Neural Networks and Sentiment Features for Extremist Content Detection in Arabic Social Media

The proliferation of extremist content on social media poses critical threats to societal stability, necessitating advanced detection mechanisms. Despite substantial research on extremist content detection in various languages, Arabic remains significantly underexplored. Recognizing the pivotal role of social media, this study introduces a novel approach to detecting extremist posts in Arabic by leveraging neural networks. The proposed models utilize Arabic Bidirectional Encoder Representations from Transformers (AraBERT), Multi-Layer Perceptron (MLP), and Sentiment Features (SFs). Among the tested models, the optimal configuration-fine-tuning AraBERT with integrated MLP and SF-achieved an impressive 98% accuracy in detecting extremist Arabic tweets. Additionally, the model demonstrated robust performance when evaluated on real-world extremist posts from VKontakte, achieving 81% accuracy. These findings underscore the effectiveness of combining AraBERT, MLP, and SF in improving extremist content detection and highlight the potential of neural network-based solutions in combating harmful online content.