Research On Insurance Q&A Matching Model Based On Knowledge Graph And Question Analysis

With the widespread rise of digital technology in traditional industries,online network services have penetrated into people’s lives.Some online insurance platforms have opened insurance knowledge section and question and answer section,where insurance customers can browse the contents of insurance products.However,due to the lack of insurance professional knowledge,it is very difficult for them to choose suitable insurance products.At present,the online insurance platform mostly adopts the way of manual reply,which has some problems such as delayed reply and untimely feedback.The online insurance platform needs intelligent customer service answer system to analyze the intention of insurance customers’ questions and match the corresponding answer text for customers,so as to reduce the time of replying customers’ questions,improve the efficiency and quality of online insurance services,and promote the further development of online insurance services.In addition,for similar semantic questions frequently asked by customers,the automatic question and answer matching method can effectively avoid the waste of customer service resources.Therefore,based on the insurance product data and manual question-and-answer data in the online insurance platforam,this paper constructs the insurance knowledge map in a top-down way,and constructs the intention classification model and entity recognition model to analyze the questions of insurance customers.On this basis,the candidate answer set is constructed,and finally,the BERT-attention-CNN-KG insurance question and answer matching model is designed,which can accurately match answers for questions consulted by insurance customers in time.The research content of this paper mainly includes the following aspects:(1)Construct knowledge map based on insurance data,obtain all kinds of structured or semistructured text data from online insurance website using crawler technology,then conduct ontology construction to determine the categories of entities,relationships and attributes,then construct BERT-biLSTM-CRF model for entity identification and relationship extraction,and use Neo4j for knowledge storage;(2)Analyze insurance questions based on deep learning algorithm,and divide the question parsing task into two parts:Firstly,the named entity recognition model(NER)of biLSTM-CRF and BERT-biLSTM-CRF was constructed,and the insurance entities in question were accurately extracted.The identification accuracy of BERT-biLSTM-CRF was 0.74.By comparing text-CNN,bi-LSTM and bi-GRU classification models,the intention types of questions were classified by bi-GRU to obtain 86%of the best discriminant effect.Then,entity recognition and intention classification models were used to construct the candidate answer sets;(3)Design an insurance-oriented question-and-answer matching model.Firstly,according to the intent category and implication entity of insurance questions,obtain some question-and-answer pairs from FAQ database as candidate question-and-answer sets,and then sort the answers of the insurance questions and candidate question-and-answer sets through the BERT-attention-CNN-KG question-and-answer matching model constructed in this chapter.The most relevant top-k answers will be pushed to insurance users as recommended answers,so as to provide targeted answers to consultation texts of insurance customers and help them solve questions in the insurance field.