Research On Multi-hop Question Answering Based On Sparse Knowledge Graphs

The essence of knowledge graph is a structured form of knowledge representation,and knowledge graph aims to model the complex relationships between things using graph structures.Currently,knowledge graphs have been widely used in many fields such as intelligent question answering,semantic search,recommender systems and decision analysis.Knowledge graph based question answering retrieves the answer to a question by parsing natural language questions and using the knowledge graph as a data source to retrieve relevant knowledge.There is already a relatively mature research on knowledge graph question answering methods in academia.In order to make knowledge graph question answering methods capable of solving problems that arise in the real world,researchers have shifted their focus from simple problem question answering research to complex problem question answering research.Simple problems can generally be answered by using only one set of knowledge graph triples,while complex problems require multiple triples for complex logical reasoning to get the answer,making question answering more difficult.Meanwhile,most of the studies only focus on single-hop or multi-hop problems on complete knowledge graphs.However,knowledge graphs in practical applications often face the Incompleteness problem,where sparse knowledge graphs lack the necessary information and paths,making inference and question answering particularly difficult.On the other hand,most KGQA models pre-select answers from specified local domains,leading to the absence of answers in non-local domains during the search.To address the above issues,the following research work is conducted in this paper:1.A new knowledge graph question and answer model Ree-KGQA is proposed for solving the multi-hop complex question and answer task of knowledge graph.The model proposed in this paper obtains the answers to complex questions by end-to-end neural networks,obtaining the embedded representations of KG entities and relations,fusing answer context information using attention mechanisms to enhance the embedded representations of questions,and reasoning through knowledge embedding methods.The model can make full use of the global relational semantics and structured information of the knowledge graph to enhance inference,thus effectively improving the performance of multihop knowledge graph question answering.This paper is validated on relevant knowledge graph datasets,and the experimental results show that superior results can be achieved compared with other models.2.Meta R-KGQA,a meta-relational learning model,is proposed to treat the knowledge graph question answering task as a few-shot learning task for solving different types of complex problems with few samples.The model constructs the required support set of meta-relational learning samples from the training set by an unsupervised retriever.The model learns generic,generalizable relational metainformation from similar problem samples and uses it for the new problem at hand,while updating the parameters of the model.The model learns different types of questions while acquiring global relational semantics,which can effectively improve the question answering performance on sparse knowledge graphs.The experimental results show that the overall performance of the model is better than other models in terms of its hits@1 value.