspaper art museum

10 joint research suspect bribe arrest university

11 Higashifukuoka Toukaidai final opponent victory

12 fertilized egg transplant consent IVF man

13 Aogakudai Soudai stage prize victory Jundai

Table 2: The overlap of the top five words by importance for the 13 topics extracted using ITA.

topic

𝑡

importance

𝑤=1 𝑤=2 𝑤=3 𝑤=4 𝑤=5

1 opponent Toyko Tanaka champion defense

2 man woman fertilized egg transplant IVF

3 China America Japan Taiwan Takahara

4 dissolution prime minister reporter Suzuki this year

5 Koike laugh decision Tomin-first political

6 Goyoukai Nihon-buyo everyone recital musical

7 Higashifukuoka Toukaidai final Toin Gakuin semifinal

8 abdication emperor citizens coverage constitution

9 Aogakudai stage prize victory Soudai outward way

10 America company influence UK production

11 party chairman case constituency incumbent party

12 Kashima Kawasaki Yokohama ma Urawa Iwata

13 Saitama Hyogo Kyoto Chiba Okayama

Examination of Document Clustering Based on Independent Topic Analysis and Word Embeddings

189

Table 3: Comparison of the matching rates between article classifications by each method and human classifications.

average num-

ber of articles

ITA LDA

The number

of articles

classified

manuall

Agreement

rate with vec-

tor classifica-

tion (%)

Agreement

rate with TF

classification

(%)

The number

of articles

classified

manuall

Agreement

rate with LDA

(%)

1-5 weeks

1423.8 423.0 64.8 81.0 439.4 40.6

10-50

weeks

※1

1680.0 505.8 69.9 83.8 518.2 43.8

Table 4: Comparison of the matching rates between the top 10 articles and topics for each method.

number of

articles

number of

matched articles

matching rate

(%)

Vector classification

129 98 76.0

LDA (coherence)

88 52 59.1

both LDA and ITA. During this process, only assign

articles that are clearly determined to belong to each

topic. Articles that are judged not to belong to any

topic or those that are determined to belong to multi-

ple topics are excluded. We had other experiment par-

ticipants perform a similar assignment task before-

hand. As a result, the agreement rate with the author's

classification was 85.6%. Therefore, we confirmed in

advance that there was no significant bias in classifi-

cation accuracy.

From these results, we will compare LDA, a topic

extraction and classification method that utilizes a

probabilistic generative model, with ITA, which lev-

erages the relationships between the extracted topics.

Furthermore, we will compare the TF classification

method that uses word importance with the Vector

Classification method that employs distributed repre-

sentations obtained from word co-occurrences.

4.2 Results

Table 1 presents the 13 topics extracted by LDA from

the article data of the first week of 2017. The param-

eters used are those that yielded the best coherence

values. Table 2 displays the topics extracted by ITA

from the same data. As shown in Table 2, ITA assigns

importance to all constituent words of each topic. Al-

lowing the user of the method to infer meaning from

the most important words of each topic, similar to

LDA. LDA exhibits a high degree of commonality in

words across topics, suggesting a substantial amount

of mutual information. In contrast, ITA extracts top-

ics with high independence, resulting in fewer com-

mon words among topics compared to LDA. Similar

results were obtained with other datasets.

Next, Table 3 shows the average number of arti-

cles and the number of articles assigned to the 13 top-

ics by human classification. And the average agree-

ment rates between classifications by each method

and human classification for the article data over one

year. Specifically for weeks 1-5 (one week each) and

weeks 10-50 (every ten weeks). The average number

of articles assigned to each topic by human classifiers

was 423.0 and 505.8 for ITA. While for LDA, the

numbers were 439.4 and 518.2, indicating no signifi-

cant difference. Furthermore, the agreement rates be-

tween articles assigned by each method and those as-

signed by human classifiers were 64.8% and 69.9%

for TF Classification. 81.0% and 83.8% for Vector

Classification, while LDA yielded 40.6% and 43.8%.

Thus, Vector Classification achieved the highest

agreement rates with human classification, whereas

LDA yielded the lowest.

