Research On Entity Alignment In Knowledge Graph Via Relation-Aware And Layer Attention Transfer

Entity Alignment refers to finding entities with the same semantic meaning in different knowledge graphs,and is a vital step in knowledge graph fusion.A vital issue in the entity alignment task is how to distinguish whether two entities in different knowledge graphs point to the same thing in the real world.In recent years,graph neural networks have become the most commonly used network framework for entity alignment tasks due to their ability to identify isomorphic subgraphs.However,most of the existing graph neural network-based entity alignment methods do not make full use of the relationship information among entities in knowledge graphs,and cannot represent the relationship diversity of knowledge graphs better.In addition,assigning the same weight to all neighboring entities makes the important neighboring entities not fully utilized.To solve the above problems,the following studies are conducted in this paper.(1)To address the problem that relational diversity cannot be fully utilized,this paper proposes a differential adjacency matrix-based relational-aware graph convolutional network entity alignment method(-REA: Relation-Aware Graph Convolutional Network Entity Alignment via a Differential Adjacency Matrix,A-REA).This network uses a differential adjacency matrix to distinguish different relationships,aggregates the feature representations of neighboring entities in the process of learning entity feature representations,and incorporates different relationship information into the feature representations of entities.The entity representation is enhanced by the specific relational semantic and structural information of entities in the knowledge graph,as a way to better compute the subgraph similarity between different knowledge graphs.(2)To address the problem that important neighbor entities cannot be fully utilized,this paper proposes Layer Attention Transfer-Based Entity Alignment in Graph Convolutional Networks(Lat EA).In order to utilize the information carried by the distant-hop neighbor entities more effectively,the network saves the attention coefficients between the entities in the previous hop through layer attention transfer,which is used to calculate the attention coefficients between the neighbor entities and the central entity in the next hop,and uses a gating mechanism to adaptively integrate the neighbor entities in different hops into the central entity representation,thus obtaining a more accurate entity representation.Finally,this paper conducts comparison experiments on the publicly available datasets DBP-15 K and DWY-100K(including three cross-linguistic sub-datasets and two cross-knowledge graph sub-datasets)to verify the effectiveness of this paper’s method on cross-linguistic and cross-knowledge graph entity alignment datasets,and verifies the effectiveness of each module through ablation experiments.Distinguishing the relationship between entities and the importance of neighboring entities is an effective signal for entity alignment,especially for structural heterogeneity between different knowledge graphs.