Transfer Learning Model for Cyberbullying Detection in Tunisian Social Networks

Due to the proliferation of smartphones connected to the Internet, many individuals, particularly young people in Arab society, have widely embraced social media platforms as the primary means of communication, interaction, and friendship formation. Technological advancements in smartphones and communication have enabled young people to stay in touch and join massive social networks worldwide. However, such networks expose young individuals to cyberbullying and offensive content, endangering their safety and emotional well-being. Although numerous solutions have been proposed for automatically detecting cyberbullying, most existing solutions have been designed for English-speaking users. Morphologically rich languages, such as Arabic, specifically the Tunisian dialect, present challenges of data scarcity. As a result, solutions developed for another language prove ineffective when applied to Arabic content. With this in mind, this study aims to enhance the effectiveness of existing cyberbullying detection models for Arabic content by designing and developing a cyberbullying detection model. A diverse set of heterogeneous classifiers derived from traditional machine learning and deep learning and transformer learning techniques were trained using annotated Arabic cyberbullying datasets collected from three different platforms (Facebook, Twitter, and YouTube). The results demonstrate the efficacy of the proposed model compared to other examined classifiers. The overall improvement achieved by the proposed model reaches 85% compared to the best-trained classifier.