IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 2, June 2024, pp. 1723~1731

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i2.pp1723-1731  1723

Hybrid model for extractive single document summarization:

utilizing BERTopic and BERT model

Maryanto, Philips, Abba Suganda Girsang

Department of Computer Science, School of Computer Science, Bina Nusantara University, Jakarta, Indonesia

Article Info ABSTRACT

Article history: Extractive text summarization has been a popular research area for many
years. The goal of this task is to generate a compact and coherent summary
Received Feb 20, 2023 of a given document, preserving the most important information. However,
Revised Oct 17, 2023 current extractive summarization methods still face several challenges such
Accepted Nov 16, 2023 as semantic drift, repetition, redundancy, and lack of coherence. A novel
approach is presented in this paper to improve the performance of an
extractive summarization model based on bidirectional encoder
Keywords: representations from transformers (BERT) by incorporating topic modeling
using the BERTopic model. Our method first utilizes BERTopic to identify
BERTopic the dominant topics in a document and then employs a BERT-based deep
BERT neural network to extract the most salient sentences related to those topics.
Cable news network Our experiments on the cable news network (CNN)/daily mail dataset
Daily mail demonstrate that our proposed method outperforms state-of-the-art
Extractive summarization BERT-based extractive summarization models in terms of recall-oriented
understudy for gisting evaluation (ROUGE) scores, which resulted in an
increase of 32.53% of ROUGE-1, 47.55% of ROUGE-2, and 16.63% of
ROUGE-L when compared to baseline BERT-based extractive
summarization models. This paper contributes to the field of extractive text
summarization, highlights the potential of topic modeling in improving
summarization results, and provides a new direction for future research.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Abba Suganda Girsang
Department of Computer Science, School of Computer Science, Bina Nusantara University
Jakarta-11480, Indonesia
Email: [email protected]

1. INTRODUCTION
In recent years, the advancement in technology has made access to information much easier. Textual
material has been thriving on the internet and has only grown ever since. The total number of scholarly
articles on the internet has reached at least 27 million by 2014 [1]. With the rapid growth of digital
information, the demand for effective automatic text summarizer has increased, leading to research of more
efficient and effective text summarization approaches [2].
Automatic text summarizer is one of the subsets of natural language processing (NLP), which is
used to create a more concise form of long textual content for easier digestion by humans [3]. Text
summarization can be done in two main approaches: extractive and abstractive. Extractive text summarization is
an approach of text summarization that extracts important features, i.e. words or sentences, from a source
document without alteration to be combined into a more concise summary [4]. On the other hand, abstractive
text summarization extracts important information from the source document and generates its own
summary, which may generate novel words in the process [5]. Based on the amount of documents involved,
automatic text summarizer can also be divided into two categories: single document and multi-document

Journal homepage: http://ijai.iaescore.com

1724  ISSN: 2252-8938

summarization. Single document summarization is associated with generating short summaries from long
text documents, while multi-document summarization works with multiple documents that are correlated
with unspecified themes. Recently, deep learning approaches have shown great potential in text summarization,
particularly with the advent of pre-trained models such as bidirectional encoder representations from
transformers (BERT) [6]. BERT is a transformer-based architecture that has achieved state-of-the-art results
in various NLP tasks, including text summarization tasks [7]–[9].
This paper proposed an extractive-based text summarization method that leverages both BERT and
BERTopic, a variant of BERT specifically designed for topic modeling. Our method first applies BERTopic
to identify the underlying topics and form sentence clusters. Then BERT is applied to summarize each
sentence cluster. Finally, the summary from each cluster is combined to form the final summary. To the best
of our knowledge, this is the first implementation of such an approach.
Experiments were conducted on the cable news network (CNN)/daily mail dataset, and the results
show that our proposed method outperforms BERT when it is used on its own in terms of recall-oriented
understudy for gisting evaluation (ROUGE) scores, a widely used evaluation metric for summarization. By
implementing these improvements, we were able to improve the performance of BERT extractive
summarization to 32.53% of ROUGE-1, 47.55% of ROUGE-2, and 16.63% of ROUGE-L. In conclusion, our
proposed method demonstrates the effectiveness of combining BERT and BERTOPIC for extractive text
summarization. The results show that BERT can effectively capture the semantic information of the text and
BERTOPIC can effectively identify the underlying topics, leading to a more informative and coherent
summary. This work provides a new direction for future research in the field of extractive text summarization
and highlights the potential of leveraging pre-trained models for this task.

