Research On Machine Reading Comprehension Method Based On Deep Learning

With the development of science and technology,the application of artificial intelligence technology in life is more and more extensive,among which the difficulty is how to make the machine can correctly understand human language.The goal of the machine reading comprehension task is to teach the machine to do reading comprehension,so that the computer has the same ability to understand the text as human beings.Existing machine reading comprehension methods usually use the attention mechanism to get the corresponding vector representation between the passage,the question and the candidate answer.However,the single attention mechanism still has some shortcomings.It may pay too much attention to some words,or neglect some words that will be useful.This will cause it to be unable to make effective use of the key information in the information interact of articles,questions and candidate answers.To solve above problems,this paper proposes a machine reading comprehension model based on multi-perspective co-matching mechanism,used in the training of language model BERT term vectors,and then in the article,questions and candidate answers interaction module uses a variety of strategies for interaction information from multiple perspective,then through the aggregation of the document class vector,according to the last layer in the forecast to predict the answer.Next,the main research content of this paper will be introduced:(1)This paper proposes a machine reading comprehension model(MSCM),which is based on the common matching of multiple matching strategies.In the information interaction layer,the model uses four different matching strategies to match the information among articles,questions and candidate answers.These four matching strategies are: 1.Full matching strategy;2.Maximum pooling matching strategy;3.Attention matching strategy;4.Max-attention matching strategy.The four different matching strategies avoid the excessive or insufficient attention on some words by the single attention calculation method,and can better match the article,question and candidate answer,so as to extract more accurate information.Experiments on RACE data sets show that the accuracy of the model is significantly improved compared to the baseline model.(2)This paper proposes a machine reading comprehension model(MPCM)based on multi-perspective common matching mechanism.This model is no longer directly use articles,the attention mechanism to direct questions and candidate answers between the key information,but design a matching function,by four different matching strategy from the perspective of multiple articles and articles and candidate answers to the questions and similarity matching,using the result of the match and the article information multiplication to complete to get related information.At the same time,the model is combined with the pre-training language model BERT to verify the applicability of the model on the higher-order basic model.Experimental results on RACE data sets show that the accuracy of this model is much better than that of other baseline models.