Abstract Existing multimodal fake news detection methods face two persistent challenges: insufficient fine-grained semantic alignment between images and text, and a lack of cross-verification against external factual knowledge. To address these limitations, this paper proposes Evidence-Augmented Scene graphs and Interactive Embedding Reasoning, a multimodal fake news detection framework that integrates evidence enhancement, scene graph reasoning, and local semantic interaction. The framework operates through a hierarchical reasoning architecture: first, an evidence alignment and filtering mechanism retrieves and integrates external textual and visual evidence from the Internet to enable fact-cross-verification, particularly for cases where the news content is internally self-consistent yet contradicts external facts. Second, visual and textual scene graphs are constructed and enriched with external knowledge graphs to achieve structured semantic modeling of multimodal content. Third, an image-enhanced bidirectional GRU network with a cross-attention mechanism realizes token-level local semantic interaction between images and text, capturing subtle cross-modal inconsistencies that may be missed by coarser-grained analysis. Experiments on Chinese and English public datasets including Weibo and Twitter demonstrate that EASIER achieves competitive accuracy and F1 scores compared with state-of-the-art baselines, with notably strong recall on fake news (95.12% on Weibo). The hierarchical reasoning framework—from factual evidence through structured semantic reasoning to local semantic interaction—enhances the detection robustness of the model in complex multimodal contexts, particularly for deception categories involving factual falsehood and image manipulation where content-only methods struggle.
Li et al. (Mon,) studied this question.