Abstract
Religious hatred is a serious problem on Arabic Twitter space and has the potential to ignite terrorism and hate crimes beyond cyber space. To the best of our knowledge, this is the first research effort investigating the problem of recognizing Arabic tweets using inflammatory and dehumanizing language to promote hatred and violence against people on the basis of religious beliefs. In this work, we create the first public Arabic dataset of tweets annotated for religious hate speech detection. We also create three public Arabic lexicons of terms related to religion along with hate scores. We then present a thorough analysis of the labeled dataset, reporting most targeted religious groups and hateful and non-hateful tweets' country of origin. The labeled dataset is then used to train seven classification models using lexicon-based, n-gram-based, and deep-learning-based approaches. These models are evaluated on new unseen dataset to assess the generalization ability of the developed classifiers. While using Gated Recurrent Units with pre-trained word embeddings provides best precision (0.76) and F-1 score (0.77), training that same neural network on additional temporal, users, and content features provides the state-of-the-art performance in terms of recall (0.84).