Research On Multiple Choice Machine Reading Comprehension

The Machine Reading Comprehension(MRC)is an important task to evaluate the ability of the machine to understand natural language.And MRC is an important indicator of the level of artificial intelligence.The multiple choice MRC task lets the machine select the correct answer from the candidate options after inputting the documents,the question.These candidate options are usually not the text fragment from the document but are usually paraphrase or summaries of the document content,or even the information obtained by reasoning the document or using external commonsense knowledge.Therefore,the multiple choice MRC task has a high requirement of semantic understanding.This paper studies the multiple choice MRC task from three aspects,i.e.,semantic matching,semantic reasoning,and external knowledge assistance.We adopt a robust multiple choice MRC model Co-Matching as the base model.The Co-Matching model concurrently matches the document with both the question and the candidate option at the interaction layer.However,the Co-Matching model has two shortcomings.Firstly,the semantic matching method is not comprehensive enough.The Co-Matching model does not take into account the fact that the question and options are usually complementary.Secondly,the model lacks reasoning.Therefore,the first work of the paper adds the complementary information of the questions and the candidate option,and then match it with the document.The experiment results show that this improvement can enhance the semantic matching ability of the model.Usually,the options of multiple choice MRC are not the text fragment from the document,and the model requires an in-depth understanding of the text meaning to reasoning.Therefore,to address the second shortcoming of the Co-Matching model,the second work of this paper is to propose and adds two kinds of multi-step reasoning based on the model of the first work.In order to make eachreasoning step depending on the integration results of several early reasoning rather than only the latest reasoning result,we add skip connections to the multi-step reasoning.The experiment results show that adding multi-step reasoning and skip connections in the model can improve model accuracy.The first two works assume that the answer only comes from the document,so the model performance is limited when answering questions that require external knowledge assistance.The third work of this paper adds external knowledge to assist the model.Through the sentence-level external knowledge integration method,the information of external knowledge,questions,articles,and candidates is used together for semantic matching and semantic reasoning.The experimental results show that using the sentence-level external knowledge to assist semantic matching and reasoning can improve the accuracy of the questions that require external knowledge reasoning.Through an in-depth analysis of the characteristics of the multiple choice MRC task,this paper effectively improves the overall performance of the multiple choice MRC model from the aspects of semantic matching,semantic inference,and external knowledge assistance.