WeakNER-JointAraBERT: Entity-Aware Multi-Task Learning for Arabic Fake News Detection with Weakly Supervised NER

The proliferation of misinformation across Arabic digital media constitutes a significant threat to social stability and informed public discourse. Despite recent advances in transformer-based natural language processing, Arabic fake news detection remains constrained by linguistic complexity, limited annotated resources, and the absence of scalable entity-level supervision. This paper proposes WeakNER-JointAraBERTs, a novel entity-aware multi-task learning framework that jointly optimizes document-level fake news classification and auxiliary named entity recognition (NER) without requiring manually annotated entity labels. The framework employs an AraBERTv02 shared encoder, a binary classification head, and a Conditional Random Field (CRF)-based NER branch trained on automatically generated BIO-format pseudo-labels produced by a pretrained Arabic NER teacher model. To rigorously position the proposed approach, this study also presents a PRISMA-inspired systematic review of 28 fake news detection studies published between 2019 and 2026, organized into five thematic clusters. Experiments conducted on the Palestine Fake News Dataset (PFN; 2368 samples, 27.5% fake) demonstrate that WeakNER-JointAraBERT achieves 92.41% accuracy, 90.4% Macro-F1, and 98.1% ROC-AUC, outperforming both AraBERTv02 and Asafaya Arabic BERT single-task baselines in accuracy, Macro-F1, and ROC-AUC. Ablation studies confirm that NER loss weight calibration (λ=0.5) is critical to performance. Error analysis reveals that entity-aware supervision reduces misclassification on organization- and location-referencing articles. The proposed framework has direct applications in Arabic content moderation platforms, social media monitoring systems, and fact-checking pipelines operating under annotation scarcity, providing a scalable foundation for misinformation detection across the Arabic-speaking world’s 422 million users.