Research And Implementation Of Knowledge Reasoning Algorithm For QA System

Knowledge graph has attracted the attention of many researchers due to its highly structured knowledge and wide range of application.As a typical application of knowl-edge graph,Knowledge Base Question Answering still has the problem that it cannot rep-resent and utilize knowledge effectively.To solve this problem,this thesis studies knowl-edge representation and question-and-answer reasoning based on knowledge graph,and implements knowledge embedding model based on entity and relation descriptions and question-and-answer model based on fact memory and knowledge graph.Among them,the knowledge embedding model based on entity and relation descriptions is used to im-prove the vector representation of knowledge,and the question-answering model based on fact memory and knowledge graph is used to deduce the answers to questions on the knowledge graph.Specifically,the main research work of this thesis is as follows:(1)Knowledge embedding model based on entity and relation descriptions is pro-posed and implemented.Most of the current knowledge embedding models do not con-sider the semantic information of triples enough,so this thesis proposes a knowledge em-bedding model based on entity and relation descriptions for this problem.In this model,not only the traditional DKRL model and TransD model are improved,but also the en-tity description embedding model combined with the attention mechanism and convolu-tional neural network and the hierarchical relation semantic embedding model are adopted when acquiring entity and relation semantics.Compared with the baseline models such as TransE,TransD and DKRL,the average ranking of this model reaches the optimal level when Bern sampling is used in link prediction,and hits@10 also reaches the optimal level of 77.4%.When Bern sampling is used in the triplet classification,there are 4 more thou-sand points than the previous optimal model(TransD model).(2)A question-answer model based on fact memory and knowledge graph is pro-posed and implemented.In the existing Knowledge Base Question Answering technolo-gies,most models cannot have high accuracy and good interpretability at the same time.Aiming at this situation,this thesis proposes a question-answer model based on fact mem-ory and knowledge graph.In this model,bidirectional triplet information,multiple scoring indexes and extended fact list are adopted to maintain the interpretability of the original fact memory network model,and at the same time,high accuracy of question answering is obtained.Compared with the two baseline models,such as the fact memory network,this model improves 1 thousandpoint in WebQuestions and 1.1 percentage points in Sim-pleQuestions.(3)A question answering system based on knowledge graph reasoning algorithm is proposed and constructed.The question answering system consists of three functional modules: question processing module,data retrieval module and question answering mod-ule,and the new model in(1)and(2)is successfully applied in the core question answering module.In practice,the question-answering system has the advantages of good visualiza-tion,high accuracy and strong interpretability.