Research On Knowledge Base Question Answering Based On Information Retrieving

Knowledge graph Question Anwering is defined as parsing and analyzing the semantic features of a given natural language question,reasoning on the knowledge graph and returning the final result.Among the existing cutting-edge research on knowledge graph Question Answering,the information retrieval-based approach has received wide attention.This approach inherits from the idea of information retrieval systems,which model QA as a process of retrieving answers and is therefore promoted by researchers as a major class of research methods.This approach is mainly divided into entity recall,entity representation,and entity selection.Among them,entity recall uses semantic information(entities or relations)in the question to recall candidate answer entities from the graph.While candidate entity representation requires the formation of joint facts with the input question based on entity embedding learning by considering the complexity of semantic composition of natural language question sentences;while entity selection selects the final answer from the set of candidate entities by ranking method.Although the existing knowledge graph Question Answering methods based on information retrieval have made some progress,there are still challenges in their solving of complex interrogative sentences.In this thesis,we will investigate the knowledge graph question and answer method based on information retrieval,mainly focusing on entity representation and entity selection.The main research contents of this thesis include.(1)A Knowledge Graph Entity Representation Learning Method Based on Multiple Embedding Representation IntegrationKnowledge graph representation learning aims to map the factual semantics of all entity nodes or relations in the graph onto a low-dimensional continuous space using representation learning methods.In this research,we propose to use the external texts of entities and relations in the knowledge graph to complement the information of the entities themselves at the semantic level,while combining the spatial adjacency features of triples to better learn the representation of entities and relations.A fully learned learning model for knowledge graph representation can provide assistance in modeling candidate entities in complex question and answer implementations,and also facilitate the optimization of the knowledge graph’s own data such as knowledge graph Complement.In this thesis,we construct a new dataset for the Chinese knowledge graph representation learning task,and compare the method of this thesis with the main representation learning methods on this dataset,and the experimental results show the effectiveness of the proposed method.Meanwhile,this thesis designs a knowledge graph question answering method based on knowledge graph representation,and experiments are conducted on a available Chinese KB-QA dataset to illustrate the effectiveness of this method.(2)Entity Selection Method Based on Global and Local Features of Knowledge GraphThe main goal of this research is to enhance the solution of complex questions answering for knowledge graphs,and for this purpose,we propose this method.On the one hand to enhance the representation learning of knowledge graph entities,while proposing a new entity selection method.For entity representations,we propose to extract the topological structure and semantic features of the knowledge graph as the global features of candidate entity nodes,and model complex Question Answering as a composite classification task based on entity representation and question representation.At the same time,the core inference paths generated by the knowledge graph during the search process are used as local features,combined with the semantic similarity of the question and answer to construct different dimensional features of the candidate entities,and finally form a hybrid feature scorer.For entities selecting,since the final inference paths may be missing,this thesis adopts an unsupervised multiple clustering based approach to design the cluster module.In turn,the final answer clusters are directly generated based on the two-class feature representations of candidate entities,which makes incomplete knowledge graph question answering possible.The results on two common English knowledge graph Question Answering datasets and one Chinese knowledge graph Question Answering dataset show that the method can effectively help solve complex question answering,especially in the case of incomplete graph knowledge.(3)The Knowledge Graph Question Answering system based on Information Retrieval.In this thesis,we combine the research of the former two research and implement an knowledge graph Question Answering system based on information retrieval to address the problem that the current knowledge graph Question Answering is limited in solving complex question answering.Among them,entity linking and entity recall are implemented with wellestablished industry methods,while the entity representation and entity selection modules are implemented based on the methods mentioned before.The system contains modules such as entity linking,entity recall,entity representation,and entity selection.The modules in the system are tested and uploaded online for users to obtain the required information.In summary,this thesis firstly proposes a Knowledge Graph Entity Representation Learning method to better represent the topological features of knowledge graphs for the knowledge graph question answeringg task,and illustrates the effectiveness of the method by link prediction experiments and question answering experiments.Secondly,this thesis investigates a Question Answering method based on global and local features of the knowledge graph,which combines representation learning and semantic parsing methods to improve the solution of complex question answering.Finally,this thesis designs and builds an knowledge graph Question Answering system based on Information Retrieval.