A Fusion of context-aware based BanglaBERT and Two-Layer Stacked LSTM Framework for Multi-Label Cyberbullying Detection

Researchers developed a hybrid AI model combining BanglaBERT and stacked LSTM networks to detect multiple types of cyberbullying in Bangla text simultaneously. The approach addresses limitations in existing single-label classification methods by recognizing that comments can contain overlapping forms of abuse like threats, hate speech, and harassment.

• →Multi-label cyberbullying detection is more realistic than single-label approaches as comments often contain overlapping abuse types.
• →The fusion architecture combines BanglaBERT-Large with two-layer stacked LSTM to capture both context and sequential dependencies.
• →Low-resource languages like Bangla face challenges due to limited robust pre-trained models for cyberbullying detection.
• →The model addresses four categories: cyberbullying, sexual harassment, threats, and spam using various sampling strategies.
• →Comprehensive evaluation includes multiple metrics and 5-fold cross-validation to assess model generalization.