COMPARATIVE ANALYSIS OF TRANSFORMER MODELS FOR SENTIMENT CLASSIFICATION IN CODE- MIXED INDIC LANGUAGES

Abstract

Multiple language usage in a single message, or code-mixed text, has increased dramatically as a result of increased social media engagement. Because of this, activities involving Natural Language Processing (NLP), such as sentiment analysis and cyberbullying identification. Models that can effectively manage linguistic variability while retaining high accuracy are needed to address these issues. We investigate transformer-based designs that improve classification performance by utilizing knowledge transfer strategies. RoBERTa, GPT-2, XLM-RoBERTa, and IndicBERT are used in our method, which enhances classification accuracy by the transfer of sharing-private information across code-mixed and monolingual tasks. Results from experiments show that our multi-task framework surpasses single-task models with high accuracy on all datasets with:IndicBERT achieved 96.86% for Hinglish, XLM-RoBERTa achieved 96.95% for Punglish, and IndicBERT obtained 97.55% for Tanglish. In order to advance reliable NLP applications in multilingual environments, this project highlights the transformers' multi-task learning capabilities in enhancing performance on low-resource and code-mixed languages.