Abstract Cyberbullying is a deliberate and widespread kind of online harassment, which is hard to detect due to its subtextual nature, sarcastic overtones, and situational expressions. Traditional methods of detection, based on the use of key-word filters, manual filtering, or primitive machine-learning algorithms, are often unsuccessful at detecting these complexities, thus limiting the quality and the overall generalizability of such systems. In this study, an integrated framework to detect cyberbullying is proposed and combines the contextual embeddings, sentiment polarity and emotional cues to enhance the interpretative richness and strength. A curated dataset of 99,990 social media posts from Kaggle was pre-processed so that anonymity and linguistic integrity could be guaranteed. Fine-tuned RoBERTa variants were used to extract emotion, sentiment and contextual representations. The resulting embeddings were subsequently reduced using a Variational Autoencoder to create a 64-dimensional latent representation, which was appended with the emotion and sentiment features to create a unified multimodal feature representation. This fused representation was classified with a regularized deep-neural-network with residual connections, batch normalization and dropout as overfitting mitigation strategies. The model achieved 99.49% accuracy and a macro F1 of 0.9941 in internal validation and 95.91% accuracy and a macro F1 of 0.9412 in external evaluation. The interpretation was maintained by using Integrated Gradients which provided clear attribution to predict results. The system was deployed as a browser-extension which demonstrates scalable, real-time detection and transparent moderation; affirming the utility of emotion-aware and context-rich models for cyberbullying detection.
Musembi et al. (Sat,) studied this question.