A Chained Deep Learning Model for Fine-Grained Cyberbullying Detection With Bystander Dynamics

Accurate detection of cyberbullying on Social Networking Sites (SNSs) is crucial for online safety, especially for individuals impacted by it. Cyberbullying language is often implicit, necessitating a comprehensive analysis of conversational context to determine intention and severity. Current studies on cyberbully detection predominantly focus on the main post, overlooking the valuable insights provided by bystander reactions. Moreover, confusion over the differences between cyber-aggression and cyberbullying can undermine the reliability of some of these studies. To ameliorate these issues, this paper addresses the gap in existing research by emphasizing the significance of bystander in fine-grained cyberbullying detection. We specifically investigate the influence of bystander for more precise identification of cyberbullying attacks. Our approach involves fine-tuning a pre-trained language model (BERT) to identify features associated with bystander, categorizing them into roles such as defender, instigator, neutral, and other. To enhance fine-grained cyberbullying detection, we propose a classifier chain that combines BERT's output with Long Short-Term Memory (LSTM) networks. Our experiments demonstrate that involving bystander roles increases the model's performance by 25.35%, achieving a final classification F1-score of 89%. Moreover, our approach accurately determines the level of aggression in cyberbullying incidents through considering the interdependency of bystander roles and cyberbullying classes labels, providing an innovative solution to the challenge of cyberbullying misclassification. In summary, our study highlights the significance of incorporating bystanders as a feature in cyberbullying detection and introduces a chained fine-grained model that surpasses conventional approaches, demonstrating promising outcomes in precisely classifying cyberbullying incidents.