Table 4 presents the results of the agreement rates

for the top 10 articles assigned to each topic when

LDA parameters were determined based on coher-

ence values. It also includes the results when using

Vector Classification. In this case, the agreement rate

for Vector Classification was again the highest. This

indicates that the results did not significantly vary

with different LDA parameters.

4.3 Discussion

From the results in Tables 1 and 2, it is evident that

the amount of common information between topics is

lower in ITA compared to LDA. However, it is also

apparent that there are topics with overlapping con-

tent between the two methods. Figure 5 illustrates the

number of matching top ten words between the 13

topics extracted by LDA and ITA. The top ten words

of topics 6, 7, 8, 11, 12, and 13 extracted by LDA

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

190

match over 70% with those of topics 1, 2, 7, 9, 11, and

12 in ITA. This suggests that the granularity of topics

extracted by ITA is not significantly different from

that of LDA. Simultaneously, it can be inferred that

the issue of redundant topics with a high number of

common words in LDA is mitigated. This allows for

a more multifaceted understanding of the overall

topic structure of the documents.

In the comparative experiment with human classi-

fication using newspaper articles, it was found that

the results of ITA and the Vector Classification based

Figure 5: The number of matching top ten words between

topics extracted by LDA and ITA.

on word2vec exhibited the highest agreement rates

with human classification. This indicates the potential

for achieving document classification that aligns

more closely with human perception through ITA and

document classification based on word embeddings.

Moreover, the lack of significant differences in the

average number of articles assigned to each topic by

humans between LDA and Independent Topic Anal-

ysis implies that there is no substantial difference in

the granularity of the extracted topics. This finding

can also be understood from the comparison in Figure

5. Specifically, it appears that discrepancies in preci-

sion arise during the classification process of articles

into topics beyond the common content identified by

LDA and ITA. The number of articles classified by

humans for each topic accounted for approximately

30% of the total articles. This characteristic of news-

paper articles, which comprehensively documents

events occurring each day, likely contributes to the

difficulty in extracting large topics with a high num-

ber of assigned articles.

Additionally, the agreement rate of TF Classifica-

tion was lower compared to Vector Classification.

ITA aims to minimize the common information

among topics in its extraction method. Consequently,

rare words that appear in a limited number of docu-

ments may be assigned a higher level of importance.

However, when calculating the similarity between

topics and documents using the TF Classification

method, it may only be high for a few documents con-

taining these important yet rare words. This can lead

to classification results that diverge from human per-

ception. Simultaneously, the similarity to articles

containing important words within the topics consist-

ently produces high values. This characteristic could

explain why the results for agreement rates were

higher than those of LDA.

While LDA estimates the agreement rates with

each topic based on word distribution, the proposed

method calculates the agreement rates between topics

and the constituent words of articles. This is achieved

using word embeddings derived from co-occurrence

learning. Methods utilizing probabilistic generative

models are known to achieve high precision in group-

ing documents. However, due to the abundance of

common words among the extracted topics and high

mutual information, a single document may corre-

spond to multiple topics. Therefore, when evaluated

based on proximity to human perception, it is plausi-

ble that precision may be lower.

Regarding the number of articles classified by hu-

mans, ITA showed little variation across weeks, while

LDA exhibited variability in its values. A common

characteristic attributed to LDA is the difficulty in de-

termining optimal parameters. Since the optimal so-

lution varies depending on the dataset and purpose,

adjustments are required for each application. Conse-

quently, experience and trial-and-error may be neces-

sary. In this study, the parameters for LDA were de-

termined based on prior research; however, the output

of LDA is highly sensitive to its parameters. There-

fore, changing the parameters could potentially alter

the experimental results. The same applies to

Word2Vec. In contrast, the results of ITA do not fluc-

tuate significantly based on user input. It is believed

that the minimal variability in the number of articles

classified by humans is a result of this characteristic.

5 CONCLUSIONS

In this paper, we conducted a classification of news-

paper articles into various topics using ITA, a method

known for its high independence in topic extraction,

along with word embeddings. We compared the clas-

sification results of our proposed method with those

obtained through LDA, a widely used topic extraction

Examination of Document Clustering Based on Independent Topic Analysis and Word Embeddings

191

method, as well as with results manually classified by

humans. The findings revealed that our proposed

method demonstrated a higher accuracy in terms of

the agreement rate between the extracted topics and

the assigned articles compared to existing methods.

