Abstract
The user-generated content on the internet including that on social media may contain offensive language and hate speech which negatively affect the mental health of the whole internet society and may lead to hate crimes. Intelligent models for automatic detection of offensive language and hate speech have attracted significant attention recently. In this paper, we propose an automatic method for detecting offensive language and fine-grained hate speech from Arabic tweets. We compare between BERT and two conventional machine learning techniques (SVM, logistic regression). We also investigate the use of sentiment analysis and emojis descriptions as appending features along with the textual content of the tweets. The experiments shows that BERT-based model gives the best results, surpassing the best benchmark systems in the literature, on all three tasks: (a) offensive language detection with 84.3% F1-score, (b) hate speech detection with 81.8% F1-score, and (c) fine-grained hate-speech recognition (e.g., race, religion, social class, etc.) with 45.1% F1-score. The use of sentiment analysis slightly improves the performance of the models when detecting offensive language and hate speech but has no positive effect on the performance of the models when recognising the type of the hate speech. The use of textual emoji description as features can improve or deteriorate the performance of the models depending on the size of the examples per class and whether the emojis are considered among distinctive features between classes or not.