Reproduce “Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables” on OpenResearch

About this paper

Haisong Gong, Weizhi Xu, Shu wu, Qiang Liu, Liang Wang

HeterFC introduces a novel model for fact checking that performs veracity reasoning over both unstructured text and structured table evidence using a word-level heterogeneous graph and Relational Graph Convolutional Networks (R-GCNs). This approach achieves a Feverous score of 0.3476 and a Label Accuracy of 0.6329 on the FEVEROUS dataset, outperforming existing transformer and homogeneous graph-based methods.

Key insights

• Modeling relationships at a fine-grained, word-level granularity within a heterogeneous graph, with distinct edge types for textual, tabular, and cross-evidence connections, is crucial for effective reasoning over mixed-type evidence.
• Relational Graph Convolutional Networks (R-GCNs) effectively propagate information across diverse relationship types, capturing richer semantic context than homogeneous or sentence-level graphs.
• A combination of attention-based claim-evidence interaction, an assisted loss for evidence relevance, and fused predictions (graph-based and PLM-based) significantly enhances the model's robustness and accuracy by integrating diverse reasoning pathways and filtering irrelevant evidence.

Summary by alphaXiv.