Hate and Offensive Speech Detection on Arabic Social Media

Abstract

We are witnessing a proliferation of hate speech on social media targeting individuals for their protected characteristics, including ethnicity, religion, gender, and nationality. Our research focuses on devising effective Arabic hate and offensive speech detection frameworks to address this serious issue. In the first part of the thesis, we aim to improve Arabic hate speech detection systems and present our efforts at building binary and multi-class (3-class and 6-class) hate and offensive speech datasets using four robust extraction strategies that we implement based on the four types of hate: religion, ethnicity, nationality, and gender. Next, we develop several 2-class, 3-class, and 6-class machine and deep learning classification models that we train on different feature spaces using a variety of feature extraction techniques. We also investigate how we can develop single and ensemble machine and deep learning models for hate speech detection and conduct extensive experiments to assess the performance of the various learned models on unseen data. The performance outcome is very encouraging compared to prior hate speech studies carried out on Arabic and English corpora. Furthermore, we examine the word-embedding models’ effect on the neural network’s performance since they were not adequately examined in the literature. Through 2-class, 3-class, and 6-class classification tasks, we investigate the impact of both word-embedding models and neural network architectures on predictive accuracy. We first train several word-embedding models on a large-scale Arabic text corpus. Next, based on our Arabic hate and offensive speech dataset, we train multiple neural networks for each detection task using the pre-trained word embeddings. This task yields a large number of learned models, which allows conducting an exhaustive comparison. One key for improving hate speech detection performance is to have a textual training corpus that is vast and confidently labeled. Thus, in the second part of this thesis, we explore how we can improve hate speech detection and leverage the abundant social media content based on the recent success of semisupervised learning techniques. In particular, we explore two new research directions: (1) adopting semi-supervised self-learning to create a large-scale hate speech corpus and use it to improve hate speech detection models; and (2) build ensemble-based semi-supervised learning systems based on the machine and deep learning models. We empirically demonstrate the effectiveness of these approaches and show that our semi-supervised approaches improve classification performance over supervised hate speech classification methods.