Research On Multi-hop Reading Comprehension Based On Relational Graph Neural Networ

Machine reading comprehension(MRC)is a task designed to test the machine’s understanding of natural language by asking the machine to answer questions according to a given supporting document.In recent years,with the development of various deep learning technologies and the proposal of reading comprehension data sets,MRC has attracted more and more attention and deployed on different platforms.For example,on the customer service platform,the customer service robot can accurately retrieve document information through MRC technology and answer the questions raised by users.In the above practical application,the answers to a large number of questions can not be inferred from a single document but must be answered after the integration of multiple documents.Therefore,researchers turn their attention to the multi-document and multi-hop reasoning reading comprehension.Multi-hop machine reading comprehension(MRC)needs to reason many times.To solve the problem of multi-hop machine reading comprehension,we need to collect multiple pieces of evidence in documents,and then use the evidence to reason and to confirm the answer.However,it is very difficult for the machine to extract the logical reasoning information.In the existing research,some scholars apply graph neural networks to multi-hop reading comprehension.The multi-hop reading comprehension model based on graph neural network transforms the logical reasoning information in the document into the association information between nodes and then reasons.However,the existing models still have the problem of insufficient interaction information extracting between nodes.Aiming at the problems,a multi-hop reading comprehension model based on an improved relational graph neural network is proposed.The main work is as follows:(1)A document selection model based on comparative sorting is proposed.In previous studies,document selection is mostly based on the text classification model of BERT,selecting the documents with high relation scores and failing to consider the degree of relationship between documents and questions or answers.Our model considers this degree of relationship,and builds a comparison matrix based on comparison ranking,and proposes a comparison ranking document selection algorithm based on the construction of a comparison matrix.Experiments are carried out on Hotpot QA dataset to verify that this model can effectively improve the performance of the model.(2)A reading comprehension model based on a multi-channel relational graph convolution network is proposed.After selecting key documents by the document selection model based on comparative ranking model,next step is multi-hop reasoning.Aiming at the problem that the edges connecting nodes are redundant,this model proposes the construction of node graph based on document and query.The multi-hop reading comprehension models based on graph neural network mostly adopt the method of graph convolution neural network or graph attention network.Because of the insufficient interactive information between nodes,a reading comprehension model based on a multi-channel relational graph convolution network is proposed.Then,the model experiment on the Hotpot QA dataset shows the effectiveness of the model.