Research Of Machine Reading Comprehension Algorithm Based On Deep Learning

As a new research field of natural language processing and question answering,machine reading comprehension has become a hot topic in academia and industry.The goal is to enable machines to read and correctly understand natural language texts and then answer given questions.The study of machine reading comprehension tasks is very important.On the one hand,reading comprehension can assess the degree of machine understanding of natural language,and can promote the development of natural language understanding technology;On the other hand,reading comprehension is one of the key technologies in question answering system in the future.Benefit from the remarkable improvement of computing power and storage capacity of the computer,as well as the powerful text feature extraction ability of deep learning method itself,research on end-to-end machine reading comprehension model has made great progress in recent years.To fully understand the semantic information of text,machine reading comprehension tasks usually utilize hierarchical network model architecture with the aim to extract information at different levels.The reading comprehension model architecture usually consists of an embedding layer,an encoding layer,an interaction layer and an output layer,while the embedding layer and the interaction layer are the most important parts of the model.In order to enhance the accuracy and efficiency of the model,the thesis aims to improve the embedding layer and interaction layer of the exist reading comprehension model.The research work in the thesis includes:(1)Studying and recurring the representative R-NET model of the machine reading comprehension task,which is used as a benchmark for the subsequent model.The experimental results show that the recurred model can achieve the effect of the thesis.(2)In the embedding layer of the model,the thesis improves the traditional word vector representation method.Traditional word vectors are context-independent,lacking lexical and semantic information.In order to enrich the semantic representation of words,this thesis add the word vector obtained by the pre-trained language model and wordlevel attention vector to the embedding layer.Experiments show that this improved method can enhance the accuracy of the model.(3)In the interaction layer of the model,the thesis uses hierarchical attention and aggregation mechanism to improve context encoding.The traditional model has a simple interaction structure,which leads to weak correlation between context and question,and poor understanding of the model.In order to extract more fine-grained text features,this thesis first aggregate the results of two bidirectional attention models,and then fuse the low-level vectors into the current vector representation.Finally,the self-attention model is used to further enhance the feature representation of the context.Experiments show that the improvement of the model in the interaction layer can enhance the semantic understanding ability of the model.(4)Finally,based on the improved method of embedding layer and interaction layer,we construct the machine reading comprehension model of this thesis,and then compared it with other public models.The experimental results show that the model is more effective than the traditional reading comprehension model.