2. RELATED WORKS
2.1. Topic modeling
Topic modeling is a text-mining technique used in NLP and computer science to discover and
extract the latent topics (i.e. word clusters) present in a large collection of text documents [10]. It is a type of
unsupervised learning, which entails that it is performed without labeled data [10]. In recent years, it has
gained significant popularity due to its efficiency in handling large volumes of unstructured text data and its
ability to automatically identify patterns and themes within that data. It has a wide range of applications,
including information retrieval, document classification, sentiment analysis, and social media analysis [11]–[13].
Several studies have also demonstrated the viability of a topic modeling approach on extractive
summarization [14]–[16]. Those studies incorporated bag-of-words approach-based algorithms, such as latent
dirichlet allocation (LDA) and latent semantic analysis (LSA). These algorithms differ in their mathematical
formulation and the assumptions they make about the underlying structure of the text data, but they all share
the goal of identifying a set of latent topics that can be used to describe the collection of text documents.
However, these techniques disregard the grammar of the words, hence losing the contextual information of
the text. This often results in a false interpretation of the text. In conclusion, topic modeling is a powerful and
widely used technique for uncovering the underlying themes and patterns in large collections of text data. Its
applications are numerous, and it continues to be an active area of research in the field of NLP and computer
science.

2.2. Bidirectional encoder representations from transformers

BERT is a pre-trained bidirectional encoder published in 2018 that is based on the transformer
architecture. BERT was intended for NLP tasks, including sentiment analysis and question-answering. Since
its initial release, BERT has become one of the most widely used language models in NLP and has been
incorporated in numerous studies. In this literature review, we will discuss some of the key contributions and
findings related to BERT.
BERT’s bidirectional training mechanism is a critical factor in it is popularity. By processing input
sequences from both left and right sides, BERT can capture the context in a way that unidirectional models
cannot. Another important contribution of BERT is its use of attention mechanisms. The attention mechanism
is highly parallelizable, unlike sequential models like recurrent neural network (RNN), which may be
difficult to scale with large corpus [17]–[18]. This allows BERT to attend to different parts of the input
sequence, regardless of the document size. As a result, BERT can better capture the relationships between
words and improve its representations. This has been shown to result in improved performance on many
benchmark NLP tasks, such as general language understanding evaluation (GLUE), multi-genre natural
language inference (MultiNLI), and stanford question answering dataset (SQUAD) v1.1 [19].
Finally, BERT has also been used as a pre-trained model in many NLP applications, including
question-answering, text classification, and text summarization. In these applications, BERT has been shown

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1723-1731

Int J Artif Intell ISSN: 2252-8938  1725

to outperform other pre-trained models and even task-specific models that have been trained from scratch [6].
In conclusion, BERT has had a significant impact on the NLP community and is effective in a wide range of
NLP tasks. Its bidirectional training mechanism and attention mechanisms have been identified as key factors
in its success, and its ability to be fine-tuned for specific tasks has made it a popular choice for NLP
practitioners.

