基于改进图节点的图神经网络多跳阅读理解研究  

Research on Multi-Hop Reading Comprehension Based on Graph Neural Network with Improved Graph Nodes

作　　者：舒冲 欧阳智 杜逆索 何庆[2] 魏琴[2] SHU Chong;OUYANG Zhi;DU Nisuo;HE Qing;WEI Qin(College of Computer Science and Technology,Guizhou University,Guiyang 550025,China;Guizhou Big Data Academy,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学计算机科学与技术学院,贵阳550025 [2]贵州大学贵州省大数据产业发展应用研究院,贵阳550025

出　　处：《计算机工程》2022年第1期99-104,共6页Computer Engineering

基　　金：国家重点研发计划(2018YFB1004300);贵州省科学技术厅重大科技计划(黔科合重大专项字[2018]3002);贵州大学培育项目(黔科合平台人才[2017]5788)。

摘　　要：多跳阅读理解需要基于问题并在多个支撑文档中寻找相关信息进行跳跃式推理来回答问题。针对当前多跳阅读理解模型中所存在的实体图内缺乏关键问题信息以及信息冗余问题,提出一种基于改进图节点的图神经网络多跳阅读理解模型。采用基于指代词的实体提取方法提取实体,将提取到的实体基于问题关联实体构建实体图。对实体图中的节点进行编码预处理,通过门机制的图卷积网络模拟得到推理信息,计算推理信息与问题信息的双向注意力并进行结果预测。在WikiHop数据集上的实验结果表明,该模型在测试集上取得了73.1%的预测准确率,相比基于图神经网络、循环神经网络和注意力机制的多跳阅读理解模型准确率更高、泛化性能更强。Multi-hop reading comprehension requires searching for question-associated information in supporting documents to perform leaping reasoning to answer the question.The entity graphs of existing multi-hop reading comprehension models lack key information relevant to the question,but contain redundant information.To address this problem,we propose a multi-hop reading comprehension model based on a graph neural network with improved graph nodes.We employ a demonstrative pronoun-based method to extract entities,and use the extracted entities to build an entity graph based on the question-associated entities.After the nodes in the entity graph are preprocessed by encoding,we use the gated Graph Convolutional Network(GCN)to obtain the inference information through simulation.The bidirectional attention between reasoning and question information is calculated,and on this basis the answer is predicted.The experimental results on the WikiHop dataset show that the model achieves a prediction accuracy of 73.1%on the test set.Compared with the multi-hop reading comprehension model based on GCN,Recurrent Neural Network(RNN)and attention mechanism,the proposed model displays higher accuracy and stronger generalization performance.

关 键 词：多跳阅读理解 实体图 问题关联实体 图卷积网络 双向注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]