An algorithmic system for arabic fake news detection using neural networks and transformer embeddings with class weighting.

Modern communication technologies have increased the speed of news dissemination. Information can now reach wide audiences almost instantly. At the same time, these technologies can reduce traditional verification mechanisms, contributing to the accelerated spread of false information and posing potential societal and political challenges. While research on fake news detection has increased, Arabic remains under-explored due to limited datasets and the language’s morphological and semantic complexity. We compiled a multi-source Arabic news dataset consisting of 7,474 articles. The dataset was meticulously validated, achieving high labeling quality (Fleiss’ Kappa = 0.80). Preliminary experiments with baseline algorithms indicated that the neural network (NN) consistently outperformed the other models. This supported its selection as the core classifier in the proposed framework. The proposed system integrates an NN classifier with CAMeLBERT embeddings for semantic feature extraction. A comprehensive comparison was conducted with other prominent Arabic Transformers, including AraBERT, AraELECTRA, and MARBERTv2. We then evaluated multiple imbalance-handling techniques, including class weighting, undersampling, oversampling, and SMOTE. Performance was assessed under different configurations, highlighting the benefits of combining contextual embeddings with resampling strategies. Experimental results indicated that the CAMeLBERT-based neural network with class weighting achieves competitive performance across the evaluated configurations, attaining an F1-score of 96.19%, accuracy of 95.52%, precision of 95.48%, and recall of 96.90% in Arabic fake news detection. These findings indicate that the proposed model provides a reliable basis for automated Arabic fact-checking systems. In addition to predictive performance, the study strengthens methodological rigor through the integration of LIME and SHAP-based interpretability analyses. Future work will focus on assessing cross-domain generalization and investigating the feasibility of real-time deployment.