Research On Entity Alignment For Knowledge Graphs

A knowledge graph is a knowledge representation that is built on a graph structure and has the advantages of scalability and flexibility.Therefore,knowledge graphs have become widely used in the field of natural language processing.Different organizations or institutions build knowledge graphs independently based on their needs to optimize their business,this might result in information duplication or complementarity across knowledge graphs.Furthermore,a single knowledge graph covers a limited amount of information,which is a limitation for knowledge-driven applications.Hence,knowledge graph fusion has emerged as a development trend.Knowledge graph entity alignment is a crucial step in realizing knowledge graph fusion,which can connect entities present in different knowledge graphs but have the same real-world meaning,thus expanding the size of the knowledge graph and advancing fields like intelligent Q&A,semantic search,and recommender systems.Currently,commonly used knowledge graph entity alignment methods include traditional symbol-based entity alignment methods and representation learning-based entity alignment methods.Traditional symbol-based methods rely on symbolic features to align entities,while representation learning-based methods utilize embedding techniques to achieve entity alignment.Although existing knowledge graph entity alignment methods have made significant progress in improving the performance of entity alignment,this research area still faces some problems.This paper conducts research based on the knowledge graph entity alignment problem,and the main research work is as follows:1.Entity alignment based on structure and semanticsCurrent studies have not sufficiently explored how to effectively combine structure and semantic information.In response to the aforementioned problems,this paper proposes a model which is called as Relation Awareness and Attribute Involvement Based Entity Alignment Model(RAEA).Firstly,to incorporate the semantic information embedded in the relations,the model constructs a dual graph for each knowledge graph to be aligned and utilizes the attention mechanism to capture the interaction information between dual graphs and original graphs,thus calculating relation-aware entity embedding.Secondly,to obtain information about neighborhood structure,the model uses graph convolutional networks to learn the structure representation of a knowledge graph.In addition,to alleviate the problems arising from the sparsity and heterogeneity of the knowledge graph structure,the model utilizes graph convolutional networks to learn the representation of entity attributes.Finally,the model uses a pre-trained language model and word embedding model to learn the textual representation of entities to enrich the embedding representation of those entities.Through validation on three publicly available datasets,the experimental results demonstrate that the model effectively combines structural and multiple semantic information,which can enhance the performance of entity alignment.2.Entity alignment fusing knowledge representation and symbolCurrent studies have not fully explored how to effectively combine the advantages of symbol-based traditional entity alignment methods and representation learning-based entity alignment methods.To respond to the above problems,this paper proposes a model which is called as Entity Alignment Model Incorporating Entity Embedding and String Similarity(EAES).Firstly,to obtain informative entity embeddings,the model utilizes graph convolutional networks to learn the structure representations and attribute representations of the entities and concatenates entity vectors and attribute vectors.Secondly,to alleviate the problem of noise propagation introduced by deepening the layers of deep neural networks,the model optimizes the network structure.In addition,to further improve the alignment performance of the model,the model calculates the string similarity of entity names using a symbol-based traditional entity alignment method.Finally,the model combines the similarity of entity embeddings with the similarity of entity name strings and performs entity alignment.Through validation on five public datasets,the experimental results demonstrate that the model effectively combines the strengths of symbol-based traditional entity alignment methods and representation learning-based entity alignment methods,which can improve the performance of entity alignment.3.Global entity alignment based on iteration and neighborhood matchingCurrent studies have not fully explored how to effectively combine neighborhood matching,expanding seed sets,and global entity alignment.In response to the above problems,this paper proposes a model which is called as Semi-Supervised Neighborhood Matching Model for Global Entity Alignment(SNGA).Firstly,the model performs entity alignment from the perspective of neighborhood matching,taking into account both the matching of first-order relations and the matching of first-order neighboring entities,in order to mitigate the negative impacts of noisy neighboring entities.Secondly,to expand the size of the training data,the model utilizes a bidirectional nearest-neighbor iterative strategy,when two entities are most similar to each other,they are considered to be aligned.Moreover,to reduce the cost of manually labeling data,this paper further explores unsupervised entity alignment.Finally,to alleviate the alignment conflict problem,the model employs a deferred acceptance algorithm to perform entity alignment from a global perspective.Through validation on three public datasets,the experimental results show that the model effectively combines neighborhood matching,expanding seed sets,and global entity alignment,which can improve the performance of entity alignment.In conclusion,to address the current problems in the field of knowledge graph entity alignment,this paper investigates entity alignment based on structure and semantics,entity alignment fusing knowledge representation and symbol,and global entity alignment based on iteration and neighborhood matching.This paper holds certain research values and meanings.