Research Of Question Answering System Based On Chinese Knowledge Graph

With the advent of the information age,people's desire to acquire knowledge on the Internet is getting higher and higher.The traditional information retrieval method based on search engine will return a large number of web pages related to the question,which not only has high requirements for the accuracy of the ranking results,but also requires manual clicks on the links to filter the information,which will undoubtedly take a certain amount of time.Therefore,the question and answer system came into being.The question and answer system can directly understand the user's question,return a simple and correct answer,and reduce the user's query cost.Knowledge graph is a new type of database,which can be regarded as a huge semantic relation network,representing the relationship between entities in the objective world,and it stores knowledge in a graph structure.Therefore,knowledge has the characteristics of high relevance,good structure,and high accuracy.The question answering system based on the knowledge graph(KBQA)combines the advantages of the above two and has attracted extensive attention in the academia and industry.At present,question answering system based on English knowledge graph(EKBQA)has formed a certain system,which mainly includes key technologies such as question classification,entity linking,relation prediction,and answer generation.However,the research of question answering system based on Chinese knowledge graph(CKBQA)started late,and there are still many problems that need to be solved urgently.In particular,many Chinese words have the characteristics of polysemous words,and the effect of entity linking is often unsatisfactory.Therefore,our method introduced in this thesis combines heuristic rules and sequence tagging model to identify candidate entities to improve the recall rate of candidate entities as much as possible,and design the statistical features of entities and the semantic features of entities to disambiguate the entities to select the subject entities in the question.The experimental results show that the method in this thesis can not only consider the different manifestations of entities,but also ensure the accuracy of the entity linking results.There are a large number of relationships in the knowledge graph,and the training set is limited in size,so there are often unseen relationships during testing.Therefore,our method predicts simple problem relationships based on pre-trained language models,improves the prediction of unseen relationships,Meanwhile,we compare the performance of different pre-trained language models on CKBQA,and analyze the experiment results.Aiming at the complex questions of multi-hop relationship and multi-entity,this thesis proposes a method of query graph generation.We designed query graph features to rank candidate query graphs and continuously iterated to select the query graph with the highest score and converted it into a query to retrieve answers.In the end,these two types of complex questions were effectively solved.Despite the continuous development of academic research based on CKBQA,there is still insufficient work in application.Our method combines the above two aspects of research to construct a knowledge graph for basic information of Dalian University of Technology,and develops the DUTQA,which provides ideas for the design and implementation of Chinese vertical domain KBQA.In terms of the realization of the Chinese open domain KBQA,a CKBQA based on PKUBase is constructed that can answer both fact-type questions and whether-type questions at the same time.In order to make full use of the information in the knowledge graph,CKBQA compares the predicted answer type of the question with the answer entity type to improve the accuracy of the result.This thesis proposes a solution to the question with multiple constraints.Finally,we encapsulate the two systems into a web application,and displays the answer results in a graph structure.