This suggests that our approach effectively avoids the

extraction of topics characterized by a high number

of common words, which is a noted issue with LDA,

allowing for a more multifaceted representation of

document topic structures.

The number of words in the topics extracted by

ITA corresponds to the number of constituent words

in the input documents, resulting in a high computa-

tional cost for calculating the similarity between top-

ics and documents. Consequently, there are current

limitations in applying ITA to larger-scale document

datasets.

In this study, LDA, the most representative

method for document classification, was chosen as a

comparison benchmark. However, recent models

such as BERTopic, which employs BERT and hierar-

chical clustering (Grootendorst, 2022), and LDA2vec,

which integrates word embeddings prior to LDA ap-

plication (Akihiro, 2019), have also been extensively

studied. Future work must include comparisons with

these state-of-the-art methods. Additionally, it is nec-

essary to explore the use of word embeddings before

applying ITA.

In the present experiment, Japanese newspaper ar-

ticles were used as the dataset. Newspaper articles

were chosen because they cover a wide variety of top-

ics, making it challenging to categorize them into a

small number of groups. As future work, we plan to

explore the application of comparison methods to

other datasets, such as review sites or documents

maintained by companies, as well as to English doc-

uments.

REFERENCES

Ministry of Internal Affairs and Communications (MIC).

(2012). Information and Communications White Paper,

2024 edition (in Japanese)

Shinobu, O. (2015). Technologies and Trends in Text Min-

ing. Japanese Journal of Statistics and Data Science

(JJSD). 28-1, pp.31‒40 (in Japanese)

Grootendorst Maarten. (2022). BERTopic : Neural topic

modeling with a class-based TF-IDF procdure. arXiv :

2203.05794

Vivek M., Mohit A., Rohit K. K. (2024). A comprehensive

and analytical review of text clustering techniques. In-

ternational Journal of Data Science and Analytics.

18:239–258

David M. Blei. Andrew Y. Ng. Michael I. Jordan. (2003).

Latent Dirichlet Allocation. The Journal of Machine

Learning Research. Vol.3, pp993-1022

Yasushi Shinohara. (1999). Independent Topic Analysis -

Extraction of Distinctive Topics through Maximization

of Independence. Communications Techniques.

OFS99-14 (in Japanese)

Takahiro Nishigaki. Kenta Yamamoto. Takashi Onoda.

(2020). Topic Tracking and Visualization Method using

Independent Topic Analysis. AIRCC Publishing Cor-

poration. Vol.10, No.3, pp. 1-18

Yida Mu. Chun Dong. Kalina Bontcheva. Xingyi Song.

(2024). Large Language Models Offer an Alternative to

the Traditional Approach of Topic Modelling

Nikolaos Aletras. Mark Stevenson. (2014). Labelling topics

using unsupervised graph-based methods. In Proceed-

ings of the 52nd Annual Meeting of the Association for

Computational Linguistics. Vol.2, pp.631–636

Mathew Gillings. Andrew Hardie. (2023). The interpreta-

tion of topic models for scholarly analysis: An evalua-

tion and critique of current practice. Digital Scholarship

in the Humanities, 38(2):530–543

Ike Vayansky. Sathish AP Kumar. (2020). A review of

topic modeling methods. Information Systems, 94:

101582

Tomas Mikolov. Kai Chen. Greg Corrado. Jeffrey Dean.

(2013). Efficient Estimation of Word Representations

in Vector Space

Provider: The Mainichi - Japan Daily News. Released by:

Nikkai Associates. (2017). CD - Mainichi Shimbun

Data Collection

Mark Steyvers. Tom Griffiths. (2007). Probabilistic Topic

Models. Lawrence Erbaum Associates

Oren Melamud. David McClosky. Siddharth Patwardhan.

(2016). Mohit Bansal:The Role of Context Types and

Dimensionality in Learning Word Embeddings.

NAACL-HLT 2016. pp.1030–1040

Akihiro Ito. Rina Shirai. Hiroshi Uehara. (2019). Topic Ex-

traction Using Similar Vocabulary with Word2Vec.

IPSJ SIG Technical Report. Vol.2019-ITS-77, No.21

(in Japanese).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

192