Improving Arabic Hate Speech Identification Using Online Machine Learning and Deep Learning Models

Abstract

Due to the rising use of social media platforms on a global scale to interact and express thoughts freely, the spread of hate speech has become very noticeable on these platforms. Governments, organizations, and academic institutions have all spent substantially on discovering effective solutions to handle this issue. Numerous researches have been performed in several languages to find automated methods for identifying hate speech, but there has been minimal work done in Arabic. The findings of a performance evaluation of two machine learning models, namely the passive-aggressive classifier (PAC) and the Bidirectional Gated Recurrent Unit (Bi-GRU) augmented with an attention layer, are investigated in this work. Proposed models are developed and evaluated using a multi-platform Arabic hate speech dataset. We employ term frequency-inverse document frequency (TF-IDF) and Arabic word embeddings for feature extraction techniques after running a variety of pre-processing steps. The experimental results reveal that the two proposed models (PAC, Bi-GRU with attention layer) provide an accuracy of 98.4% and 99.1%, respectively, outperforming existing methods reported in the literature.