Using Deep Learning Techniques to Detect Hate and Abusive Language in Arabic Tweets

Billions of people spend hours on social media platforms every day. While there are numerous known benefits of social media, hate speech and abusive language on social media platforms have become an increasingly serious social problem affecting individuals and societies’ psychological state. Detecting and preventing hate speech and abusive language is a crucial task for healthy and safety digital communication. To overcome this important social problem, we propose four various deep learning and pre-trained models: (i) Ensemble deep learning model, (ii) Multilingual BERT model, (iii) Arabic BERT model, and (iv) ALBERT model to detect hate speech and abusive language in Arabic text. To that end, we utilized the L-HSAB Arabic dataset to build our models, and we utilized the OSACT dataset to evaluate the generalizability of our model’s architecture. Our model tackles two classification tasks: a binary classification and a multiclass classification task. Our Arabic Bidirectional Encoder Representations from Transformers (BERT)-based model achieved the best performance on the binary classification task with an F1-score of 90.8%. While our Multilingual BET-based model obtained the best result on the multi-class classification task with an F1-score of 80.0%. Finally, the generalizability experiment of our best-performing mod-el on the binary classification task achieved an F1-score of 90.2% on the OSACT dataset.