Research On Entity Alignment With Structural Heterogeneity Based On Complex Network

Entity alignment,the key technology in the process of knowledge integration,aims to match entities that refer to the same real-world identity from different knowledge graphs.However,the structural heterogeneity between knowledge graphs poses great challenges to entity alignment task.From the perspective of data structure,knowledge graph,the heterogeneous information network with high complexity,is a kind of complex network.Hence,considering the impact of structural heterogeneity on entity alignment task,this thesis conducts an in-depth study based on the topological properties of complex networks.As a result,this thesis proposes corresponding entity alignment algorithms from the perspectives of topological structure and semantic content.The main contributions of this thesis are as follows:(1)This thesis proposes a novel degree aware based adversarial graph convolutional network(DAGCN)for entity alignment.In this work,we define the first metric,heterogeneity of degree(HED),to quantify the structural heterogeneity between knowledge graphs.Based on the topological properties of complex networks,this metric quantifies the structural heterogeneity from degree difference feature between equivalent entities and fills the void of quantitative research and representation of structural heterogeneity in entity alignment task.To mitigate the impact of heterogeneity,a model that integrates GCN and GAN is proposed to generate degree level irrelevant embedding by utilizing a degree aware adversarial training framework.In the multi-tasking framework,the alignment-oriented and degree aware process jointly improve the final embedding and alignment results.Finally,the rationality of HED and the effectiveness of DAGCN in mitigating the impact of structural heterogeneity are also proved by the experimental results on WK31-15k and DBP 15k datasets.(2)This thesis proposes a graph semantics based neighboring attentional network(GSNAN)for entity alignment.In this work,structural heterogeneity is characterized as the problem of common neighbors between equivalent entities and is further explored from semantic content.Compared with other general complex network models,knowledge graphs have rich semantic content and can be used as important supplementary information to address the structural heterogeneity problem.First,the framework utilizes a pre-trained word vector to extract semantic representation in knowledge graphs and uses the semantic features of entities to enhance the performance of entity alignment.Then considering the relation semantic similarity of common neighbors,a neighboring semantic based graph attentional model is proposed.This model conducts a weighted combination according to the neighboring importance and further alleviates the impact of structural heterogeneity.In the experimental results on the WK31-15k dataset,the Hits@1 score of the GSNAN model reaches a maximum of 66.10%,which is better than other state-of-the-art models.