Comprehensive Integration of Machine Learning, Deep Learning, and Fine-Tuned Large Language Models for Malayalam News Categorization

The comprehension of news is substantially improved when it is conveyed in an individual’s native language. The increasing volume of news articles, driven by global events, necessitates the effective organization and categorization of vast amounts of information. Dravidian languages, such as Malayalam, characterized by their intricate morphology and diacritical features, pose distinctive challenges in this domain. This research embarks on an innovative initiative to classify Malayalam news articles into four principal categories: business, entertainment, sports, and technology. A synthesis of Deep Learning and Machine Learning methodologies, alongside diverse embedding techniques, is utilized to assess the precision and effectiveness of the categorization endeavor. The ensemble classifiers XGBoost and Stacking, when integrated with the XLM RoBERTa word embedding, attained the highest accuracy rate of 0.91. Similarly, XLM RoBERTa and FastText word embedding techniques, attained a performance with an accuracy rate of 0.93 for almost all the deep learning model. Furthermore, large language models were leveraged, investigating both zero-shot learning and fine-tuning strategies. The fine-tuned BLOOM model reached an accuracy level of 0.89 The research additionally presents a comprehensive examination of the importance of large language models in facilitating the categorization process. Furthermore, the research provides a comprehensive examination of the role of large language models in enhancing the categorization process. By advancing the discipline of Natural Language Processing, this investigation provides significant insights into the utilization of sophisticated technologies for regional news analysis, underscoring their potential to enhance information management and accessibility.