Research On Machine Reading Comprehension Based On Deep Learning Neural Network

Machine reading comprehension is one of the most important directions in the field of natural language processing,which has also spawned a large number of neural network models for machine reading comprehension tasks.However,most models use LSTM or GRU based structures.Although this structure works well,due to the nature of RNN,this structure is time consuming,and most models are used in the application of coarse-grained attention mechanism.This will have a certain impact on the results of the machine to answer questions.Therefore,this paper proposes a new fine-grained model structure.This structure uses the SRU that is comparable in speed to CNN as the main structure of the network and uses a fine-grained attention mechanism to align the representation of the passage and the question,it can better combine the information of the passage and the question to infer the answer.In addition this paper also proposes a new network architecture,called Doubled Simple Recurrent Units(DSRU),which is a two-layer SRU variant,in which the second layer SRU uses contextual representation to calculate the current hidden state.After the input layer,the structure of the paper is designed with a variety of alignment structures,including passage align question layer,question align passage layer,and self-alignment layer.These alignment structures are designed according to the idea of attention mechanism.This paper conducted experiments on Stanford Question and Answering dataset.Finally,on the SQuAD 1,1 version of the dataset,the proposed model structure achieved an F1 score of 85.1 on the cross-validation set,and on the later upgraded dataset SQuAD 2.0.The single model structure of this paper achieved an F1 score of 65.95 on its cross-validation set.This score has surpassed the two baseline models given by Stanford officially,but it is still lower than the best-performing baseline model.So it still has more improvement space.