Transformer Based Classification of Closely Related Dravidian Languages in Code-Mixed Data

In recent years, code mixing, the blending of words and phrases from multiple languages within a single conversation has become increasingly common in multilingual contexts, particularly in India. This work investigates the effectiveness of transformer based models in identifying languages within code mixed data that includes English and three closely related Dravidian languages: Malayalam, Kannada, and Tamil, all written in English script. A comprehensive dataset was developed with five distinct labels: Tn-en (Tamil-English), Kn-en (Kannada-English), Ml-en (Malayalam-English), and En (English). Transformer models, including BERT, CamemBERT, mBERT, DistilBERT, and ELECTRA, were employed for classification, with CamemBERT achieving the highest validation accuracy of 97.62% due to its ability to handle morphologically rich languages effectively. The findings highlight the potential of transformer models in enhancing natural language processing (NLP) applications for code-mixed text, particularly in managing the complexities of closely related multilingual environments.