Evidence Decomposition Graph Network for Fact Verification

Anonymous

Abstract: Fact verification is the task to verify a given claim according to extracted evidence sentences. Most existing works use whole evidence sentences or break them into phrases to perform evidence interaction, where evidence is treated either too coarsely or over fragmented. We also find that many models suffer from exposure bias, which finally leads to them only paying attention to the evidence ranked higher by previous steps while failing to recognize crucial pieces from all candidates. In this paper, we propose an Evidence Decomposition Graph Network (EDGN), which decomposes each evidence sentence, especially the complex ones, into several simple sentences, highlighting the required key information without losing sentence structure and meaning. EDGN also absorbs a simple but effective evidence shuffling method to mitigate exposure bias. Experiments on the FEVER benchmark show our model can take all evidence candidates into account, distill necessary key information from complex evidence, and outperform existing methods in the literature. We will release our code to the community for further exploration.