Research On Machine Reading Comprehension Technology Based On Deep Learning

Machine reading comprehension is an important and challenging task in the field of natural language processing.It enables machines to understand the semantic information in an article by reading it,and automatically give answers when asked questions related to the article.In traditional machine reading comprehension,artificial features and statistical learning methods are used for modeling and different joint probabilistic mapping functions are used for matching.Limited by data sets and machine hardware,the performance of the model is not satisfactory.Recently,along with the release of high quality large-scale data sets and the development of deep learning technology,machine reading comprehension uses neural network to automatically learn the features of articles and problems,overcoming the limitation of manual feature selection in traditional machine reading comprehension methods.But through analyzing the details of the specific model,we found that there are still some problems to be solved in the state-of-the-art reading comprehension methods,mainly in the following two aspects.First,in the basic model BiDaf,when the word in the answer appears in the question,the accuracy is higher,but when the free translation or synonym replacement is used in the question,the accuracy of the model is significantly reduced.And the attention mechanism alone is not deep enough for the information interaction between the article and the question,and the understanding degree of the model is relatively simple.Aiming at this problem,this paper proposes a multi-level attention reading comprehension model integrating the semantical information.The semantical information enables the model to obtain different semantic information,and reduces the excessive attention of local information and the deviation error of attention.Second,although the widely adopted end-to-end machine reading comprehension method can learn certain text structure information,the reading comprehension of long text content is easily disturbed by irrelevant word information,thus unable to dig out deeper semantic information.Aiming at this problem,this article proposes a method that combines dependency syntactic knowledge into machine reading comprehension,the pretrained word embedding can not only introduce a large number of prior knowledge in reading comprehension model,the accuracy of the reading comprehension model is effectively improved.