Research On Representation Learning-Based Knowledge Graph Reasoning

Knowledge graphs are structured semantic knowledge bases,which are widely used in various fields,such as search engines,personalized recommendations,medical diagnoses,and financial services.However,lots of facts are missing in knowledge graphs,which limits their practical applications.Therefore,developing knowledge graph reasoning techniques—which infer new knowledge based on existing knowledge—is of great significance.Recently,representation learning-based knowledge graph reasoning has attracted increasing attention.Generally,representation learning-based knowledge graph reasoning techniques can be divided into three categories:shallow graph embeddings,deep graph embeddings,and complex logic embeddings,of which the first two techniques aim to perform single-step reasoning,and the last one aims to perform multi-step reasoning.Despite the great achievements of the three categories of reasoning techniques,they still face various issues.First,it is important for shallow graph embeddings to model different semantic features of knowledge graphs.However,existing models still struggle to model some critical features,such as semantic hierarchies and semantic similarity.Second,deep graph embeddings improve reasoning performance by introducing neural networks.However,the introduced neural networks also reduce reasoning efficiency.Third,complex logical embeddings have to deal with various multi-step reasoning scenarios,while current models have limited applicability.To improve reasoning performance,computational efficiency and applicability of existing methods,this dissertation conducts in-depth research in the following four aspects.(1)In shallow graph embeddings,this dissertation studies how to model semantic hierarchies in knowledge graphs.This dissertation proposes a hierarchy-aware knowledge graph embedding method,which uses the Cartesian product of two-dimensional polar coordinate systems as the representation space.More specifically,the proposed model uses the radius coordinates to model entities at different levels of the hierarchy and the angular coordinates to distinguish entities at the same level of the hierarchy.Experiments demonstrate that the proposed model significantly outperforms previous state-of-the-art methods in the single-step knowledge graph reasoning task and effectively captures semantic hierarchies in knowledge graphs.(2)In shallow graph embeddings,this dissertation further studies how to model semantic similarity in knowledge graphs.This dissertation proposes a duality-induced regularizer for tensor factorization-based models,of which the scoring functions have difficulty preserving entities’ semantic similarity in representation spaces.The regularizer is mainly based on an observation:for an existing tensor factorization-based model,there is often another distance-based model closely associated with it,which can be used as effective constraints for entity embeddings.Theoretically,the proposed regularizer gives an upper bound of the tensor nuclear 2-norm of the adjacency tensor of knowledge graphs.Experiments demonstrate that it is widely applicable and effective in improving the reasoning performance of tensor factorization-based models.(3)In deep graph embeddings,this dissertation studies how to improve the computational efficiency of existing models with graph convolutional networks(GCNs)as the encoding module.This dissertation conducts extensive experiments to demonstrate that explicit graph structure modeling in GCNs does not significantly impact reasoning performance,which is in contrast to the common belief.Based on the observation,this work removes the redundant parts in GCNs and proposes a simple yet effective framework for single-step knowledge graph reasoning,which implicitly models graph structure when updating parameters with gradient descent.Experiments demonstrate that the proposed framework leads to comparable performance with GCN-based reasoning methods while being more computationally efficient.(4)In complex logic embeddings,this dissertation studies how to extend existing methods’ application scenarios.This dissertation proposes a cone embedding method for multi-step knowledge graph reasoning.Specifically,the proposed cone embeddings use the Cartesian product of two-dimensional cones as the representation space,which can model any entity sets in knowledge graphs.Then,to model logical conjunction,disjunction,and negation,this dissertation proposes an encoding module based on cones’intersection,union,and complement.Moreover,this dissertation proposes a scoring function based on the set inclusion relationship of cones.To our knowledge,the proposed cone embedding model is the first geometry-based complex logic embedding method that can simultaneously model conjunction,disjunction,and negation.Experiments demonstrate that the proposed model is widely applicable in various multi-step reasoning scenarios and significantly outperforms existing state-of-the-art methods.