EmoFuse: Emotion-Aware Fusion and Adaptive Expert Routing for Multi-Domain Arabic Fake News Detection

Abstract

Arabic fake-news detection must account for linguistic variation and differences between content domains. We propose EmoFuse, a MARBERT-based framework that combines two independently fine-tuned veracity experts with an emotion-specialized contextual representation and a domain-supervised routing network. Each expert fuses a 128-dimensional veracity vector with a 768-dimensional emotion representation, while the gate assigns input-dependent weights to expert predictions. Adaptive routing improves Macro-F1 by an average of 4.78 percentage points over uniform expert averaging. The primary system achieves a macro-F1 of 0.8735 on the historical official test set. The results support adaptive routing across the two observed dataset-defined domains, but they do not provide consistent evidence that emotion specialization adds complementary information beyond veracity-orientated or generic contextual representations.