Cyberbullying Detection in Arabic Text Using Different Deep Learning Approaches

The rise of social media has enabled the rapid spread of user-generated content across various domains, including advertising, entertainment, politics, and economics. However, this growth has also facilitated the increase of harmful behaviors, notably cyberbullying. Addressing this issue requires advanced emotional and sentiment analysis techniques. In this study, the ArCyC (Arabic Cyberbullying Corpus) were integrated with Twitter data to develop robust models for cyberbullying detection. Several deep learning models have been suggested which are: Deep Neural Net. (DNN), Convolutional Neural Nets. (CNN), Recurrent Neural Net. (RNN), Hybrid CNN+RNN, BERT & AraBERT. Text and Emoji have been experimented in the dataset. Models’ performance was evaluated based on accuracy. Experiments using the ArCyC dataset demonstrated that both textual and symbolic elements contributed significantly to classification accuracy. In contrast, analysis with the ArCyC dataset revealed that textual features had a more dominant influence due to the limited use of emojis. The results underscore the effectiveness of deep learning approaches in detecting cyberbullying within Arabic social media content. AraBERT for text has obtained the highest accuracy equal to 95%, similarly LSTM obtained the same accuracy for both text and emoji’s