CNGT: Co-attention Networks with Graph Transformer for Fact Verification

Jing Yuan, Chen Chen, Chunyan Hou, Xiaojie Yuan

Abstract

Fact verification is a challenging task that requires retrieving evidence from a corpus and verifying claims. This paper proposes Co-attention Networks with Graph Transformer (CNGT), a novel end-to-end reasoning framework for fact verification. CNGT constructs an evidence graph given a claim and retrieved evidence, uses a graph transformer to capture semantic interactions among the claim and evidence, and learns global node representations of the evidence graph via self-attention mechanisms and block networks. Deep co-attention networks integrate and reason on the evidence and claim simultaneously. Experiments on FEVER, a public large-scale benchmark dataset, demonstrate that CNGT achieves a 72.84% FEVER score and a 76.93% label accuracy score, outperforming state-of-the-art baselines. CNGT has de-noising and integrated reasoning abilities and case studies show that it can explain reasoning at the evidence level.
BibTeX
@inproceedings{yuan2023cngt,
  author = {Jing Yuan and Chen Chen and Chunyan Hou and Xiaojie Yuan},
  title = {{CNGT: Co-attention Networks with Graph Transformer for Fact Verification}},
  booktitle = {ADMA},
  year = {2023},
  doi = {10.1007/978-3-031-46664-9\_39},
}

Loading related content...

Figure