Research Of Iterative Entity Alignment Method Based On Graph Representation Learning

Under the background of the rapid development of Internet information,knowledge graphs,as a data structure that can efficiently organize data,are widely used in various high-intelligence natural language tasks.However,the information content of a single knowledge graph is small,and sometimes it is difficult to meet the task requirements.Therefore,it is necessary to use knowledge graph fusion technology to fuse multiple different knowledge graphs to expand the coverage of information,thereby improving the performance of its downstream tasks.The primary task of knowledge graph fusion is entity alignment.Entity alignment refers to finding entities with different surface forms but the same meaning in different knowledge graphs.How to judge whether two entities with different surface forms are equivalent is the difficulty of the entity alignment task.Most of the existing entity alignment methods mainly rely on the structural information of the knowledge graph and some pre-aligned entities that have been labeled and represent the entities as low-dimensional vectors.The vectors in the embedding space are used to determine whether the entities are aligned by calculating the similarity between the vectors.These entity alignment methods only consider the relationship triples in the knowledge graph,while ignoring the attribute triples that also contain rich information.In addition,the quantity and quality of pre-aligned entities will directly affect the performance of entity alignment.This paper studies these two problems,and the main work is as follows:1.This paper proposes a joint embedding entity alignment model based on graph representation learning(JEGCN).The model utilizes both relationship triples and attribute triples in the knowledge graph.First,the multi-layer graph convolutional network model is used to embed the structural information contained in the relationship triples,and the highway network,entity name initialization strategy,and the nearest neighbors negative sampling strategy improves the representation ability of the structure embedding.Secondly,the model calculates the relationship embedding according to the association between the entity and the relationship in the relationship triples,and generates the structure-relationship joint embedding of the entity,and trains the joint embedding.Again,the model uses a multi-layer graph convolutional network model to embed attributes in attribute triples.Finally,the structure-relationship joint embedding is connected with its attribute embedding to generate the structure-relationship-attribute joint embedding.The experimental results show that JEGCN has a good performance on the DBP15K dataset.The alignment accuracy Hits@1 on the DBP15K_FR-EN dataset reaches 90.26%,and the Hits@10score reaches 96.66%.2.JEGCN belongs to supervised learning and has a strong dependence on pre-aligned entities,so this paper proposes an iterative entity alignment method based on graph representation learning,named JEGCN?iter.This method is a semi-supervised learning method that uses a graph convolutional network model for modeling and training,generates new aligned entities,and adds them to the training data to guide the subsequent training process.To improve the reliability of the aligned entities produced by the model,this paper proposes a method for setting thresholds and a mutual nearest entities strategy,and adopts a re-initialization strategy to reduce error propagation probability during iteration.The Hits@1 of JEGCN?iter on the DBP15K dataset is on average about 10%higher than that of JEGCN,which proves that the iterative alignment method can effectively improve the effect of entity alignment and reduce the dependency of the entity alignment model on pre-aligned entities.3.Based on the iterative entity alignment of semi-supervised learning,this paper further proposes an unsupervised entity alignment method JEGCN?usv that does not require pre-aligned entities at all.The method jointly calculates the distance between entities through the entity's semantic embedding distance and Levenshtein distance,thereby generating a preliminary set of aligned entities,and then using these aligned entities to start the iterative alignment training process.Experimental results show that this unsupervised entity alignment method performs close to the supervised JEGCN model on the DBP15K dataset,and even surpasses some supervised and semi-supervised entity alignment methods.