Emotion Trajectory Modeling: A Hybrid Transformer-CNN Approach for Cyberbullying Detection

Cyberbullying is often limited to the analysis of static and individual messages, an approach that is less effective at detecting the situational nature of online bullying and its escalation. The proposed study presents another theory Emotion Trajectory Modeling, which assesses the temporal dynamics of emotional states in a whole conversation. To perform this analysis, a hybrid deep learning model, which combines Transformer network and Convolutional Neural Network (CNN), is suggested. Transformer component to infer long-range contextual dependencies across the conversational sequence is used to model the emotional arc over time. At the same time, the CNN derives strong local features and hierarchical patterns of individual messages, distinguishable textual identifiers of abusive language. Such a two-architecture system will allow the system to detect sudden, negative changes in the emotional tone of a conversation, as they are one of the key indicators of bullying. The Multi-Label Twitter Dataset on Online Abuse was used to train and test the model. The classification accuracy is proven by experimental validation to be 96%. Such a good performance confirms the hypothesis that a dynamic analysis based on trajectories is better than the sole inspection of messages. The combination of global context awareness (Transformer) with local feature extraction (CNN) introduces a more accurate and reliable system of identifying complex types of cyberbullying.