Research On Multiple-choice Machine Reading Comprehension Based On Graph Convolutional Neural Networks

Machine reading comprehension is a crucial task in the field of natural language processing,where the goal is to enable computers to understand and answer relevant questions from human language.The task usually consists of four subtasks: extractive reading comprehension,cloze style question answer,multiple choice question answer,and free-form question and answer.With the constant development of artificial intelligence,machine reading comprehension technology has been widely used in the fields of intelligent customer service,intelligent search,medical diagnosis and so on.However,despite the significant progress,current machine reading comprehension models still have limitations in dealing with semantic logic inference problems.Related studies have been conducted in this regard to improve the effectiveness of the model by modeling passages,questions and candidate options mainly through attention mechanisms.But the attention mechanism suffers from the problem of paying too much attention to some words or ignoring key words,which causes the model to ignore key information in the learning and comprehension process.Here,we focus on the multiple choice task in machine reading comprehension,which requires a machine to answer a question based on the content of a given passage and select the best option from a set of candidates.For the purpose of addressing the shortcomings of existing models in dealing with semantic logic inference problems,our paper will conduct an in-depth study of multiple choice models in terms of optimizing the accuracy of mainstream machine reading comprehension models,the details of which are as follows:(1)Term similarity-aware model for multiple-choice question answer evidence sentence extraction.Multiple-choice question answer aims to automatically select the correct option from the candidates given the passage and the question.However,traditional approaches typically model attention mechanisms based on whole text information or manually mark key sentences for weakly supervised learning,which leads to models that extensively focus on redundant information and generate expensive manual annotations.We consider unsupervised evidence sentence extraction in this paper to precisely identify evidence sentences and minimize the impact of redundant information,while avoiding expensive manual tagging.Specifically,we propose a term similarity-aware evidence sentence extraction model,which dynamically distills key information by term similarity and intelligently selects sentences from passage that are more relevant to the questions as a collection of evidence sentences.(2)A multiple-choice answer prediction model based on graph convolutional neural networks.Nowadays,passage modeling is unsatisfactory when inputting longer texts.To minimize the information lost in the modeling process due to long texts,we model the evidence sentence set and construct an entity graph to represent the evidence sentence set in the passage in the form of a graph.Firstly,we perform the operations of sentence division,word division and word form reduction on the passages with Spacy and NLTK toolkits;then,we match the candidates by comparing sentences containing the same vertices through a series of coding techniques,and aggregate the matched features by an improved graph convolutional neural network to complete the answer prediction of the multiple-choice task.(3)Finally,to evaluate the effectiveness of evidence sentence extraction and entity graph construction,we apply the model to the typical pre-trained language model BERT,which is coded and evaluated on the RACE and Dream dataset benchmarks,with an average of 231.1words per piece for middle school reading comprehension and 353.1 words per piece for high school reading comprehension in the RACE dataset.The proposed method is extensively evaluated on two publicly available datasets,and the results show a significant improvement over various state-of-the-art methods for machine reading comprehension matching.It is verified that the model proposed in the paper achieves substantial performance improvements over the current baseline.