Research On Multiple-Choice Question Of Prose Reading Comprehension

Today,research on reading comprehension has become a hot topic in question-answering system research.Different with traditional question-answering system that retrieves answers in large-scale document sets,reading comprehension focuses more on deep understanding of a single document,which needs more natural language processing other than information retrieval technology.Most of the current research on reading comprehension is specific to choice question,especially to the simple factual question with entity-oriented short answers.However,in the real college entrance examination,reading materials and questions are very complex.For example,prose is a common material in Matriculation Chinese Test of Beijing(BJMCT)that few traditional NLP researches on.And question type includes both conventional questions and choice questions,which is flexible and complicated.This paper researches the solution to choice question in prose reading comprehension of BJMCT.This paper is mainly divided into four parts:(1)Analysis on choice question of prose reading comprehension from three aspects including forms,contents and difficulties.(2)Answering strategies based on language model.This paper uses one-gram language model and bi-gram language model to compute article's support to every options,and then sorting options based on the results to answer the question.The recall of the solution is 32%.(3)Answering strategies based on heuristic information.The main idea of this solution includes:(1)dividing options into text statement clause and idea-understanding clause based on option division information;(2)primary exclusion of options based on sentiment polarity;(3)using abnormal information to compute idea-understanding clause and then sorting options based on the results.The recall of the solution is 60%.(4)SVM-based multi-label classification of options.This paper studies with two methods of class definition and automatic classification:(1)dividing options into five classifications of conclusion type,cause-effect type,relationship inference type,emotion type and language appreciation type;(2)representing options by bag-of-words model with vocabulary,semantics and grammar and word2 vector additive model trained by neural network,which uses SVM-based Binary Relevance algorithm and SVM-based Classifier Chains algorithm to make multi-label classification.The complete matching of label is 40%.This paper's contribution is as follows:(1)It presents answering strategies based on language model and heuristic information.(2)It proposes classification task of option,defines an option category and tries SVM-based multi-label classification.