Research On Knowledge Graph Completion Technology Based On Graph Neural Networks

With the proposal and rapid development of a new generation of artificial intelligence,cognitive intelligence has become a hot spot of current research.As a container of knowledge,knowledge graphs provide knowledge organization management and application support services for intelligent applications,and become the core of cognitive intelligence development.Knowledge graphs describe the real world containing several objective facts,but due to the limitations of its construction process(most of which are constructed manually or semi-automatically),there are a lot of missing knowledge and hidden knowledge not discovered.Therefore,it is of great significance to use knowledge graph completion technology to discover implicit links and complete missing knowledge.Knowledge graph completion refers to discovering link features based on the existing knowledge in the graph,so as to infer potential associations and complete the missing triple.Current mainstream knowledge graph completion techniques,such as rule-based completion models,representation learning-based completion models,and neural network-based completion models,have achieved good results in knowledge graph completion.However,most of the existing methods regard the knowledge graph as a simple graph,mainly relying on the convolution of the spectral graph or the information of the first-order neighborhood,ignoring the local structure in the knowledge graph,such as some loop information with special semantics,and at the same time Long-distance association information is not considered.For solving problems mentioned above,this thesis considers how to obtain and utilize various semantic information and long-distance effective information under these local structures to improve the knowledge graph completion effect.The following contents are the work of this thesis:(1)The existing graph neural network does not consider the special local structure,to solve this problem,a local structure aware graph neural network is proposed,and three local structures are defined.These local structures rich in semantic information are fused in the entity embedding learning stage;(2)The existing graph neural network can only use close-range information,which is unreasonable for highly heterogeneous graph structured data such as knowledge graphs.For solving this problem,this thesis proposed a long-distance information fusion mechanism to aggregate remote related information of entities and improve the effect of knowledge map completion;(3)Based on the above completion algorithm model,an intelligent question answering system based on knowledge graph completion is designed and implemented,which realizes the functions of question answering,data query,and knowledge graph completion.