Chinese Machine Reading Comprehension Based On Multi-hop Attention

For the past few years,As artificial intelligence's develop,the task of machine reading comprehension under natural language processing has become a hot topic of research.The emergence of a large number of Chinese data sets has contributed to a research climax of Chinese machine reading comprehension tasks.The core content of machine reading comprehension task research is text content reasoning,and most of the research work is based on a single-layer network model structure,which is characterized by first using bidirectional LSTM or bidirectional GRU to obtain word vector representation of the document and word vector representation of the query,and then use the obtained word vector representation to interact with the document and question information,and finally use the attention mechanism to predict the correct answer.Although the effect of this model structure is relatively significant,due to the single-layer network model,the text content reasoning will not be deep enough,to some extent,it will cause part of the text word information to be missing,which will have a certain impact on the effect of the machine doing the cloze task.In view of the above situation,after studying the Chinese machine reading comprehension model based on deep learning,this paper proposes a novel machine reading comprehension model based on multi-hop attention.In some related works,the effect of deep learning mode on text content reasoning is significantly better than traditional learning mode.There are two main types of deep learning technologies currently studied:the first is Multi-hop Architectures)Technology,which allows the model to perform multiple iterative calculations on documents and questions to achieve the purpose of deep reasoning of text content,and the effectiveness of today's multi-hop reasoning technology has been verified in relevant task models;the second is Attention Mechanisms(Attention Mechanisms)technology,this mechanism allows the model to focus on the part of the document related to the problem,according to the different parts of the document and the relevance of the query to re-weight.Based on these technologies,this paper adopts a multi-layer network model to handle the cloze task of Chinese machine reading comprehension,and integrates multi-hop architecture and attention mechanism,and proposes a Chinese machine reading comprehension model based on multi-hop attention.Through the design of the multi-hop architecture,the model better integrates the query information into the semantic information of the article words to achieve the semantic reasoning of the article.In the final answer prediction stage,the model will update t word vector representation of the document and word vector representation of the query after the update Dot product calculation,the obtained results form a two-dimensional matrix and the matrix is operated to better filter the final correct answer,so as to better complete the machine reading comprehension task.This paper has conducted experiments on the first Chinese fill-in-the-blank machine reading comprehension data sets PD&CFT(People Daily,Children's Fairy Tale Datasets)proposed by IFLYTEK Joint Laboratory of Harbin University of Technology.The accuracy of the model structure used in this paper is 65.8%and 68.5%in the validation set and test set of PD data set,and 43.2%and 35.2%in the CFT data set.The corresponding comparative experiments show that the accuracy of answer prediction will be improved when using the Chinese machine reading comprehension model based on multi-hop attention proposed in this paper to do machine reading comprehension cloze tasks.