基于实体图神经网络的事实核实方法

Fact Verification Method Based on Entity Graph Neural Network

针对信息真伪识别问题,研究了基于实体图神经网络的事实核实方法。首先,对通过事实句子提取的实体进行文档检索,利用加强长短期记忆网络进行证据筛选得到相应证据集;然后,利用实体识别获取证据中的相关实体,并通过构建的实体图和图注意力神经网络机制实现实体信息的传播更新;最后,融合实体信息和证据句子信息进行标签预测。在事实提取与验证(FEVER)数据集上的试验结果表明,与3种基准模型相比,该方法有效提升了标签预测的准确性,并在实体信息充分的验证集上表现更佳,其模型在推理层数为3时取得最佳效果。该方法既可提取关键实体,又可捕获实体间信息关联,为提升信息真伪识别技术提供参考。

Aiming at the problem of information identification,a fact verification method based on an entity graph neural network is proposed.Firstly,the entities extracted from the fact sentences are used for document retrieval,and the corresponding evidence set is obtained by the enhanced long-short term memory network.Then,the related entities in the evidence are obtained by the entity recogni⁃tion,and the entity information is transmitted and updated by the constructed entity graph and graph attention neural network mechanism.Finally,the entity information and the evidence sentence infor⁃mation are integrated to predict the label.Experimental results on the fact extraction and VERification(FEVER)data set show that the proposal can effectively improve the accuracy of the label prediction compared with the benchmark model,and has better performance on the verification set with sufficient entity information,and the model achieves the best effect with the number of reasoning layers as 3.The method can accurately extract the key entities as well as their relationships,providing a solution for improving the information authenticity recognition technology.