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

Abstract

The complexity, concealment, and context-dependence of online harassment occurring through social networking sites have made it a growing concern. Conventional detection systems frequently ignore bystander participation and with respect to online conversations in favour of treating digital abuse as a straight forward binary problem. This paper introduces a chained neural architecture that combines a bidirectional transformer encoder known as BERT along with the Long Short-Term Memory (LSTM) network to detect fine-grained levels of cyber bullying with the incorporation of by standard dynamics. The model first identifies the part played by standers defenders, instigators, neutrals, or others and then uses this output to predict the aggression level of a cyberbullying incident. Using the publicly available CYBY23 dataset, the introduced model yields an F1-score of 89%, improving detection performance by25.35% compared to traditional binary classifiers. The results show that considering bystander involvement improves context understanding and reduces misclassification between aggression and bullying.