Embedding-Based Topic Modeling for Hindi Across Diverse Texts

With advancements in language models and text embedding techniques, embedding-based topic models have gained prominence over traditional statistical approaches by capturing deeper semantic structures in text. While models such as the Embedded Topic Model (ETM) and BERTopic have demonstrated superior performance for English, their effectiveness for Hindi remains largely unexamined. Moreover, no prior study systematically evaluates these models across diverse data formats, particularly for Hindi. This research bridges that gap by assessing ETM and BERTopic with various multilingual and Hindi-specific embedding models, comparing them against traditional methods like Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization (NMF). We conduct evaluations on heterogeneous Hindi text datasets, including short-form reviews, conversational YouTube comments, and long-form news articles, to investigate the impact of linguistic structure and text format on topic modeling performance. Coherence scores are analyzed across a wide range of topic counts, revealing that BERTopic performs best with multilingual embeddings, while ETM benefits from Hindi-specific embeddings. Additionally, embedding-based models consistently outperform traditional methods, particularly in short-text scenarios where statistical models struggle. These findings underscore the critical role of embedding selection and data structure in topic modeling effectiveness, offering key insights into optimizing topic models for Hindi and other lowresource languages.