Public health fact verification based on semantic hierarchy graph

The task of fact verification is to evaluate the truth of a claim-focused on plausible evidence, verifying the truth of a claim in public health is challenging because the need for reasoning across various retrievable pieces of evidence. Most fact-checking studies typically focus only on political claims, while few studies address fact verification on other topics. Therefore, this paper chooses to study fact verification in public health. To support this case study, this paper employs PUBHEALTH, a dataset containing claims related to public health. In this study, this paper presents an approach based on a semantic hierarchy. Different from most methods, which represent evidence sentences by concatenating or merging the characteristics of isolated evidence sentences, our method obtains the complex semantic hierarchy of evidence sentences through semantic analysis. Specifically, based on the XLNet model, the graph structure is initially utilized to redefine the relative distance of words. Then, a graph attention network and multi-layer graph convolutional network are used to propagate and aggregate information from neighboring nodes on the graph. Experiments on the PUBHEALTH dataset show that the proposed method achieves 76.2% in F1 score, which is substantially higher than the current highest score, indicating the effectiveness of the fact verification task based on semantic hierarchy.