Tunisian Dialect Hate Speech and Abuse Detection with BERT-Based Models

With the rapid growth of social media platforms, harmful content such as hate speech and abusive language has become increasingly widespread online. This type of content can hurt individuals and create negative and unsafe online spaces. While many studies have been done to detect such harmful language in widely spoken languages like English or Modern Standard Arabic, less attention has been given to local dialects that have fewer resources. One of these dialects is Tunisian dialect, which is different from Modern Standard Arabic and can be harder for language models to understand. In this paper, we focus on detecting hate speech and abusive language in Tunisian dialect using deep learning methods. We use a special dataset called T-HSAB, which contains user comments in Tunisian dialect that have been labeled as hateful or abusive. We train and test several BERT-based language models, including AraBERT, ARBERT, MarBERT, and TunBERT. These models were made to understand Arabic and its different forms. Among them, TunBERT, which was trained specifically on Tunisian dialect, gives the best results when identifying toxic comments.