Multiple-Choice Question Answering For Chinese Reading Comprehension

Machine reading comprehension is the current research hotspots of artificial intelligence,which involves many core technologies such as natural language processing and machine learning.It has been widely used in chatting robot and question answering system.Its task is to teach machine to understand a given passage and then to answer the question related to it,which pays more attention to semantic analysis between question and background material.Focusing on micro-reading mode automatic answering technologies,this paper studied the multiple-choice question for Chinese literature reading comprehension in Beijing College Entrance Examination.First,the characteristics of the multiple-choice question was analyzed,and the automatic answering process was decomposed into two parts: candidate sentence selection and textual entailment recognition.Secondly,a variety of candidate sentence selection methods and a multiple-to-one textual entailment model based on hierarchical neural network was proposed.Finally,the model was evaluated on the Beijing College Entrance Examination papers in the past ten years.The specific works was as follows:(1)Multiple-choice question and option analysisThis paper analyzed the characteristics of the multiple-choice question for literature reading comprehension in college entrance examination and its solution.After a comprehensive and detailed analysis of the multiple-choice questions,the types of multiple-choice questions,and answering techniques,the automatic answering process can be decomposed into candidate sentences selection and textual entailment.(2)Option candidate sentence selectionThis paper considered the option candidate sentence selection as a semantic relevance method and proposed several sentence similarity methods,namely sentence similarity based on word embedding,topical distribution similarity based on LDA,sentence similarity based on topical wordembedding,and word-overlapping.The experiments on Gaokao Multiple-Choice Dataset show that the method based on word embedding obtained best,and F-measure reaches 40.2%.(3)Textual entailment recognitionAiming at the characteristics of multiple-to-one textual entailment in reading comprehension,this paper proposes a multiple-to-one textual entailment model based on hierarchical neural network,which can effectively integrate the semantic information between multiple sentences.Option segments often correspond to more than one sentence in the background material.First,option is divide into clauses by rules.Then a hierarchical neural network was used to learn semantic information for option clause sets and candidate sentence sets.Finally,we used those two sets' semantic information to classify,which is combined by several heuristics including concatenation,element-wise product and difference.Our method gained a accuracy of 58.92% on Beijing College entrance examination reading comprehension entailment data set.(4)Multiple-choice question solution SystemCandidate sentence selection and textual entailment was used to construct a multiple-choice questions reading comprehension solution system.This paper also exploit several automatic answering strategies for the system.The experiments showed that the best strategy has achieved a accuracy rate of 57.14% on real Beijing College Entrance Examination papers in the past ten years.