Research On Chinese Machine Reading Comprehension Technology Based On CNN

Machine reading comprehension is one of the most cutting-edge and popular research directions in the field of natural language processing.It aims to build a model by using computer,so that computers can read articles,analyze semantics and answer questions like human beings,which has important research value and practical value.From composition automatic scoring to intelligent finance,from intelligent customer service to search engine,machine reading comprehension technology can automate a lot of time-consuming and laborious manual analysis,greatly improving the productivity of society.With the continuous improvement of hardware computing power and explosive growth of big data,and the development of deep learning technology,the research of machine reading comprehension has made great progress.In some specific tasks,the answer of computer model can be comparable to the human level.The machine reading comprehension model needs to fully understand the semantic information of the question and the article.Usually,three layers of hierarchical network architecture are used,namely the coding layer,the interaction layer and the decoding layer.In this thesis,we improve the coding layer,decoding layer and interaction layer of the existing model for the understanding of multi segment extraction Chinese machine reading,so as to improve the accuracy of the model and the efficiency of model training and prediction.The main research work is as follows:(1)In the coding layer and decoding layer of the model,this thesis improves the word vector coding method of baseline model and the decoding method of CRF sequence annotation.This thesis uses the pre-trained word vector adding character vector to fine-tune,this word and character mixed embedding method,which not only contains rich semantic features,but also takes into account the flexible characteristics of the character vector,combined with the decoding method of half-pointer and semitagging.It can effectively reduce the probability of error in answer boundary prediction.The experimental results show that the improved method of the coding layer and the decoding layer increases the F1 value of the model by 0.83%.(2)In the interaction layer of the model,this thesis uses expansion gate convolution instead of LSTM,expansion convolution in the baseline model to expand the receptive field of the network,threshold convolution controls the flow of information to achieve the purpose of multi-channel transmission,and then uses residual network to combine gated convolution and expansion gate convolution,namely expansion gate convolution.Problem features and material features are interactively integrated by attention mechanism.The experimental results show that this improvement can increase the training speed of the model by nearly 13 times and the prediction speed by nearly 28 times when the accuracy reaches or exceeds LSTM.(3)This thesis explores the use of BERT to obtain context coding to solve the ambiguity problem,but the experimental results show that the effect is not very obvious on the dataset that is not clearly labeled.(4)In the process of model training,this thesis introduces the weighted moving average and the latest RAdam optimizer to improve the stability and convergence speed of the model training.