RERG: Reinforced evidence reasoning with graph neural network for table-based fact verification

Table-based fact verification aims to check whether a statement is entailed by the content of relevant table. Existing works mainly either parse a statement with logical form or design a table-aware neural network to represent the statement-table pair. However, they fail to directly exploit guidance signals to capture enough evidence from the table, which lead to performance degradation. Thus, to investigate how to select potential key words from the table for fact verification, we propose a Reinforced Evidence Reasoning framework with Graph neural network (RERG), which simulates human inference process of focusing on some words at each step. Specifically, we employ a Transformer-based graph neural network to represent multi-granularity features. Then, we design a monitor node and connect it with some potential key nodes by reinforcement learning on each graph layer, according to the feedback of the reward. In this way, the monitor node can be used to predict the label, which has aggregated various key information through multiple graph layers. Besides, we add secondary updating after the attention mechanism to enhance information aggregation of each graph layer. Experimental results on two benchmark datasets TABFACT and INFOTABS show performance improvements over state-of-the-art baselines and the feasibility of selecting some meaningful evidences during graph reasoning.