Multi-turn Question Answering System Based On Knowledge Graph

The rapid development of artificial intelligence makes the question-answering system get further development,and the application of question-answering system based on artificial intelligence is also everywhere.With the appearance of knowledge graph,It can be used for storage large-scale data.It gives a new possibility to the Q&A system,and retrieval based Q&A has become the mainstream application of Q&A based on knowledge graph.The research focus of this thesis is to apply knowledge graph to multiple turns of interaction and realize multiple turns of question and answering.At present,the single turn retrieval question and answering based on knowledge graph has been relatively mature,but there is less exploration for multiple turns of question and answering.Because users can not give all keywords at once in the process of obtaining information,in order to provide users with more friendly help and user-friendly guidance,so as to accurately locate user needs and return to the real needs of users.Therefore,it is necessary to have multiple rounds of question and answering.The whole system includes the construction of the construction of knowledge graph,the design of multi-turn question answering model and the realization of the system.First of all,this thesis extracts entities from multi-source heterogeneous data,in which the web data obtained by crawlers is extracted by regular method,while the free text is extracted by knowledge sample tree.Then,the extracted entities are transformed into triples and stored in the neo4 j graph database which is dynamically built based on docker-compose.Next,through the construction of human-computer interaction historical record storage and transmission function,the historical communication content is recorded,which is used to analyze the current round of dialogue and generate the reply based on slot model or random walk model.Finally,the knowledge graph construction and multiple turns of question and answering model are combined to a formed program to realize the construction of robot by alerting the configuration,and display it as a web page.The multi-turn question-answering system proposed in this thesis gives a complete implementation scheme,but there is still something to be optimized.Among them,the map building link can add synonyms to the entity alignment,and multiple turns of Q&A model also need to add word2 vec,so as to better realize the function of similar problem identification.