Research On The Solution Of Chinese Cue Questions In High School Entrance Examination By Machine Reading Comprehension

Machine Reading Comprehension has always been a focus of research in the field of natural language processing.The goal of MRC is to make the computer have the ability to understand articles as well as human beings through the comprehensive use of natural language processing and comprehension technologies such as text representation,comprehension and reasoning.Since the articles and questions are in the form of human language,which involves the understanding ability of morphology,syntax,grammar and semantics,this task becomes an important evaluation to measure the machine’s understanding ability of natural language.In order to depict the deeper comprehension ability of the machine,this paper studies the quizzed task of reading comprehension cues in the college entrance examination.This paper collects the Chinese reading comprehension questions of the college entrance examination in recent 15 years,and marks the answer candidate sentences manually.Based on the data set,this paper studies cueing questions and proposes an appropriate problem-solving model.The main work of this paper includes:(1)Data set construction.In this paper,real questions and simulation questions of Chinese reading comprehension in the college entrance examination in the past 15 years were collected,and the cue-type questions were analyzed and sorted out to form the cuetype question data set.The cue-type question data set was manually marked with answer candidate sentences for the formed cue-type question data set.(2)Machine reading comprehension model based on event representation.Firstly,the text event graph is extracted from the reading material by the cue phrase.Then,the Text Rank algorithm is used to select the cue-related events by considering the time elements,emotional elements and the importance of each word in the document comprehensively.Finally,the answer to the question is constructed according to the selected clues.Experimental results show that the proposed method can effectively solve cueing questions.(3)Extractive machine reading comprehension model based on pre-training language model.Firstly,paragraph topic similarity is used to combine article paragraphs,and then emotion and time characteristics are incorporated into the pre-training language model.Finally,the semantic similarity of paragraphs and questions is used to select top-6paragraphs as candidate paragraphs,and the answer candidate sentences of candidate paragraphs are combined as the final answer sentences.Experimental results show that the model proposed in this paper can effectively solve cueing questions and improve the performance of the model.(4)Generative machine reading comprehension model based on pre-training language model.First,the paragraph retrieval module is used to get the top-k paragraph with high relevance to the question.Then,the fragment extraction module is used to extract the answer candidate fragments in the paragraph and get the fragment set.Finally,text generation module is used to get the final generated answer.The experimental results show that the performance of the generated model is improved compared with that of the extracted model.