2.3. BERTopic
Topic modeling is a technique originated from text mining that allows latent topics to be extracted
from a document or a collection of documents. The aim is to uncover the underlying structure of the text by
discovering the words and phrases that are most representative of a particular topic. Topic modeling has
numerous applications in various domains, such as text classification, information retrieval, and text
summarization. Recent advancement in topic modeling is the discovery of BERTopic, which leverages the
power of BERT, a transformer-based deep learning model. BERTopic is a pre-trained language model that
has been fine-tuned on a large corpus of text data and is specifically designed for topic modeling. Leveraging
BERT’s attention mechanism, BERTopic is able to capture semantic relationships between words and
phrases in the text, hence having a better understanding of the text, which leads to improved performance on
topic modeling tasks [20]. In a recent study, BERTopic was compared to LDA and other topic modeling
methods on several datasets for text classification and text summarization. The results showed that
BERTopic outperformed the other methods in terms of accuracy and coherence, indicating that it is well-
suited for text classification and summarization tasks [21].
BERTopic works in a few steps: first, the document is embedded using a sentence transformer
model. Next, the embeddings will be reduced in dimensions using uniform manifold approximation and
projection (UMAP). This process will simplify the embedding, hence reducing computation complexities for
the next process. A clustering technique, in this case hierarchical density-based spatial clustering of
applications with noise (HDBSCAN) is then applied to the reduced embeddings to form clusters of
documents. Finally, term frequency-inverse document frequency (TF-IDF) will be applied and the most
relevant topic representations will be extracted.
In addition to its performance, BERTopic has several other advantages over traditional topic
modeling methods. Firstly, BERTopic is highly scalable and can be applied to large datasets, making it ideal
for big data applications [22]. Secondly, BERTopic is easy to use and does not require extensive
hyperparameter tuning, making it a user-friendly tool for topic modeling. Finally, BERTopic is highly
interpretable, as it provides not only the topics but also the words and phrases that are most representative of
each topic, making it a valuable tool for understanding the underlying structure of the text [23]. To conclude,
BERTopic is a powerful tool for topic modeling that leverages the strengths of BERT to provide improved
performance over traditional topic modeling methods. Its scalability, ease of use, and interpretability make it
a valuable tool for a variety of NLP applications, including text classification, information retrieval, and text
summarization.

2.4. Recall-oriented understudy for gisting evaluation

ROUGE is a metric that is widely used to evaluate text summarization tasks. ROUGE works by
measuring the overlapping phrases between the generated summary and reference summary. It is based on
the concept of precision and recall, in which the former measures how much of the produced summary is
relevant to the reference summary, meanwhile the latter measures how much of the reference summary the
produced summary is capturing [24].
Due to its simplicity, ROUGE has been widely used in text summarization. It is determined by
counting the overlapped n-grams, word sequences, and word pairs between the generated summary and the
reference summary. ROUGE offers a variety of metrics, such as ROUGE-L, which evaluates the length of
the longest common subsequence between the two summaries, and ROUGE-N, which measures the overlap
in n-grams. However, despite its popularity, ROUGE possesses some major limitations. Due to how it works,
ROUGE only measures recall, disregarding other important aspects of text summarization such as coherence,
fluency, and relevance. Additionally, ROUGE also doesn’t understand the concept of synonyms, hence not
suitable for evaluating abstractive summarization where new words are generated.
In conclusion, ROUGE is used widely as an evaluation metric for extractive text summarization.
Although ROUGE has some limitations such as it only measures recall and does not take account of
coherence, relevance, and fluency which are important to summarization. it is still an important tool for
evaluating extractive summarization performance.

2.5. Cable news network/daily mail dataset

The CNN/daily mail dataset is a widely used benchmark dataset for the task of text summarization.
It was introduced in the paper “abstractive text summarization using sequence-to-sequence RNNs and
Hybrid model for extractive single document summarization: utilizing BERTopic and … (Maryanto)
1726  ISSN: 2252-8938

beyond” and it consists of news articles and their corresponding highlights or summaries [25]. Since its
introduction, the CNN/daily mail dataset has been used extensively in the field of text summarization to
evaluate the performance of various models.

