Structure And Rule Information Based Knowledge Graph Representation Learning

With the gradual transition of artificial intelligence to a more intelligent stage of cognitive intelligence,knowledge graphs are playing an increasingly important role in the construction of cognitive intelligence applications,such as intelligent search,intelligent dialogue,and personalized recommendation.The existing knowledge graphs are represented by symbols,where entities and relations are uniquely marked with uniform resource identifiers.Faced with knowledge graphs containing massive triplets,the form of symbolized knowledge representation makes it difficult for computers to easily and efficiently perform tasks such as knowledge computation,knowledge detection,and reasoning completion,and also hinders the application of knowledge graphs in other fields of artificial intelligence.To solve the above mentioned problems,the knowledge graph representation learning method is proposed.This kind of method has the advantages of high efficiency and convenience in knowledge computation and reasoning completion tasks,and has quickly become a hot research topic in the field of knowledge graph at this stage.However,it is inappropriate to think of knowledge graphs as a way of representation learning on general datasets consisting of a large number of triplets.How to model the inherent graph-structured characteristics of knowledge graph and its implicit logical rules and relational patterns has a very important influence on the construction of high-quality reasoning completion models.Based on this,this dissertation systematically studies and discusses several key issues related to modeling structural and rule information in knowledge graph representation learning methods.The innovations and main contributions can be summarized as follows:First,this dissertation proposes a graph structure-based knowledge graph representation learning model G2 SKGE,aiming to model the complex contextual structure information around entities in the knowledge graph.The model creatively proposes a link-based(i.e.,the triplet)contextual structure information fusion mechanism,and leverages the link embeddings to obtain the fusion embeddings for entities.To more rationally and fully integrate the links around the entity,the proposed model conducts a systematic study on leveraging contextual structure information,from the aspects of the encoding method,the participation manner and the number of links.The experimental results on the real knowledge graph datasets show that: the G2 SKGE model can adaptively select important links to participate in the construction of fusion embeddings for entities,so as to obtain high-quality entity embeddings.Compared with multiple representative baseline models,the G2 SKGE model achieves significant performance improvements on both the tasks of knowledge graph completion and triplet classification.Second,this dissertation proposes a relational path-based knowledge graph representation learning model Inter ERP,which aims to learn the semantic path structure between entity pairs in the knowledge graph.From the perspective of enhancing the interactions between different embeddings,the model proposes a new learning method for obtaining representations of entities,relations and relational paths.The interactions between the embeddings of different elements in the knowledge graph imply the semantic association between them.By introducing an interaction matrix,and with the help of the Inception network that is widely used in the image field,the proposed model can capture the interactions between entities and relations within a triplet,and the interactions between different relations along the path.The experimental results on several benchmark datasets prove that the performance of the proposed Inter ERP model can match or outperform baseline models on the tasks of knowledge graph completion and triplet classification.Third,this dissertation proposes a logical rule-based knowledge graph representation learning model JSSKGE,which aims to deal with the problem that the discretized and symbolized logical rules are difficult to model.The model designs a soft logic rule mining algorithm SPCRE,to automatically extract the required soft logic rules from the knowledge graph.Soft logic rules are discrete and symbolic,while knowledge graph embedding models require continuous and numerical feature vectors as input,which brings great challenges to the construction of embedding learning models based on logic rules.To conquer this challenge,on the one hand,the JSSKGE model introduces the concept of argument embeddings of relations;on the other hand,the model further subdivides logic rules into five types of sub-rules,namely,reflexive rules,symmetry rules,equivalent rules,inverse rules,and transitive rules.These five types of sub-rules can derive the restrictive constraints on the argument embeddings of relations.In this way,we can construct a knowledge graph representation learn model that combined with soft logic rules.The experimental results on real datasets demonstrate the effectiveness and superiority of the proposed JSSKGE model.Fourth,this dissertation proposes a logical pattern-based hyperbolic knowledge graph representation learning model Hyper JSS,which aims to solve the problem of modeling the hidden characteristics of the conceptual layer at the upper level of the knowledge graph.The proposed model embeds entities and relations into the hyperbolic space,to represent the knowledge graph that always exhibits a hierarchical structure pattern in semantic concepts,such as the hierarchical structure pattern among directors,movies,and actors.Compared with the traditional Euclidean space model,the hyperbolic space model has the natural advantages for modeling hierarchical structure pattern data.Furthermore,the Hyper JSS model employs Givens transformations such as rotation and reflection transformations to model the logical patterns of relations,such as the symmetry relation similar＿to and the antisymmetric relation has＿part.Combined with the previous work on the modeling of graph structure and logical rule information,the proposed model can achieve the goal of modeling knowledge graph data from multiple dimensions.The experimental results on the benchmark datasets prove that the proposed Hyper JSS model achieves significant improvements on the tasks of knowledge graph completion and triplet classification.In summary,this dissertation effectively solves the problems of modeling structural and rule information in the knowledge graph and lays a solid foundation for the real reasoning applications of knowledge graph representation learning methods and other artificial intelligence applications in the future.