Explainable Real-Time Cyberbullying Detection Using Transformer Models

Abstract

Cyberbullying has become a serious issue on online communication platforms, often affecting the emotional wellbeing and mental health of users. With the increasing use of digital communication systems, there is a growing need for intelligent methods to identify and control harmful online behaviour in real time. This paper presents an explainable realtime cyberbullying detection system integrated into a web-based chat application. The proposed system uses a transformerbased model to identify toxic, abusive, and hateful messages during user interactions. The application is developed using the Django web framework and incorporates the trained model to analyse messages before storing them in the database. To improve transparency and understanding of the model predictions, the system also provides explainable outputs for detected harmful messages. Whenever a message is classified as toxic, the warning count of the corresponding user is increased automatically. If the number of warnings reaches five, the user account is blocked to reduce repeated misuse of the platform. By combining transformer-based text analysis, explainable prediction mechanisms, and automated moderation, the proposed system helps create a safer and more responsible online communication environment.