Cyberbullying detection in low-resource languages remains critically underexplored despite the urgent need for automated content moderation systems. Bengali, the world’s seventh most spoken language with over 230 million speakers, particularly suffers from inadequate computational resources for cyberbullying detection, leaving millions of users vulnerable to online harassment. We address this gap by introducing BanCyB , a comprehensive multi-label Bangla cyberbullying corpus comprising 10,000 manually annotated social media texts across four distinct yet overlapping categories: Trolling, Insult, Hate Speech, and Targeted Harassment. Unlike existing binary classification approaches, our multi-label framework captures the complex, overlapping nature of cyberbullying behaviors prevalent in real-world scenarios. We propose a Hybrid Transformer–LSTM Feature Fusion (HTLFF) architecture that integrates BanglaBERT’s deep contextual representations with LSTM-based sequential modeling to effectively capture both semantic nuances and temporal dependencies in Bangla cyberbullying text. Comprehensive evaluation shows that HTLFF achieved the higher scores on the test set under our evaluation protocol, obtaining a 92.37% F1-score for binary classification and 87.05% macro-F1 and 88.03% micro-F1 for multi-label classification compared to traditional machine learning, deep learning, and transformer-based baselines. The LIME-based and Layer Integrated Gradients (LIG) interpretability analyses reveal category-specific linguistic patterns, offering deeper insight into how different forms of Bangla cyberbullying manifest in text. These findings highlight the improved effectiveness of HTLFF within our evaluation setting and provide valuable resources for advancing NLP research in low-resource languages. The codebase is available through the publicly accessible GitHub repository at https://github.com/mynul11/BanCyB.
Hasan et al. (Sun,) studied this question.