Evaluating Traditional And Transformer-Based Topic Models For Indian Financial News Analytics

Vrishali A. Chakkarwar; Dr. Sharvari C. Tamane

doi:10.29284/fh59ny80

Authors

Vrishali A. Chakkarwar Research Scholar, Department of Computer Science and Engineering, Dr. Babasaheb Ambedkar Marathwada University, Chhatrapati Sambhajinagar -431004, Maharashtra, India
Dr. Sharvari C. Tamane Director & Professor, University Department of Information & Communication Technology MGM University, Chhatrapati Sambhajinagar , Maharashtra, India

DOI:

https://doi.org/10.29284/fh59ny80

Keywords:

Topic Modeling, Latent Dirichlet Allocation, Non-negative Matrix Factorization, BERTopic, Unsupervised Learning.

Abstract

Topic modeling allows the unsupervised learning of latent thematic representations in large textual collections which can in turn form the basis of semantic analyses and later uses such as trend analysis and classification. Considering the complexities of the Indian financial markets and the limitations of purely quantitative models, the study will conduct a comparative evaluation of the Latent Dirichlet Allocation (LDA), Nonnegative Matrix Factorization (NMF), and BERTopic, on a dataset of 50,000 headlines and descriptions of Indian business news that were scraped from Business Standard between 2003 and May 2020. LDA is described as a probabilistic generative model, NMF as a non-negative ma-trix factorization model, and BERTopic as a transformer architecture that combines Sen-tenceBERT embedding with UMAP dimensionality reduction, HDBSCAN clustering, and class-based TF-IDF weighting.

The qualitative analysis was conducted through the comparative analysis of salient words and word clouds. The alternative methods may not be as effective as BERTopic when evaluating contextual coherence and defining clear topical boundaries because it al-lows the discrimination between closely related financial subdomains. The findings empha-size the effectiveness of embedding-based topic modeling approaches in achieving greater interpretability and advancing the scope of financial text analytics.

References

[1] H. Li, ”Latent Dirichlet Allocation,” Springer Nature Singapore, 2023, pp. 439–471. doi: 10.1007/978-981-99-3917-6 20.

[2] S. Mifrah, ”Topic Modeling Coherence: A Comparative Study between LDA and NMF Models using COVID’19 Corpus,” IJATCSE, vol. 9, no. 4, pp. 5756–5761, Aug. 2020, doi: 10.30534/ijatcse/2020/231942020.

[3] R. Kannan, G. Ballard, and H. Park, ”A high-performance parallel algorithm for non-negative matrix factorization,” SIGPLAN Not., vol. 51, no. 8, pp. 1–11, Feb. 2016, doi: 10.1145/3016078.2851152.

[4] R. Kannan, G. Ballard, and H. Park, ”MPI-FAUN: An MPI-Based Framework for Alternating-Updating Nonnegative Matrix Factorization,” IEEE Trans. Knowl. Data Eng., vol. 30, no. 3, pp. 544–558, Mar. 2018, doi: 10.1109/tkde.2017.2767592.

[5] M. W. Berry and M. Browne, ”Email Surveillance Using Non-negative Matrix Factor-ization,” Comput Math Organiz Theor, vol. 11, no. 3, pp. 249–264, Oct. 2005, doi: 10.1007/s10588-005-5380-5.

[6] M. Luo, F. Nie, X. Chang, Y. Yang, A. Hauptmann, and Q. Zheng, ”Probabilistic Non-Negative Matrix Factorization and Its Robust Extensions for Topic Modeling,” AAAI, vol. 31, no. 1, Feb. 2017, doi: 10.1609/aaai.v31i1.10832.

[7] S. Fathi Hafshejani and Z. Moaberfard, ”Initialization for non-negative matrix factoriza-tion: a comprehensive review,” Int J Data Sci Anal, vol. 16, no. 1, pp. 119–134, Nov. 2022, doi: 10.1007/s41060-022-00370-9.

[8] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, ”BERT: Pre-training of Deep Bidi-rectional Transformers for Language Understanding,” Association for Computational Lin-guistics, Jan. 2019, pp. 4171–4186. doi: 10.18653/v1/n19-1423.

[9] A. C. Nanayakkara and G. A. D. M. Thennakoon, ”Enhancing social media content anal-ysis with advanced topic modeling techniques: A comparative study,” International Jour-nal on Advances in ICT for Emerging Countries, vol. 17, no. 1, pp. 40–47, May 2024, doi: 10.4038/icter.v17i1.7276.

[10] L. George and P. Sumathy, ”An integrated clustering and BERT framework for improved topic modeling,” International Journal of Information Technology, vol. 15, no. 4, pp. 2187–2195, Jan. 2023, doi: 10.1007/s41870-023-01268-w.

[11] M. Grootendorst, ”BERTopic: Neural topic modeling with a class-based TF-IDF proce-dure,” Cornell University, Mar. 11, 2022. doi: 10.48550/arxiv.2203.05794.

[12] J. Blanco-Portals, F. Peiro´, and S. Estrade´, ”Strategies for EELS data analysis: Introducing UMAP and HDBSCAN for dimensionality reduction and clustering,” Microscopy and Mi-croanalysis, vol. 28, no. 1, pp. 109–122, Nov. 2021, doi: 10.1017/S1431927621013696.

[13] K. A. Coulter, ”The impact of news media on Bitcoin prices: Modelling data-driven dis-courses in the crypto-economy with natural language processing,” Royal Society Open Science, vol. 9, no. 4, Apr. 2022, Art. no. 220276, doi: 10.1098/rsos.220276.