Robust Cyberbullying Detection in Diverse Textual Noise

In the digital age, cyberbullying has been highlighted by studies in sociology and psychology as a pervasive and complex phenomenon. Despite advancements in big data and machine learning, Cyberbullying Detection faces challenges due to the diversity of textual noise and ambiguity in online communications. Existing methods, relying on fixed rules for noise mitigation, struggle with the dynamic complexity of online text. In our study, we propose the Coherence Capsule Net (CoCapsNet), which integrates the capsule networks with coherent robustness training for cyberbullying detection. Specifically, We utilize Attention-based Compositional Embedding (ACE) for text embedding, which balances sentence semantic representation and efficiency for cyberbullying detection on massive social media data. Then, for fine-grained semantics mining from noisy texts, we use the capsule network for capturing the key features of the text by iteratively aggregating lower-level capsule vectors through k-means routing. Finally, we employ a coherent robustness training method based on R-Drop regularization to enhance the models robustness against textual inconsistencies and noise in the cyberbullying detection domain. We conduct experiments in four datasets to validate our CoCapsNet approach, the experiments demonstrate our CoCapsNet outperforms the baseline models, indicating CoCapsNet's advanced effectiveness in cyberbullying detection.