Research On Fact Verification Based On Fine-grained Reasoning

At present,people can use various communication media and platforms to publish and search for information on the Internet.Although this reflects the openness of information,it also leads to the wide spread of a large amount of information on the Internet,in which there are many rumors whose sources are difficult to determine.The continuous spread of rumors has caused many negative effects on social credibility.In this regard,the fact verification task aims to provide an automated verification system that judges a given claim,and gives the judgment results of "SUPPORT","REFUTE",and "NOT ENOUGH INFO",thereby verifying the credibility of Internet information.In order to solve the task of fact verification,the model should have good information retrieval and sentence selection capabilities,so as to provide sufficient data input for subsequent fact determination.The fact judgment model should also be able to carry out sufficient information transmission and reasoning in the sentences,and adopt fine-grained strategies to improve the effect of the model.In addition,the information needs to be identified in this period,so that the reasoning process pays more attention to the key information and ignores the noise information.Existing models often focus on only part of the above requirements,resulting in sub-optimal effects and a lack of generalization capabilities.This paper proposes a hierarchical graph attention model HGAR with key-value(k-v)level reasoning ability.HGAR uses a fully connected graph structure for information transfer and reasoning between sentences,introduces a graph attention mechanism to distinguish key information and noise information during each information transfer process,and further improves model performance through a k-v level fine-grained reasoning layer.HGAR also adds a pooling mechanism to each layer,which can not only reduce model parameters without affecting model performance,but further filter out information noise from the graph.This article has carried out sufficient experiments on the FEVER data set,and the experimental results have verified the effectiveness of HGAR.In addition,this paper also validates the effect of each module of the model in solving the task of fact verification through a well-designed ablation experiment.