3. METHOD
In this paper, we attempted to implement BERTopic to model the topic of each sub-document, from
which summary sentences will be decided based on their similarity with the topic. However, this method is
not feasible because, for shorter documents, the topic tends to not be generated. Currently, we are working on
alleviating this issue by exploring other methods. In this case, we are replacing the BERTopic clustering
technique from HDBSCAN to k-nearest neighbor (KNN) algorithm as HDBSCAN does not force documents
into a cluster where they might not belong [26]. This is irrelevant to our case where every sentence in the
news should have correlation with one another. As a result, this will improve our resulting topic
representation. Figure 1 shows how our algorithm works.

Figure 1. Illustrates the proposed method

First, the original article are split into an array of sentences. These sentences are fed to our custom
BERTopic model which will result in three clusters of topics that contain sentences. These topic clusters are
then summarized by the BERT extractive summarization model, and one sentence is produced from each.
These sentences are then joined together to form a final summarization. Following is an example of how a
text will be processed by our method:
Step 1: break the original article dataset into sentences, as shown in Figure 2.
Step 2: feed the sentences into the BERTopic model, as shown in Figure 3.
Step 3: compile sentences in clusters of topics, as shown in Figure 4.
Step 4: summarize each topic and join into the final summarization, as shown in Figure 5.
Step 5: compare with base BERT extractive summarization model result and compute ROUGE-1, ROUGE-2,
ROUGE-L. In this paper, we compared the summary generated from our proposed model with the original
BERT model. The results of this comparison are shown in Figure 6.

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1723-1731

Int J Artif Intell ISSN: 2252-8938  1727

After testing on a sample dataset, we ran the model on 10.000 CNN/daily mail test dataset to perform a
summarization of the given article and evaluate each ROUGE-1, ROUGE-2, and ROUGE-L F1-scores before
averaging them to get the final model performance.

Figure 2. Illustrates text splitting process

Figure 3. Topic extracted from text

Hybrid model for extractive single document summarization: utilizing BERTopic and … (Maryanto)
1728  ISSN: 2252-8938

Figure 4. Sentence topic clusters

Figure 5. Summary being generated from each sentence cluster

Figure 6. Summary generated from various model

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1723-1731

Int J Artif Intell ISSN: 2252-8938  1729

4. RESULTS AND DISCUSSION

Evaluation results of the proposed method are presented in Table 1. In this study, we attempted three
different runs using two, three, and four numbers of topics respectively. One summary sentence is selected
from each topic. Hence the summary length is directly proportional to the amount of topics. As can be seen in
the table, the best result is obtained when three topics are selected. According to Table 1, the F1-score for
ROUGE-2 increases proportionately with the number of topics, or in other words. It means that the greater
the number of topic clusters, which means an increase in summary length in our approach, the overlapping of
bigrams between the system and reference summaries also increases. In this experiment, the highest value of
overlapping bigrams is 14.48%, contained in the summary results of four topics.

Table 1. Evaluation of the model for different number of topics count

Number of topics ROUGE-1 (%) ROUGE-2 (%) ROUGE-L (%)
2 34.02 13.20 21.74
3 35.59 14.27 21.79
4 35.13 14.48 21.02

We evaluate our model and baseline BERT summarization model which results are shown in
Table 2. Compared with baseline models, our proposed solution using BERTopic topic modeling achieves
better scores in overall metrics. This result is expected because our model focuses on clustering sentences
into their relevant latent topics before proceeding to summarization. This will generally produce a
summarization where each sentence represents each main event in the article, and thus it achieves good
performance on this comparison.
- Baseline model: as seen from Table 3, the summary generated by the baseline model manages to include
one sentence that contains the relevant information to the golden summary, pertaining to how the suspect
is found guilty, while the rest were deviated. The longest pair of consecutive words that overlap with the
golden summary is two (bigrams), with 3 occurrences, while the pair of consecutive words with the
highest occurrence belongs to unigram, with 15. It is worth mentioning that a large portion of unigrams
that occurred in the baseline summary are stop words, not keywords.
- Proposed model: the summary generated by the proposed model manages to include most sentences that
are relevant to the golden summary. Our proposed model manages to reach 11-grams with one
occurrence. Additionally, the model also manages to reach a high number of unigrams, with a minimum
amount of stop words. The summary generated was of high proximity to the golden summary, though still
lacking fluency.

