Research On Large-scale Knowledge Graph Embedding Methods

Knowledge graph is a system of explicitly representing of human knowledge.As one of the important issues of artificial intelligence(AI)in recent years,it has been widely applied to semantic search,human-computer interaction,decision support and other AI application scenarios.However,in various knowledge-driven applications,it is necessary to represent the elements in knowledge graph as low-dimensional dense vectors by knowledge graph embed-ding.It makes up the drawbacks in explicitly representing knowledge,and meets the demand of wide inference,analysis and prediction for knowledge.Although numerous knowledge graph embedding models have been proposed,there are obvious flaws in inaccuracy and semantics-insufficiency of knowledge representation:(1)Models based on translation or distance metrics are unable to model complex relations better;(2)Some models only use the structure informa-tion of triples in knowledge graph;(3)Most models ignore the differences between concepts and instances in knowledge graph.To solve these issues and improve the quality of knowledge representation,this paper explores how to effectively use multi-source information in knowl-edge graph,as a complement to structure information of triples.The main researches are as follows.1.Proposing a knowledge graph embedding model with entity types TransET:For the issue that most knowledge graph embedding models ignore entity types,a projection function based on representations of entity types,in the form of circular convolution,is designed to construct different representations of entities while belonging to different types.Translation-based method is then used to learn the structure information of triples composed of mapped entities as well as relations.To increase the differences between entities belonging to similar type,while take into account some similarities between them,the types of sampled entities are limited with a certain probability when negative sampling.TransET can enrich the semantics of knowledge representation,and can better modelling complex relations.2.Proposing a knowledge graph embedding model for jointly learning concepts and in-stances JECI and JECI++:For the issue that most existing knowledge graph embedding models ignore differences between concepts and instances,a prediction function based on neighbor in-formation and belonging concepts,in the form of a circular convolution,is designed to predict target instances.This function links concepts and instances,thus they can be learned j ointly.To address the low generalizability and high complexity of JECI,JECI++simplifies the hierarchi-cal concepts and incorporates relations into neighbors.Both in JECI and JECI++,the quality of knowledge representation is further improved by limiting the concepts which sampled instances belongs to while negative sampling.These two models can not only tackle the poor quality of knowledge representation caused by differences between concepts and instances,but also alle-viate the problem that instances having similar reltions and belonging to similar concepts gather in embedding space.The models presented in this paper are systematically evaluated and compared with the baselines by link prediction and triple classification,on the datasets extracted from Freebase,DBpedia,YAGO and other real-world knowledge graphs.The experimental results of TransET are 2.2%to 9.8%higher than the optimal baseline model and the experimental results of JECI++are 1.7%to 18.6%higher than the optimal baseline model.These results fully demonstrate the following conclusions.(1)Information contained in entity types is beneficial to learning knowledge representation.(2)Concepts and instances are necessary to be differentiated for more precise representation of knowledge.(3)Neighbor information of instances can improve the recognition of similar instances.(4)The circular convolution adequately captures the se-mantics between objects including instances,relations and concepts.