A Deep Learning Framework for Automatic Detection of Hate Speech Embedded in Arabic Tweets

In this paper, we investigate the ability of CNN, CNN-LSTM, and BiLSTM-CNN deep learning networks to automatically classify or discover hateful content posted on social media. These deep networks were trained and tested using ArHS dataset which consists of 9833 tweets that were annotated to suite hateful speech detection in Arabic. To the best of our knowledge, this is the largest Arabic dataset which handles the subclasses of hate speech. Moreover, we investigate the performance on two existing Arabic hate speech datasets along with ArHS dataset resulting in a combined dataset which consists of 23,678 tweets. Three types of experiment are reported: first, the binary classification of tweets into Hate or Normal, second, ternary classification of tweets into (Hate, Abusive, or Normal), and lastly, multi-class classification of tweets into (Misogyny, Racism, Religious Discrimination, Abusive, and Normal). Using the ArHS dataset, in the binary classification task, the CNN model outperformed other models and achieved an accuracy of 81%. In the ternary classification task, both the CNN and BiLSTM-CNN models achieved the best accuracy of 74%. Lastly, in the multi-class classification task, CNN-LSTM and the BiLSTM-CNN models both achieved the best results with an accuracy of 73%. On the Combined dataset, in the binary classification task, the BiLSTM-CNN achieved an accuracy of 73%. In the ternary classification task, BiLSTM-CNN achieved the best accuracy of 67%. Lastly, in the multi-class classification task, the CNN-LSTM and the BiLSTM-CNN achieved the best accuracy of 65%.