Arabic hate speech detection system based on AraBERT

Tunisia has entered a phase freedom of speech with access to social media since the Jasmine Revolution in 2011. Toxic contents such as abusive and hateful speeches have become omnipresent on Tunisian social media. Considering the side effects of these toxic contents on the psychology of users, it is necessary to detect them automatically. The dialect of Tunisian is underrepresented. As a consequence, there is not enough data set. In this paper, we present the data collection process with the aim of having a Tunisian reference dataset, to evaluate different models of hate speech and abuse detection. We also present our neural network model based on AraBERT. Our experimental results on our dataset shows that the AraBERT model performs better with an F1 score of 0.99.