A Novel Machine Learning-based Framework for Detecting Religious Arabic Hatred Speech in Social Networks

Abstract

Social media platforms generate a huge amount of data every day. However, liberty of speech through these networks could easily help in spreading hatred. Hate speech is a severe concern endangering the cohesion and structure of civil societies. With the increase in hate and sarcasm among the people who contact others over the internet in this era, there is a dire need for utilizing artificial intelligence (AI) technology innovation that would face this problem. The rampant spread of hate can dangerously break society and severely damage marginalized people or groups. Thus, the identification of hate speech is essential and becoming more challenging, where the recognition of hate speech on time is crucial in stopping its dissemination. The capacity of the Arabic morphology and the scarcity of resources for the Arabic language makes the task of distinguishing hate speech even more demanding. For fast identification of Arabic hate speech in social network comments, this work presents a comprehensive framework with eight machine learning (ML) and deep learning (DL) algorithms, namely Gradient Boosting (GB), K-Nearest Neighbor (K-NN), Logistic Regression (LR), Naive Bayes (NB), Passive Aggressive Classifier (PAC), Support Vector Machine (SVM), Ara-BERT, and BERT-AJGT are implemented. Two representation techniques have been used in the proposed framework in order to extract features: a bag of words followed by BERT-based context text representations. Based on the result and discussion part, context text representation techniques with Ara-BERT and BERT-AJGT outperform all other ML models and related work with accuracy equal to 79% for both models.