SimEA: an entity alignment model based on multi-aspect similarity features

Entity alignment (EA) links equivalent entities across different knowledge graphs (KGs) to integrate information and support downstream tasks. Most current methods typically rely solely on either structural or attribute information, failing to fully utilise the abundant multi-aspect information available in KGs. With the rise of graph neural networks, a growing number of entity alignment methods adopt Graph Convolutional Networks (GCNs). However, aggregating dissimilar neighbours and attribute values can introduce noise due to the heterogeneity of KGs, ultimately degrading alignment performance. In this article, we propose an EA model, SimEA, which integrates multi-aspect similarity features. SimEA employs a Multilingual Bidirectional Encoder Representations from Transformers (BERT) to learn embeddings from entity names/descriptions, neighbours, and attribute values. A similarity feature module is then applied to transform and aggregate these similarity features, followed by a scoring module to evaluate entity pair alignment likelihood. Experimental results demonstrate that SimEA significantly outperforms baseline methods on the hit@1 metric on the open dataset, confirming its effectiveness in entity alignment.