Reseach On Neural Machine Reading Comprehension

Natural language comprehension is a key ability of human cognition and a prerequisite for acquiring knowledge,and letting machines learn to understand human languages or words is also a difficulty and challenge in the field of artificial intelligence(AI).Machine reading comprehension(MRC)is an important branch in the field of natural language processing(NLP)and a sub-direction of question answering system.The significance lies in making machines understand text semantics and have reasoning ability.Given original text and questions related,the main process of machine reading comprehension is to understand the context semantics and infer the output to match the answer.Compared to the question answering system,it does not rely on an external knowledge base system,but searches for the connection between sentences from a small scope.The traditional NLP technology hopes that the machine can possess the basic ability to capture the meaning,part of speech,and named entities in the text like humans.These tasks don't require complex semantic understanding and reasoning and can be done well through common machine learning or statistical learning methods.Further,we hope that the machine can understand the article or paragraph,according to the context or additional knowledge for analysis and thinking.This requires extracting higher-level,finer-grained text features in more complex scenarios.Neural machine reading comprehension is a new stage in the development of MRC field,that is,the use of neural network models is more suitable for identifying lexical matching and interpretation than traditional feature-based classifiers.End-to-end deep learning models with better performance have become mainstream models recently.All parameters are obtained through optimization without relying on any downstream language features,eliminating the need to construct a large number of artificial features,and have certain anti-interference ability.This paper designs a deep learning-based machine reading comprehension model based on the mainstream reading comprehension framework,which can be used to solve the ”span extraction” type of reading comprehension problems.Aiming at the design flaws of the mainstream model BERT in the ”embedding” module,a random concealment optimization of adjacent word segmentation was proposed.At the same time,due to the differences between Chinese and English languages,in conjunction with Chinese word segmentation tools,the Chinese training sample generation strategy optimization was introduced,which improved the performance of the model under the Chinese dataset.In addition,based on the characteristics of the segment extraction reading comprehension task,this paper modified the training loss function and introduced the boundary-assisted objective function to improve the performance of the model under the segment extraction task.We conducted experiments on typical Chinese and English span extraction datasets.By comparing the test results of the model designed in this paper with the test results of mainstream models such as BERT,R-NET,BiDAF,and Match-LSTM recently,the feasibility of the experimental design model and the effectiveness of optimization was verified.The best results of the experimental model on the Chinese Wikipedia dataset CMRC 2018 reached 70.5(EM)and 87.4(F1),while the typical English span extraction task dataset SQuAD 2.0 reached 84.6(EM)and 87.1(F1).In addition,based on the deep learning machine reading comprehension model,this paper explores the ”black box model” of deep learning models to provide model interpretability.We explain the observation mode of the attention mechanism involved in the model by visualizing the weights,and point out six more obvious attention behavior modes existing in different levels of the model.In addition,we summarize the accuracy of different types of problems from a macro perspective to summarize and analyze the key abstract comprehension capabilities learned by current NLP models.