Abusive Language Detection in Bengali Social Media: A Multi-Modal Ensemble Approach

The growth of Bengali social media content requires robust automated moderation for abusive language. We present a multi-modal ensemble for Bengali abusive-language detection that combines word-level and character-level TF-IDF features with linguistic cues in a stacking model of LinearSVC, Logistic Regression, and XGBoost. We also enrich the BD-SHS dataset with a 15-category abusive-language taxonomy. On the held-out BD-SHS test set, the ensemble reaches 91.2% accuracy and a 90.7% F1-score, with category-level F1 from 0.882 to 0.938 (p<0.001 vs. ablated variants). The system is efficient enough for near real-time use and targets Bengali morphology, code-switching, and cultural context. We discuss practical implications for moderation in low-resource settings.