Table 2. The performance of thebaseline BERT summarization model compared to our proposed method
Model ROUGE-1 (%) ROUGE-2 (%) ROUGE-L (%)
BERT 26.86 9.67 18.68
BERT+BERTopic 35.59 14.27 21.79

Table 3. Comparison of summaries generated by models with golden summary

Golden summary Baseline BERT summarization model Our proposed model
“Dzhokhar Tsarnaev is “After deliberating for 11½ hours, jurors found “Seventeen of the 30 counts were capital
found guilty on all 30 Dzhokhar Tsarnaev guilty on Wednesday of all 30 charges, meaning he is eligible for the death
charges he faced. counts he faced in the Boston Marathon bombing trial. penalty. After the verdict, CNN legal analyst
Seventeen counts were He was there, Tsarnaev's defense attorney Judy Clarke Paul Callan said Clarke now faces an uphill
capital charges, conceded as the trial opened, but many say the defense battle. After deliberating for 11½ hours,
meaning he is eligible strategy always had been to focus on persuading the jurors found Dzhokhar Tsarnaev guilty on
for the death penalty.” jury to spare Tsarnaev's life. The younger man, Clarke Wednesday of all 30 counts he faced in the
said, was only following his older brother.” Boston Marathon bombing trial.”

5. CONCLUSION
Based on the result of the experiment, it can be concluded that we can achieve higher performance
in extractive summarization by leveraging BERTopic to cluster the original text before performing
summarization. We were able to improve the performance of BERT in the extractive summarization tasks by
32.53% for ROUGE-1, 47.55% for ROUGE-2, and 16.63% for ROUGE-L. Although this method has yielded
a high performance, there is still room for improvement and further research. In this research paper, we only
experimented using BERTopic as the topic modeling model and BERT as the summarization model, while

Hybrid model for extractive single document summarization: utilizing BERTopic and … (Maryanto)
1730  ISSN: 2252-8938

there are other newer models such as ALBERT, TinyBERT, or DistilBERT which can perform
summarization at similar performance without using too many resources. Future work also can experiment
with using our method for abstractive text summarization to further improve the result of the abstractive text
summarization process.

