Design of a Real-Time, Multilingual, Emotion-Aware Cyberbullying Detection System using Multi-Teacher Knowledge Distillation and Explainable AI

Abstract

Social media cyberbullying has propagated rapidly and is being experienced by individuals worldwide. It tends to be expressed using sarcasm, emotional language, and multiple languages, making it difficult to determine the identity of the perpetrator. Although automated detection systems are becoming increasingly prevalent, the majority of existing systems suffer from language issues, function only in offline batch mode, and are black-box models that cannot be interpreted. These constraints make it more difficult to intervene with speed and transparency.

This paper offers a real-time, multilingual system for detecting cyberbullying, using explainable AI, emotion and sarcasm detection, and Multi-Teacher Knowledge Distillation (MTKD) to address shortcomings.

The system leverages an ensemble of transformer-based teacher models, like mBERT, XLM-R, and IndicBERT, to capture language-specific features. Then, the models collaborate to produce a lightweight XG Boost classifier. To assist with the interpretation of context, additional layers are incorporated to identify sarcasm and emotion. SHAP (SHapley Additive Explanations) is employed to provide each prediction token-level interpretability. Algorithmic and architectural design of a system that would form a transparent, efficient, and deployable solution to detect cyberbullying in different emotional and linguistic contexts is the focus of this study.