Research On Knowledge Graph Representation Learning With Confidence

With the continuous development of artificial intelligence technology,KG has become the core data support for structured knowledge-driven intelligent applications.KG is essentially a kind of semantic network.Its nodes represent entities or concepts,and edges represent the semantic relationship between entities or concepts.KG contains a large amount of structured knowledge,shaped like a triple: <left entity,relationship,right entity>,for example: <Obama,born in,Hawaii>;a tuple: <entity,entity type>,for example: <Obama,Characters>.KG helps machines understand knowledge and semantic information and are widely used in intelligent applications such as Semantic Search,Question Answering,and Intelligent Customer Service.However,due to the limited accuracy of knowledge extraction technology in the automatic construction of KGs,KG faces very serious noise problems,such as knowledge triplet noise and entity type labeling noise.Therefore,the knowledge graph modeling technology for noise has important application and research value.Knowledge Graph Representation Learning as the core technology of knowledge graph has become a research focus in the field of knowledge graph.Representation Learning is a method based on machine learning.By constructing a semantic model of entities and relationships,the representation learning technology embeds entities and relationships into a vector space and uses vectors to represent them,so as to capture the semantic information or essential geometric structure of entities and relationships,to achieve entity classification or knowledge reasoning.However,traditional knowledge graph representation learning models face two major problems:(1)they ignore the noise problem,and there are shortcomings that will cause errors in subsequent application systems;(2)they only focus on knowledge graph relationship prediction and ignore knowledge graph entity type reasoning.In view of the above problems,this paper proposes a confidence-based representation learning model(Trust E),which aims to solve the problem of entity type representation learning in a noisy environment.The model can detect the entity type noise that may exist in the knowledge map and implement Entity type reasoning.Specifically,we first consider that entities and entity types are semantically different and have complex relationships,so we use a projection matrix to project entities and entity types into different semantic spaces,and then use the constructed two-tuple(entity,entity Type)confidence to improve representation learning of entity types.In order to make the confidence more universal,this paper only considers the internal structure information of the knowledge graph,and proposes two types of confidence:(1)Local tuple trustworthiness(LT)based on the local information of the tuple;Global triple trustworthiness(GT)based on triples related to the binary.Finally,in this paper,three experiments,entity type noise detection,entity type prediction,and entity type classification,were performed in two real-world datasets FB15 k ET and YAGO43 k ET.The experimental results show that the effectiveness of the Trust E model is significantly better than other latest benchmark models.The Trust E model can learn a better entity type representation in a noisy environment and implement entity type reasoning.