Research and Development of an Innovative CNN-BiLSTM with Attention Model for Code-Mixed Sarcasm Detection

Abstract

Tanglish, that is, the use of Tamil language with Latin alphabet and English words sprinkled throughout, is considered the prevalent style of texting and commenting among the Tamil community in the online environment. Sarcasm detection in this form of language is especially difficult through automatic processes due to the importance of tone and context clues that cannot be accounted for solely through lexical components. Previous methods have relied heavily on huge transformer models like mBERT and MURIL to gain impressive performance results but suffer from heavy computational requirements that are impractical at times. This paper proposes an innovative light-weight CNN-BiLSTM-Attention model for sarcasm detection in code-mixed Tamil-English text, specifically designed to solve this problem without using large transformer models.Tested on the DravidianCodeMix FIRE 2023 Tamil-English dataset (4,329 annotated comments), the proposed model reaches a macro F1-score of 0.83, outperforming five baseline models including standalone CNN (0.78) and BiLSTM (0.80) and surpassing the previously best-reported transformer-based result, MURIL, at 0.781. This research and development show that a carefully designed, lightweight architecture can match or exceed transformer-level performance on this task while remaining far cheaper to train and deploy, making it a practical option for real-world Tanglish sarcasm detection on resource-constrained platforms.