Topic Modeling for Nepali Political News Using Probabilistic, Algebraic and Transformer-Based Methods

Identifying latent themes within large text collections is essential for understanding discourse structures across domains. This study presents a comparative evaluation of topic modeling techniques applied to Nepali political news, including Latent Dirichlet Allocation (LDA), Non-Negative Matrix Factorization (NMF), Latent Semantic Analysis (LSA), and BERTopic, a Transformer-based neural topic modeling approach. A largescale dataset of Nepali political news articles from 2018 to 2023 was collected and preprocessed using natural language processing (NLP) techniques. Experimental results indicate that LDA outperforms neural topic modeling approaches in the Nepali language context due to the limitations of pre-trained Transformer models for Nepali. Coherence score evaluations confirm LDA's superior topic consistency and interpretability compared to alternative methods. The study identifies key political themes and their temporal trends, offering insights into the evolution of political discourse in Nepal. To the best of our knowledge, this is the first comprehensive study evaluating multiple topic modeling techniques for Nepali text analysis. The findings suggest that while neural models show promise, traditional probabilistic methods remain more effective for low-resource languages. Future work includes enhancing neural topic modeling through fine-tuned Nepali language models and expanding the dataset for broader generalizability.