Graph Structure Information For Knowledge Representation Learning

The knowledge graph stores real-world factual representations in the form of a network of triples,which is the cornerstone of many downstream artificial intelligence applications,search,question answering,and personalized recommendation etc..With the continuous ex?pansion of the knowledge base,network-based symbol representation method faces problems such as high computational complexity and poor scalability in the process of semantic com-puting and knowledge reasoning.knowledge representation learning is proposed to deal with those problem and has received widespread attention from industry and academia.It aims to map entities and relationships into low-dimensional dense vectors,so that the structure inside the knowledge base can be well preserved,and semantic computing and knowledge reasoning can be efficiently calculated in low-dimensional space.The existing knowledge representation learning methods mainly focus on the interaction pattern inside a single triplet of the knowledge graph,ignoring the graph structure informa-tion.Graph structure information can dig out a richer interaction pattern of the knowledge base,and describe the semantics of entities and relations more deeply and comprehensively.Therefore,it is of great practical significance to study knowledge representation learning with on Graph Structure Information.This dissertation starts with different graph struc-ture information.In-depth research has been carried out to explore a variety of knowledge representation learning methods for fusion graph structure information.The contents of this dissertation are summarized as below:·We propose a concise but effective model,Context-enhanced Knowledge Graph Em-bedding(CKGE),for joint knowledge base embedding with neighborhood context.For the neighborhood node context,we explore the idea of using language models to rep-resent the common relationship between the entity and its neighboring entities,so as to extract the structural equivalence contained in the neighborhood context.Finally,we designed a joint knowledge representation learning module to fuse the interaction information inside the triplet itself with the neighbor information of the entity.·We present a novel combined model based on multi-step relational path information,Text-enhanced Knowledge Graph Embedding(TKGE),to reason over entities,relations and text.In order to infer direct relationships between entity pairs by using multi-step rela-tion paths between entity pairs,we improved the long and short-term memory networks to utilize the information of entities and relations on the relation paths to model the direct relations.Aiming at the multiple multi-step relation paths between entity pairs we introduced a soft attention mechanism to learn the consistency degree of the seman-tic relationship and direct relationship between the entity pair.In addition,in order to alleviate the problem of sparseness of the knowledge graph,we extracted a large number of text relations to expand the knowledge base relation set,and constructed a more complete knowledge graph data set·We present another novel representation framework based on triplet context,Context-dependent Representation of Knowledge Graphs(CRKG),to utilize the diversity of graph's structural information for knowledge representation.In order to more fully model the graph structure semantics of the triples,we define the information about the neighborhood context of the entity and the multi-step relation path information between the entity pairs as triplet context.Considering the uniqueness of various se-mantics in triplet context,we design distinct knowledge representation learning strate-gies to learn the graph structure semantics of triples.At the same time,we propose a unified framework to fuse the interaction information within the triples and the triplet context.·We propose a knowledge representation learning algorithm based on graph node con-text,denoted Structure Aware Graph Convolutional Network(SAGCN),which lever-ages structural information for modeling the highly multi-relational data characteristic of realistic knowledge graphs.In order to better describe graph structure semantics of the entity,we not only consider the information of the adjacent entity nodes of the en-tity,but also consider the information of the adjacent relation edges of the entity.The combination of these two types of information is the graph node context.Aiming at the graph node context,we designed an innovative graph convolutional neural network to construct a representation of entities's graph structure.In addition,we designed corresponding decoders for entity classification and knowledge graph completion tasksWe perform evaluations on the proposed model on classical tasks such as knowledge graph completion,triad classification,and entity classification.The experimental results show that the knowledge representation learning method based on graph structure informa-tion increases significantly more than other classic baselines,and it also shows that the rich semantics contained in graph structure information can better model knowledge representa-tion.In addition,the knowledge representation learning model based on graph convolutional neural network has achieved better performance than other benchmark models on the task of few shot knowledge graph completion,which illustrates that graph convolutional network can model the semantics of graph structure well.