Detecting Offensive Language in Tamil YouTube Comments

Online content moderation faces significant challenges in identifying offensive language across diverse linguistic environments. This study compares the performance of five advanced BERT models mBERT, BERT Base, BERT Large, DistilBERT, and RoBERTa in detecting offensive language specifically in Tamil YouTube comments. Through meticulous evaluation using metrics like accuracy, precision, recall, and F1-score, we analyze the nuanced processing capabilities of these models tailored to address the complexities of the Tamil language. Among the models assessed, mBERT emerges as a computationally efficient option due to its multilingual efficacy, while RoBERTa demonstrates impressive accuracy with fine-tuning on a Tamil-specific corpus. BERT Base exhibits solid performance, albeit with higher computational demands, and BERT Large showcases enhanced accuracy at the cost of increased resource requirements. DistilBERT presents a memory-efficient alternative, maintaining reasonable performance with a smaller model size. Nonetheless, all models face challenges inherent in processing Tamil YouTube comments, such as code-mixing and informal language usage. This comparative analysis not only aids in selecting the most suitable BERT model for Tamil language content moderation but also sheds light on the broader landscape of offensive language detection in multilingual contexts.