Entity alignment (EA) which links equivalent entities across different knowledge graphs (KGs) plays a crucial role in knowledge fusion. In recent years, graph neural networks (GNNs) have been successfully applied in many embedding-based EA methods. However, existing GNN-based methods either suffer from the structural heterogeneity issue that especially appears in the real KG distributions or ignore the heterogeneous representation learning for unseen (unlabeled) entities, which would lead the model to overfit on few alignment seeds (i.e., training data) and thus cause unsatisfactory alignment performance. To enhance the EA ability, we propose GAEA, a novel EA approach based on graph augmentation. In this model, we design a simple Entity-Relation (ER) Encoder to generate latent representations for entities via jointly modeling comprehensive structural information and rich relation semantics. Moreover, we use graph augmentation to create two graph views for margin-based alignment learning and contrastive entity representation learning, thus mitigating structural heterogeneity and further improving the model's alignment performance. Extensive experiments conducted on benchmark datasets demonstrate the effectiveness of our method.

翻译：实体对齐（EA）旨在跨不同知识图谱（KGs）建立等价实体关联，在知识融合中扮演关键角色。近年来，图神经网络（GNNs）已成功应用于许多基于嵌入的实体对齐方法。然而，现有基于GNN的方法要么面临实际知识图谱分布中尤为突出的结构异质性难题，要么忽视了对未见（未标注）实体的异构表示学习，这会导致模型在少量对齐种子（即训练数据）上过拟合，从而造成对齐性能欠佳。为提升实体对齐能力，我们提出GAEA——一种基于图增强的新型实体对齐方法。该模型设计了简洁的实体-关系（ER）编码器，通过联合建模全面的结构信息与丰富的语义关系来生成实体的潜在表示。此外，我们利用图增强创建两种图视角，分别用于基于间隔的对齐学习与对比实体表示学习，从而缓解结构异质性并进一步提升模型的对齐性能。在基准数据集上开展的大量实验验证了该方法的有效性。