Research On Machine Reading Comprehension And Its Application In Question Answering System In The Field Of "Four Insurances And One Housing Fund"

Machine reading comprehension is to understand the semantics of articles through artificial intelligence technology.It is an important technology to construct a question answering system.Current machine reading comprehension models usually obtain answers according to given questions and paragraphs containing answers,However,in the field of "Insurance and Housing Fund ",there are few questions and matching documents pairs can be used,it is such a lack of data."Insurance and Housing Fund" is a very important social security system in China.The majority of insured people have an urgent need for knowledge acquisition in this field.Building a question answering system through machine learning technology in this field of can greatly facilitate the insured people.At present,there is no question answering system with the text of "Insurance and Housing Fund" policies and regulations as the document library.The main work of this paper is as follows:According to the characteristics of the text structure of policies and regulations,this paper puts forward the method of constructing policies and regulations metadata,and constructs the metadata of policies and regulations with the rule-based machine reading and understanding method.A machine reading comprehension model is constructed,which only needs documents for input,including question generation module and answer extraction module.Through the question generation module and using metadata,457988 question are generated based on policy,regulation documents to form question-answer pairs.The answer extraction data set in the field of "Insurance and Housing Fund" is constructed,and a good answer extraction algorithm is trained.Taking the constructed question-answer pairs and the text paragraphs of policies and regulations as the knowledge base,this paper constructs the question and answer system in the field of "Insurance and Housing Fund".When getting the answer,question-answer pairs are matched first,and return the answer if matching is successful.Otherwise,the text paragraphs are sorted by information retrieval and policy and regulation metadata,and then the final answers are returned by the answer extraction algorithm.