Cyberbullying Detection in Hinglish Text Using MURIL and Explainable AI

Code-mixed languages require specialized AI models: The paper demonstrates that existing cyberbullying detection systems designed for single languages perform poorly on Hinglish (Hindi-English mixed text). The MURIL model specifically designed for Indian languages outperformed general multilingual models like RoBERTa by 1.36 to 13.07 percentage points, showing that one-size-fits-all AI solutions can miss important safety issues in diverse linguistic communities.

Explainability is essential for content moderation: The framework includes "attribution analysis and cross-linguistic pattern recognition" to show why content was flagged as cyberbullying. This transparency is crucial for responsible AI deployment, as it allows platforms to explain moderation decisions to users and helps identify potential biases in automated systems.

Cultural context and nuance remain challenging for AI: The failure analysis reveals that AI systems struggle with "context-dependent interpretation, cultural understanding, and cross-linguistic sarcasm." This highlights a key limitation: automated systems cannot yet fully replace human judgment in moderation, particularly for culturally-specific forms of harassment, and should be used as tools to assist rather than replace human reviewers.

Performance varies significantly across different datasets: Accuracy ranged from 75.41% to 94.63% depending on the dataset used, indicating that cyberbullying detection systems may perform inconsistently across different platforms or user communities, which has implications for equitable protection across diverse online spaces.