Explainable Multi-Granularity Attribution Reasoning Framework for Fake News Detection

Fake news detection has garnered the attention of an increasing number of researchers in recent years, particularly in the context of multimodal fake news that combines text and images. However, existing methods only focus on cross-modal feature fusion guided by a consistency matrix and make predictions based on shallow semantic modeling. This type of method relies on the invisible interaction of internal features to output results, and thus has a strong black-box characteristic. Furthermore, the meticulously fabricated fake news that exists in the real world is heterogeneous. In order to improve the influence of fake news, forgers will package it in many ways to deceive the existing detectors. To address the above issues, we first investigate the hidden psychological motivations behind fake news, and propose an Explainable Multi-granularity Attribution Reasoning framework for Fake News Detection, named EMAR-FND. Specifically, four fine-grained hierarchical reasoning networks are included in the proposed framework. They perform attribution reasoning for fake news from different perspectives, aiming to find subtle manipulation details. Finally, the intermediate layer features from different perspectives are aggregated through a multi-granularity information fusion module. Experimental results demonstrate that our proposed EMAR-FND outperforms existing state-of-the-art fake news detection methods under the same settings. Furthermore, we further verify the explainability of the proposed model through a discrimination performance experiment.