Research On Entity Embeddings In Knowledge Graph Based On Network Role

Knowledge graph is a knowledge base of graph structures that describes the concepts of entities in the physical world and their interrelationships.Entity embeddings in knowledge graphs aim to characterize the entity features and generate dense and low-dimensional vector representations.So far,knowledge graph can be analyzed from the perspective of numerical calculation instead of original graph structures.Entity embeddings are widely used in downstream tasks related to machine learning.Existing knowledge representation learning approaches mostly model entities and relationship by the triplets of knowledge graphs,ignoring complex relationship between entities,such as multi-step paths and multi-order neighbors of entities.Graph-based representation learning approaches generally believe that entities that are "close" in the graph tend to share features in the vector space and entities that are "far" in the graph will tend to be separated.This kind of approaches ignores structural similarity of entities at a long distance and influences of entity attributes.Entity paths composed of entities with the same role can be used to characterize the entity features comprehensively in the graph structure.The main contents are as follows:(1)Propose methods of network role discovery in knowledge graphs.The concept of network role is used to unify multiple semantic similarities between entities.Different entity similarities will correspond to different methods of role discovery.Four network roles of entities in the knowledge graph are proposed,including homophily role,role based on similar attributes,role based on similar structures,and role based on similar entity centrality.(2)Propose modeling method on entity path based on network role.The random walk method is used to generate entity paths.Two shallow neural network models are used to perform unsupervised feature extraction from entity paths to generate entity embeddings.(3)Apply entity embedding method based on network role to entity profiling in knowledge graphs.The core work of entity profiling is to use entity embeddings to calculate entity similarity,and further measure the distinctiveness of entity labels.The entity profiling visualization system helps users to identify entity by distinctive labels instead of redundant entity description information.In this thesis,the evaluations of downstream tasks such as entity similarity,entity classification and link prediction are performed on the datasets,such as open-domain knowledge graphs DBpedia and other domain-related knowledge graphs.Multi-role entities embeddings for different knowledge graphs can be saved offline or linked to other datasets as background knowledge,which have engineering value.