A Hybrid DeBERTa and Gated Broad Learning System for Cyberbullying Detection in English Text

The proliferation of online communication platforms has created unprecedented opportunities for global connectivity while simultaneously enabling harmful behaviors such as cyberbullying, which affects approximately 54.4% of teenagers according to recent research. This paper presents a hybrid architecture that combines the contextual understanding capabilities of transformer-based models with the pattern recognition strengths of broad learning systems for effective cyberbullying detection. This approach integrates a modified DeBERTa model augmented with Squeeze-and-Excitation blocks and sentiment analysis capabilities with a Gated Broad Learning System (GBLS) classifier, creating a synergistic framework that outperforms existing approaches across multiple benchmark datasets. The proposed ModifiedDeBERTa + GBLS model achieved good performance on four English datasets: 79.3% accuracy on HateXplain, 95.41% accuracy on SOSNet, 91.37% accuracy on Mendeley-I, and 94.67% accuracy on Mendeley-II. Beyond performance gains, the framework incorporates comprehensive explainability mechanisms including token-level attribution analysis, LIME-based local interpretations, and confidence calibration, addressing critical transparency requirements in automated content moderation. Ablation studies confirm the meaningful contribution of each architectural component, while failure case analysis reveals specific challenges in detecting implicit bias and sarcastic content, providing valuable insights for future improvements in cyberbullying detection systems.