| Machine reading comprehension is an important research direction in the field of natural language processing.The goal of machine reading comprehension is to give machines the same text reading comprehension ability as human beings,that is,to build computer algorithms so that they can understand natural language text and correctly answer questions related to text content.Machine reading comprehension can be divided into cloze filling,multiple choice,fragment extraction and free answering according to its application scenarios.Among them,extraction machine reading comprehension is the foundation of other machine reading comprehension types,so it is one of the important research branches at present.The extraction machine reading comprehension algorithm can identify the answer text related to a given question from the natural language text,and extract the answer to the question,so as to reduce the cost of information acquisition.It is an important basic task in the fields of automatic question answering,information extraction,knowledge engineering and so on,and has high research value.Although machine reading comprehension has made a breakthrough in recent years,there are still two deficiencies in the real complex text with complex relationships,diverse structures and cross-semantics.First of all,most of the existing research data sets of machine reading comprehension are built on simple text segments,that is,long articles need to be divided into multiple semantically irrelevant text segments,which is quite different from the reading comprehension texts involved in practical applications.Secondly,most current researches on machine reading comprehension use language models based on pre-training,such as BERT,to encode the context information of documents and problems.However,most of these models are oriented to short texts with limited input length,and their effects are not great when processing complex texts.The model with better effect,such as XLNet,is complex and large in structure,and has some problems such as heavy computation burden and even infeasible in routine application.Therefore,this paper aims to build an effective extractive machine reading comprehension model for complex text processing based on the existing deep learning model architecture.Firstly,this paper makes an in-depth analysis of the currently commonly used reference data sets of machine reading comprehension,summarizes their composition rules and complexity,and constructs two complex text MRC data sets for their deficiencies:(1)Independently create CTC data set based on Chinese reading comprehension corpus of primary and secondary school students,select real Chinese reading comprehension text of primary and secondary school students,refer to the annotation format of benchmark data set,and use crowdsourcing to mark it;(2)The complex text dataset Merged＿DRCD is constructed from the text in the baseline dataset DRCD.Secondly,the machine reading comprehension model based on convolutional neural network and self-attention mechanism and the machine reading comprehension model based on gated self-matching attention mechanism are studied.The standard evaluation index is used to carry out confirmatory experiments on the benchmark data set to verify the correctness of the model implementation and the applicability of Chinese context.Then experiments are carried out based on self-assembled data sets to explore the challenge of complexity of real complex text to extraction machine reading comprehension.Finally,based on the existing model,two improvements are made to realize the complex text oriented extraction machine reading comprehension.(1)Combining key sentence extraction with answer prediction,complex text is purified by Text Rank and key sentence recognition algorithms based on keyword matching.(2)Combining question classification with answer prediction,more information is provided to the model by predicting the categories of questions and articles,so as to improve the overall prediction accuracy of the model.The experimental results show that both improvements can effectively improve the accuracy of the machine reading comprehension model of complex text,which shows the effectiveness of purifying complex text and considering problem categories.There are two main innovations in this paper:(1)Aiming at the noise problems existing in real complex texts,a text purification algorithm based on key sentence extraction is proposed to improve the accuracy of machine reading comprehension model of complex texts.(2)In view of the large difference in accuracy of machine reading comprehension model in different types of questions,it is proposed to add problem classification module to the model to increase the encoding of problem type information,which effectively improves the accuracy of machine reading comprehension model of complex text.In this paper,we first investigate the relevant methods,models and data sets of extractive machine reading comprehension tasks,and then study the mainstream models and technologies of extractive machine reading comprehension and reproduce them.Two Chinese complex text machine reading comprehension data sets are constructed with reference to mainstream reference data sets.In view of the shortcomings of existing methods in dealing with complex text,two aspects of context purification and problem type coding are improved.The experimental results show that the key sentence extraction algorithm of text purification and question type coding can effectively improve the accuracy of complex text-oriented extraction machine reading comprehension model on the self-built data set and the benchmark data set. |
