Arabic hate speech detection based on BERT models variants

Hate speech on social media, particularly in Arabic contexts, poses a serious threat to user well-being and online interaction quality. The complexity of Arabic, with its wide range of dialects, necessitates advanced detection systems to identify harmful communication targeting individuals or groups based on ethnicity, religion, gender, or nationality. To address this challenge, three key Twitter datasets — OSACT-5, LAHS, and arHate — are examined to represent diverse dialects and sociopolitical contexts. These datasets provide valuable resources for investigating the binary classification task of Arabic hate speech detection. Leveraging BERT transformers with their bidirectional contextual understanding enables the capture of nuanced meanings in Arabic expressions, thereby enhancing classification accuracy. ix prominent Arabic BERT variants — MarBERT, BERT-multilingual, QARiB, CAMeLBERT, AraBERTv0.2-Twitter, and SaudiBERT — are systematically evaluated across the selected datasets. To enhance performance, hyperparameter optimization using Grid Search and Bayesian methods is conducted only on the top three performing models. The optimized models achieve strong results: on OSACT-5, QARiB attains 93.2% accuracy with an F1-score of 82%; on LAHS, QARiB reaches 85.0% accuracy and 87.0% F1; and on arHate, SaudiBERT achieves 94.5% accuracy with a 91.8% F1-score. These results highlight the robustness and adaptability of optimized Arabic BERT models for hate speech detection across diverse dialects and imbalanced datasets, contributing to more reliable moderation of harmful content in Arabic social media. • A comprehensive evaluation of six Arabic BERT models for hate speech detection. • Research addresses challenges from dialectal variations and noisy user-generated content. • A new error analysis identifies the strengths and weaknesses of each BERT model. • The study contributes to safer online environments for Arabic-speaking users.