Complementary Graph Learning and Prompt-based Cross-modal Generation for Missing-modality Fake News Detection

Multi-modal fake news detection (MFND) has attracted increasing attention. However, due to information loading failure or access restriction, incomplete modality makes joint multi-modal information extraction be challenging. Existing MFND methods with missing-modality focus on specific missing-modality cases, and how to flexibly deal with various missing-modality cases of news in the real world has not been well studied. In this paper, we propose a novel fake news detection approach named Complementary Graph learning and Prompt-based cross-modal Generation network (CG-PG), which contains two main modules: a complementary graph learning module and a prompt-based cross-modal generation module. The complementary graph learning module explores structural complementary information in image and text graphs to implement cross-modal information propagation. To further recover the information loss caused by missing modalities, the prompt-based cross-modal generation module generates representations of the missing modality from available modalities and imposes task-related constraints on the representations with the missing-aware prompts. Experimental results on the public Weibo and Fakeddit datasets under various missing-modality cases show that CG-PG outperforms state-of-the-art related works.