Entity Alignment Between Knowledge Graphs Via Graph Neural Network

Entity alignment seeks to discover identical entities in different knowledge graphs(KGs).Recently,graph neural networks(GNNs)have emerged as a powerful paradigm for embedding-based entity alignment due to their capability of identifying isomorphic subgraphs.GNNs generate the representation of an entity by aggregating the features of the entity itself and its neighbours.However,in real-world KGs,identical entities usu-ally have non-isomorphic neighborhood subgraphs,which would cause GNNs to yield different representations for them and hurt entity alignment performance.To resolve such heterogeneity and improve the expressiveness of GNNs on identifying similar en-tities,this thesis investigates two different models from two perspectives(i.e.,entity relations and attributes),respectively.Specifically,this thesis proposes a multi-hop re-lational neighbourhood aggregation model and an inductive attribute aggregation model for entity embedding learning,respectively.This thesis further integrates the two mod-els to the entity alignment framework in an end-to-end manner to alleviate the influence of structural heterogeneity on entity alignment.In terms of relational neighbourhood aggregation,this thesis proposes Ali Net,which expands the overlap between neighborhood subgraphs by introducing the multi-hop neighbor information of entities.It uses an attention mechanism to highlight the multi-hop neighbours closely related to entity alignment and reduce the noise caused by unrelated neighbors.Then,Ali Net employs the gating mechanism to combine the one-hop and multi-hop neighbour information to obtain the final representations of en-tities.In addition to the widely-used relational neighbours,this thesis further explores the effectiveness of entity attributes.This thesis proposes an inductive entity embedding model(namely Ai E)based on the attribute-aware attention mechanism.Ai E learns the embeddings for attribute values through pre-trained language models and extracts the values'common features through convolution operations to generate attribute represen-tations.It further uses a GNN based on attribute co-occurrence for attribute smoothing representation and uses an attention mechanism to aggregate attribute embeddings to obtain inductive representations for entities.Finally,this thesis proposes an entity alignment framework based on GNNs,which can integrate the two models Ali Net and Ai E,respectively.Moreover,it can also com-bine them together as a unified end to end model called Ali Net_{Ai E},where Ai E gen-erates the initial attribute-aware representations for entities and Ali Net aggregates the neighborhood information to obtain the final entity embeddings.The two models can enhance each other to improve entity alignment.This thesis conducts experiments on three entity alignment benchmark datasets,i.e.,DBP15K,DWY100K and Open EA.The results demonstrate the effectiveness of Ali Net and Ai E and the combined variant Ali Net_{Ai E}.Besides,experiments and analyses demonstrate the enhancement effect of the attribute information on GNNs and entity alignment.This thesis also conducts ablation studies to show the effectiveness of each module in the entity alignment framework.