REFERENCES
[1] M. Khabsa and C. L. Giles, “The number of scholarly documents on the public web,” PLoS ONE, vol. 9, no. 5, pp. 1-6, May
2014, doi: 10.1371/journal.pone.0093949.
[2] G. Salton, A. Singhal, M. Mitra, and C. Buckley, “Automatic text structuring and summarization,” Information Processing and
Management, vol. 33, no. 2, pp. 193–207, Mar. 1997, doi: 10.1016/S0306-4573(96)00062-3.
[3] P. Patil, S. Dalmia, S. A. A. Ansari, T. Aul, and V. Bhatnagar, “Automatic text summarizer,” in Proceedings of the 2014
International Conference on Advances in Computing, Communications and Informatics, ICACCI 2014, Sep. 2014, pp. 1530–
1534, doi: 10.1109/ICACCI.2014.6968629.
[4] N. Moratanch and S. Chitrakala, “A survey on extractive text summarization,” in 2017 International Conference on Computer,
Communication and Signal Processing (ICCCSP), Jan. 2017, pp. 1–6, doi: 10.1109/ICCCSP.2017.7944061.
[5] N. Moratanch and S. Chitrakala, “A survey on abstractive text summarization,” in 2016 International Conference on Circuit,
Power and Computing Technologies (ICCPCT), Mar. 2016, pp. 1–7, doi: 10.1109/ICCPCT.2016.7530193.
[6] J. Devlin, M. W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language
understanding,” NAACL HLT 2019 - 2019 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies - Proceedings of the Conference, vol. 1, pp. 4171–4186, 2019.
[7] D. William, S. Achmad, D. Suhartono, and A. P. Gema, “Leveraging BERT with extractive summarization for depression
detection on social media,” in 2022 International Seminar on Intelligent Technology and Its Applications: Advanced Innovations
of Electrical Systems for Humanity, ISITIA 2022-Proceeding, Jul. 2022, pp. 63–68, doi: 10.1109/ISITIA56226.2022.9855370.
[8] T. Niu, C. Xiong, and R. Socher, “Deleter: leveraging BERT to perform unsupervised successive text compression,” arXiv-
Computer Science, pp. 1-5, 2019.
[9] D. Miller, “Leveraging BERT for extractive text summarization on lectures,” arXiv-Computer Science, pp. 1-7, 2019.
[10] G. Miner, D. Delen, J. Elder, A. Fast, T. Hill, and R. A. Nisbet, Practical text mining and statistical analysis for non-structured
text data applications. Massachusetts, USA: Academic Press, 2012.
[11] F. Jian, J. X. Huang, J. Zhao, T. He, and P. Hu, “A simple enhancement for ad-hoc information retrieval via topic modelling,” in
SIGIR 2016-Proceedings of the 39th International ACM SIGIR Conference on Research and Development in Information
Retrieval, Jul. 2016, pp. 733–736, doi: 10.1145/2911451.2914748.
[12] Thielmann, C. Weisser, A. Krenz, and B. Säfken, “Unsupervised document classification integrating web scraping, one-class
SVM and LDA topic modelling,” Journal of Applied Statistics, vol. 50, no. 3, pp. 574–591, Apr. 2023, doi:
10.1080/02664763.2021.1919063.
[13] M. Choirul Rahmadan, A. Nizar Hidayanto, D. Swadani Ekasari, B. Purwandari, and Theresiawati, “Sentiment analysis and topic
modelling using the LDA method related to the flood disaster in Jakarta on Twitter,” in Proceedings-2nd International
Conference on Informatics, Multimedia, Cyber, and Information System, ICIMCIS 2020, Nov. 2020, pp. 126–130, doi:
10.1109/ICIMCIS51567.2020.9354320.
[14] D. F. O. Onah, E. L. L. Pang, and M. El-Haj, “A data-driven latent semantic analysis for automatic text summarization using LDA
topic modelling,” in Proceedings-2022 IEEE International Conference on Big Data, Big Data 2022, Dec. 2022, pp. 2771–2780,
doi: 10.1109/BigData55660.2022.10020259.
[15] K. A. R. Issam, S. Patel*, and S. C. N., “Topic modeling based extractive text summarization,” International Journal of
Innovative Technology and Exploring Engineering, vol. 9, no. 6, pp. 1710–1719, Apr. 2020, doi: 10.35940/ijitee.f4611.049620.
[16] R. Rani and D. K. Lobiyal, “An extractive text summarization approach using tagged-LDA based topic modeling,” Multimedia
Tools and Applications, vol. 80, no. 3, pp. 3275–3305, Sep. 2021, doi: 10.1007/s11042-020-09549-3.
[17] A. Vaswani et al., “Attention is all you need,” Advances in Neural Information Processing Systems, vol. 30, pp. 1-11, 2017.
[18] T. Mikolov, M. Karafiát, L. Burget, C. Jan, and S. Khudanpur, “Recurrent neural network based language model,” in Proceedings
of the 11th Annual Conference of the International Speech Communication Association, INTERSPEECH 2010, Sep. 2010, pp.
1045–1048, doi: 10.21437/interspeech.2010-343.
[19] Y. Liu and M. Lapata, “Text summarization with pretrained encoders,” in EMNLP-IJCNLP 2019-2019 Conference on Empirical
Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, Proceedings
of the Conference, 2019, pp. 3730–3740, doi: 10.18653/v1/d19-1387.
[20] M. Grootendorst, “BERTopic: Neural topic modeling with a class-based TF-IDF procedure,” arXiv-Computer Science, pp. 1-10,
2022.
[21] A. Abuzayed and H. Al-Khalifa, “BERT for Arabic topic modeling: an experimental study on BERTopic technique,” Procedia
CIRP, vol. 189, pp. 191–194, 2021, doi: 10.1016/j.procs.2021.05.096.
[22] G. Hristova and N. Netov, “Media coverage and public perception of distance learning during the COVID-19 pandemic: a topic
modeling approach based on BERTopic,” in Proceedings-2022 IEEE International Conference on Big Data, Big Data 2022, Dec.
2022, pp. 2259–2264, doi: 10.1109/BigData55660.2022.10020466.
[23] A. Thielmann, C. Weisser, T. Kneib, and B. Safken, “Coherence based document clustering,” in Proceedings-17th IEEE
International Conference on Semantic Computing, ICSC 2023, Feb. 2023, pp. 9–16, doi: 10.1109/ICSC56153.2023.00009.
[24] C.-Y. Lin, “ROUGE: A package for automatic evaluation of summaries,” in Proceedings of the Workshop on Text Summarization
Branches Out, 2004, pp. 74-81.
[25] R. Nallapati, B. Zhou, C. dos Santos, Ç. Gulçehre, and B. Xiang, “Abstractive text summarization using sequence-to-sequence
RNNs and beyond,” in CoNLL 2016-20th SIGNLL Conference on Computational Natural Language Learning, Proceedings, 2016,
pp. 280–290, doi: 10.18653/v1/k16-1028.
[26] L. McInnes, J. Healy, and S. Astels, “hdbscan: Hierarchical density based clustering,” The Journal of Open Source Software, vol.
2, no. 11, pp. 1-2, Mar. 2017, doi: 10.21105/joss.00205.

