INTEGRATING SEMI-SUPERVISED LEARNING AND TRANSFORMER MODELS FOR OFFENSIVE LANGUAGE DETECTION IN ENGLISH–TAMIL CODE-MIXED SOCIAL MEDIA

In todays world, social media has evolved far bigger than the press media, becoming the dominant medium for public expression. Despite the extraordinary growth, social media involves the prevalence of offensive language and harmful content. As there is no real content moderation or compliance checks, social media has become an open market for trading offenses. lack of stringent content moderation and compliance checks, social media has become a place where individuals are able to express things that are objectionable. In terms of translation requirements, particularly in low-resource code-mixed language pairs, such as English–Tamil, detection of offensive verbiage is quite a challenge due to factors, such as variations in scripts, irregular spellings, and non-availability of reliable large annotated corpora. This paper describes automated annotation pipeline for the detection of offensive language in English–Tamil translations using the XLM-RoBERTa-large transformer model. The goal of the proposed system is to generate high-quality multi-class offensive-language labels—without manual annotation. This model has been optimized for multi-class classifications in seven categories: profanity, hate speech, sarcasm, threats, harassment/bullying, disrespect/dismissive, and non-offensive. We hypothesize that a semi-supervised, confidence-tiered annotation process, combined with lexicon-based category mapping, public datasets that are labelled, machine-translated resources, and synthetic data generation, can achieve annotation quality comparable to human labelling. The imbalance in classes is addressed through back-translation, generation of synthetic data, and optimization of class-weighted loss. Experimental results indicate that the model can achieve a Macro-F1 score of 92.4% and a Cohen’s Kappa of 0.92, demonstrating near-human annotation quality.