Research On Machine Reading Comprehension Based On Multiple Documents

With the development of deep learning,researchers have made further in-depth research on artificial intelligence.Natural language processing is a very important branch in the field of artificial intelligence,and its wide application in various industries has important practical significance.And machine reading comprehension is also an important research direction in the field of natural language processing.The DuReader dataset is a data set proposed by Baidu in 2018.As the largest Chinese data set,DuReader has the characteristics of large data size,authentic data sources,and multiple documents for each question.For this kind of multi-document machine reading and comprehension task,the main research of this article has three aspects: 1.Consider how to extract information that effective for the problem from long text sequences;2.Research how to better semantic encoding of context and questions;3.Research how to make the contextual information and question information better merge.The main research contents of this paper are as follows:First of all,for the problem of too long data in the DuReader dataset,the length of the text is usually beyond the acceptable range of the general machine reading comprehension model,and multiple documents may contain information that is irrelevant to the problem and ultimately affect the model results.This paper proposes a paragraph extraction strategy to extract paragraphs related to the problem as the text input to our training model.First,calculate the degree of relevance between the question and each paragraph,and use the F1 score to indicate high or low relevance.We select the top 5 paragraphs and splice them with the document title.If the selected paragraph length exceeds the preset input length of the model,the paragraph is cut into a text sequence that does not exceed the maximum preset length.Secondly,a multi-document machine reading comprehension model is constructed.The model is divided into six layers,the first layer is the paragraph extraction layer,and the paragraphs related to the problem are extracted by the above method.The second layer is the word embedding layer,which converts words in the text sequence into vector form and input into the model.In this paper,by training the Word2 vec model,the Word2 vec model is used to vectorize the text to obtain a 300-dimensional text vector representation.The third layer is the coding layer,which uses the gated recurrent network Bi-GRU to code the context and the question separately to obtain a vectorized representation of the context and the question.The fourth layer is the interaction layer,which uses two-way attention to calculate the context-to-problem attention and the problem-to-context attention,so as to better integrate the context information and the problem information.The fifth layer is the analysis layer,and Bi-GRU is used again to encode the output of the previous layer to obtain a vector representation of the context of the fusion of the problem information.The sixth layer is the answer prediction layer,which uses the pointer network to predict the start and end positions of the answer.Finally,the feasibility of the multi-document machine reading comprehension model is studied.The DuReader dataset combined with the deep learning framework Tensor Flow is used to conduct experiments.Through the experimental evaluation criteria,the results are compared with the baseline model results provided by the dataset.The results show that the prediction results of the model proposed in this article are relatively better.The baseline model has been improved.