Class imbalance-sensitive approach based on PLMs for the detection of cyberbullying in English and Arabic datasets
Azzeddine Rachid Benaissa, Azza Harbaoui, Hajjami Henda Ben Ghezala · Behaviour and Information Technology · 2024
Social Networking increases allowed the spreading of cyberbullying worldwide. The latter invaded cyberspace, kids and adolescents are no more safe in their virtual playgrounds. Indeed, online bullying is attracting considerable concern due to the societal and health issues it causes, ranging from depression, anxiety, and low self-esteem to sui cide attempts. Automatic cyberbullying detection is becoming a vital factor in protecting individuals’ lives. It has received much attention in the last decade. Researchers use machine learning and deep learning models to detect online bullying content. An automatic cyberbullying detection model would flag any bullying text as efficiently as possible. Yet, several challenges lie ahead for the development of such a robust model. Our study discerned class imbalance and bullying text representation as being the major issues concerning cyberbullying classification. In this context, we tried to handle the class imbalance problem through data augmentation, cost-sensitive learning, and lever- aging a Computer Vision loss function for the task. Moreover, we consider a prominent solution for bullying content representation, which consists of fine-tuning Pre-trained Language Models for cyberbullying detection and using these latter as feature extractors for Multichannel ConvNets and Bidirectional LSTMs. The results show the effectiveness of the proposed models, which outperform several past works and provide high Recall values (78%–96%) on English and Arabic datasets.