Leveraging Contextualized and Word Embeddings for Arabic Hate Speech Detection

With the increasing spread of social media platforms, the issue of hate speech has become a significant concern. The anonymity offered by these platforms, combined with the lack of strict regulations, has led to a rise in harmful content, and Arabic content on these platforms is no different. This study focuses on detecting hate speech in Arabic by leveraging semantic embeddings and deep learning techniques to address these issues. The study combines embeddings generated from AraBERT and AraVec to capture contextual and static features of Arabic text with Long Short-Term Memory (LSTM) and Multi-Head Attention-based models. The models were trained and tested using the arHate dataset. The results showed that the LSTM model achieved an accuracy of 91.6%, while the Multi-Head Attention model achieved an accuracy of 91.4%. Further evaluation conducted using a newly collected Arabic tweet dataset demonstrated strong performance of the models, with the LSTM model achieving 95.6% accuracy and the Attention-based model achieving 95.2%.