Dynamic Graph Attention Recurrent Neural Networks for Fact Verification

Fact verification is a recently introduced task, which is quired to judge the authenticity of a claim. Previous works mainly leverage extracting the semantic relations of claim and evidences, e.g., using the graph to model their relation. However, these graph-based models ignore the features evolution over different step graph propagation. In addition, they fail to capture the sequence features of evidence graph. To deal with the above issues, we propose a dynamic graph attention recurrent neural network for fact verification. Specifically, we design a dynamic memory mechanism to retain the node features in the updating process. Moreover, we propose a graph attention recurrent neural networks to aggregate features based on the logical order. Experimental results on FEVER dataset demonstrate the effectiveness of our model, especially in the multi-evidence scenario.