Research Of Dual Graph Model For Knowledge Base Question Answering

The knowledge graph is a graph composed of entities as nodes and relations as edges.Knowledge base question answering is a direction of widespread concern in recent years.Knowledge base question answering is to give a question and find the corresponding answer in the knowledge graph.According to the times of inferences in the knowledge graph,it is divided into single-hop reasoning and multi-hop reasoning.Single-hop knowledge base question answering only needs to search for the answer within one hop adjacent to the subject entity to get the result;multi-hop knowledge question answering requires reasoning on multiple edges of the knowledge graph.Compared with single-hop reasoning,multi-hop reasoning is more complicated.In general,knowledge base question answering mainly faces the following problems: one is the sparseness of the knowledge graph,which is usually incomplete and lacks many links;the other is the complexity of the question semantics,how to fully represent the semantics of complex questions;what's more,the extended search for the graph structure is a problem,how to achieve better search and "jump" far enough and wide enough in the knowledge graph.For these problems,this thesis combines the methods of semantic analysis and information extraction,and proposes a Dual Graph Model based on graph neural networks.On the one hand,for the complexity of the semantics of the question,using the question and related entities builds a semantic sub-graph as the matching heuristic information of the answer entity of the knowledge graph;On the other hand,the nodes and edges of the semantic subgraph are used to match the entities and relations of the knowledge graph,and the matched entities and relations use Beam Search Algorithm and Heuristic Balanced Algorithm to filter related entities and relations for building knowledge embedding graph.Constructing a knowledge embedding graph not only alleviates the sparsity problem of the knowledge graph,but also enables searching and matching in a relatively larger range.The model in this thesis is trained and tested on the widely used public dataset MetaQA,and achieved good results on both MetaQA KG-Full and MetaQA KG-50(Sparse knowledge graph).Compared with Embed KGQA model,the 3-hop QA of the model at MetaQA KGFull and MetaQA KG-50 increased by 0.7% and 1.8%.Overall,Dual Graph Model better represents the semantics of the question and improves the accuracy of the question and answer in the case of sparse knowledge graphs.