Fine-Tuning of Multilingual Models for Sentiment Classification in Code-Mixed Indian Language Texts

Abstract

We use XLM (Cross-lingual Language Model), a transformer-based model, to perform sentiment analysis on Kannada-English code-mixed texts. The model was fine-tuned for sentiment analysis using the KanCMD dataset. We assessed the model’s performance on English-only and Kannada-only scripts. Also, Malayalam and Tamil datasets were used to evaluate the model. Our work shows that transformer-based architectures for sequential classification tasks, at least for sentiment analysis, perform better than traditional machine learning solutions for code-mixed data.