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1723-1731

Int J Artif Intell ISSN: 2252-8938  1731

BIOGRAPHIES OF AUTHORS

Maryanto is a student enrolled in the Bina Nusantara University Graduate

Program (BGP) and is currently in his seventh semester at Bina Nusantara University, located
in Jakarta, Indonesia. In 2021, he worked as a software engineer at Coding Studio, and starting
in 2022, he has been employed as a software engineer at tiket.com. His professional interests
lie in machine learning, software engineering, web programming, and artificial intelligence.
He can be contacted at email: [email protected].

Philips is a student enrolled in the Bina Nusantara University Graduate Program

(BGP) and is currently in his seventh semester at Bina Nusantara University, located in
Jakarta, Indonesia. From 2022 until 2023, he worked as a frontend developer at Fotoyu, and
starting from 2023, he has been employed as a UI/UX designer at Kotakode. His professional
interests lie in machine learning, UI/UX design, web programming, and artificial intelligence.
He can be contacted at email: [email protected].

Abba Suganda Girsang has been serving as a lecturer at the Master in Computer
Science program at Bina Nusantara University in Jakarta, Indonesia, since 2015. He earned his
Ph.D. in 2015 from the Institute of Computer and Communication Engineering within the
Department of Electrical Engineering at National Cheng Kung University in Tainan, Taiwan.
He completed his undergraduate studies in Electrical Engineering at Gadjah Mada University
(UGM) in Yogyakarta, Indonesia, in 2000. Following that, he pursued his master’s degree in
Computer Science at the same university from 2006 to 2008. His professional journey
includes roles as a staff consultant programmer at Bethesda Hospital in Yogyakarta in 2001
and as a web developer from 2002 to 2003. Later, he joined the faculty of the Department of
Informatics Engineering at Janabadra University, where he worked as a lecturer from 2003 to
2015. His research interests encompass swarm intelligence, combinatorial optimization, and
decision support systems. He can be contacted via email at [email protected].

Hybrid model for extractive single document summarization: utilizing BERTopic and … (Maryanto)