Chinese cyberbullying detection via multi-feature prompt learning

Abstract

Most existing research on cyberbullying detection reduces the task to comment-level hate speech classification, overlooking its collective, event-driven, and dynamic nature. To address this limitation, we propose CDMPL, a multi-feature prompt learning framework for Chinese cyberbullying incident detection. CDMPL integrates textual features (semantic summaries generated by large language models) with temporal features (temporal behavioral patterns captured by a Dynamic Attention Truncation mechanism and an Autoformer-based time series encoder). These fused features are reformulated as a masked language modeling task, enhancing interpretability and robustness. We evaluate CDMPL on the only publicly available Chinese incident-level dataset, consisting of 86 cyberbullying and 109 non-cyberbullying events. Under a 15-shot setting, CDMPL achieves 84.79% accuracy and 84.05% F1-score, exceeding state-of-the-art large language models such as Doubao (71.91% F1) and ChatGPT-4o (70.06% F1) by more than 12%. Ablation studies further demonstrate the necessity of jointly modeling semantic and temporal features. To our knowledge, this is the first systematic study of Chinese incident-level cyberbullying detection, providing methodological and empirical advances for early detection and governance.