Research And Implementation Of Muti-documents Reading Comprehension Based On Graph Convolutional Network

With the explosive growth of information,in order to quickly get the results we want from the complex information,multi text reading comprehension technology provides us with a very good solution.Its research goal is to enable the machine to make reasonable answers to specific questions on the basis of understanding the contents of various documents,reduce the time consumption of consulting materials,reading text,summarizing,and improve people's work efficiency.Graph Convolution Network has obvious advantages in obtaining long-distance semantic features in text,and has made remarkable achievements in English data sets.However,there are three challenges in applying it to the task of Chinese Multi-documents Reading Comprehension: First,how to transform documents data into undirected graph,including the selection of nodes and the determination of connection relationship;Second,how to extract semantic features related to the problem from the text;Third,how to obtain the clue information from multiple documents.In view of the above three challenges,this paper constructs a Bi-directional Attention Sentence Graph Convolutional Network(BASG)through in-depth analysis of the task of multidocuments reading comprehension.In BASG model,sentences in text are regarded as nodes in undirected graph,and three connection strategies are proposed to connect all nodes to get undirected graph.In this paper,we use BERT to get the word semantic coding of each sentence firstly,and then use Text CNN to extract the sentence semantic code of each sentence from the word semantic coding information,so as to obtain the feature representation of each node.Then,we use GAT network to update the feature representation of each node to obtain the semantic knowledge of the full text.Finally,Bi-Attention mechanism is used to fuse the problem embedding information and graph embedding information to predict the answer to the question.In addition,in order to obtain the optimal number of output nodes,this paper tests the scoring results of BASG model when the number of output nodes is different,and the number of output nodes in the optimal model is selected as the final number.The model and algorithm designed in this paper are trained and tested on the Du Reader datasets.The test results show that BASG model can achieve better prediction results than Match-LSTM model and Bi DAF model.