Concept-enhanced heterogeneous graph network for fact verification.

Fact verification is extremely challenging in natural language processing tasks, requiring the retrieval of multiple evidence sentences from trustworthy corpora to ascertain the accuracy of a given claim. Although the current methods have achieved satisfactory performance, many of them ignore multi-granularity information or fail to fully leverage multi-granularity information, and lack inherent concept information. To tackle the issues, we propose the Concept-Enhanced Heterogeneous Graph Network (Concept-HGN) for fact verification. First, our Concept-HGN model constructs a heterogeneous graph to aggregate clues from the scattered text across multiple evidence sentences. By building different heterogeneous nodes into an integral unified graph, this hierarchical node granularity enables Concept-HGN to be more effectively applied to fact verification tasks. Then, Concept-HGN leverages the intrinsic concepts of entities from YAGO, guiding fact verification and boosting the fact verification performance. We conducted performance evaluation experiments on the FEVER and UKP Snopes datasets. On the FEVER dataset, our proposed Concept-HGN model achieved 80.26 % and 77.68 % on LA and FS, respectively. On the UKP Snopes dataset, the accuracy and macro F1 also reached 65.7 % and 61.9 %, respectively. The experimental results on these datasets indicate that the Concept-HGN model proposed in this paper outperforms the baseline models and achieves state-of-the-art performance on the task of fact verification.