Improving Sentiment Analysis of Tamil-English Code-Mixed Sentences

Abstract

This paper investigates sentiment analysis for Tamil-English code-mixed text, a common feature of social media communication in multilingual regions. Code-mixing in Romanized Tamil introduces challenges such as inconsistent spelling, transliteration, and noisy syntax that traditional models are not designed to handle. Using the FIRE-DravidianCodeMix 2020 dataset, we evaluated sentiment classification with lexicon-based methods, classical machine learning models, deep learning (LSTM), the multilingual transformer RemBERT, and hybrid approaches combining lexicon-based features with machine learning models. Results showed that classical models such as Logistic Regression, Naive Bayes, and SVM achieved the most stable performance, reaching around 69% accuracy with weighted F1-scores near 0.60. Deep learning and transformer models offered no clear advantage, with both LSTM and RemBERT performing slightly lower than the classical models, plateauing near 67% accuracy and weighted F1-scores around 0.54. These results emphasize that lightweight statistical models remain the most reliable in noisy and resource-constrained code-mixed environments, while deep learning and transformer architectures require greater adaptation to succeed.