Short video rumor detection based on causal graph

In recent years, the short video industry has experienced rapid growth, leading to the emergence of numerous knowledge-based rumors. These rumors often disguise themselves as professional knowledge, making it difficult for fact-checkers to identify their falsehoods without external expertise. Furthermore, existing Chinese short video rumor datasets lack support from external knowledge. To solve that problem and apply to a real-world scenario, this paper constructs a Chinese short video rumor dataset from Douyin, which is the largest short video platform in China, and build a related rumor evidence base. To further characterize the knowledge association between short video entities which is important for the interpretation of knowledge distortion, this paper also constructs causal relationships between entities using causal discovery algorithms. Finally, to tackle and visualize the knowledge distortion in social media short videos, this paper proposes a Causal Short Video Rumor Pretrain Model (CSVRPM). This model obtains relevant causal subgraphs from the causal knowledge repository and integrates the causal relationships within these subgraphs using an attention mechanism in the short video rumor detection model. The experiment results show that the model outperforms some state-of-the-art approaches and greatly improves the interpretability of short video rumor detection results.