A Context Aware Deep Learning Framework for Multi-Level Arabic Fake News Detection Using Hybrid Linguistic Representations

Abstract

The rapid spread of false information on social media has made it even more important to be able to spot false news, especially in Arabic-speaking areas where language is more complicated and news is organized in a hierarchy. Recent research show that deep learning and transformer-based models work well, but most of them only look at performance metrics and don't look at how alternative contextual representations affect detection behavior. This study presents a context-aware deep learning architecture that amalgamates semantic representations from AraBERT with syntactic characteristics obtained from part-of-speech tagging and emotional indicators to facilitate multi-level Arabic false news detection. The JoNewsFake dataset is used in an ablation-based experimental design to look at how each contextual component affects the main category, subcategory, and fake/real classification levels. The findings indicate that although semantic embeddings offer a solid base, the integration of syntactic and emotional context markedly increases robustness, diminishes ambiguity in nuanced subcategories, and facilitates the differentiation between authentic and fabricated news. Analysis led by visualization shows again how adding information can help fix mistakes that come from rhetorical framing and implicit sentiment. This study transitions from benchmark-oriented evaluation to contextual effect analysis, underscoring the necessity of thorough context modeling for dependable and interpretable Arabic false news detection systems.