BILINGUAL CYBERBULLYING DETECTION IN ENGLISH AND HINDI: A COMPARATIVE TRANSFORMER ARCHITECTURE STUDY

Cyberbullying represents a critical public health crisis, particularly for young adults (18-35) in India, where the psychological impact is severe and detection is hindered by linguistic diversity and resource scarcity in Hindi. This paper addresses the necessity for robust, automated cyberbullying detection across both high-resource English and low-resource Hindi text environments. We propose the Bilingual Architecture Selection Framework (BASF), featuring the Optimized RoBERTa Transfer Algorithm (ORTA), a novel methodology for deploying optimal transformer architectures based on language resource level. We conduct a comprehensive empirical benchmark comparing four distinct transformer architectures—RoBERTa, XLNet, DistilBERT, and GPT-2—on a large English multi-class corpus and a small Hindi binary corpus. The results show a significant architectural trade-off: the highly efficient DistilBERT achieved the highest Macro F1-Score on the English task (86.43%), validating the use of distilled models for high-resource efficiency. Crucially, the RoBERTa model, operating under the ORTA, achieved a decisive and counter-intuitive peak performance of 92.88% on the low-resource Hindi corpus, vastly outperforming XLNet (68.61%) and GPT-2 (67.70%). This finding demonstrates that RoBERTa's superior optimization is the most critical factor for successful cross-lingual knowledge transfer, establishing the ORTA/RoBERTa as the preferred model for maximizing efficacy in the resource-constrained Hindi domain. The BASF provides an evidence-based roadmap for designing dual-strategy bilingual cyberbullying detection systems.