实体对齐是整合多源知识图谱的核心步骤，在知识图谱融合中具有重要意义。然而，现有研究在充分利用和有效融合多种实体特征信息方面仍然存在不足。为了解决这一问题，提出一种基于最优传输的多视角特征融合实体对齐方法。具体而言，在特征学习阶段，首先利用预训练语言模型提取实体名称和属性的语义特征，以缓解知识图谱结构稀疏和结构差异的问题。随后，通过计算实体名称的字符串相似性进一步提升对齐性能。接着，引入关系感知图注意力网络，捕捉实体之间的多重关系，从而获取跨图谱的结构特征信息。在特征融合阶段，采用Wasserstein距离整合多视角特征进行实体对齐，并通过Fused Gromov-Wasserstein距离，将多视角特征成本矩阵与图谱内部结构成本矩阵相结合，得到最终的对齐结果。在公开的跨语言数据集ZH-EN、JA-EN和FR-EN上，所提方法的Hits@1评价指标分别达到0.989、0.992和0.999。与其他模型相比，所提方法在对齐效果上表现更优。

Entity alignment, a critical step in integrating multi-source knowledge graphs, plays a vital role in knowledge graph fusion. However, existing studies still exhibit limitations in effectively utilizing and integrating diverse entity feature information. To address this issue, an optimal transport-based multi-perspective feature fusion method was proposed for entity alignment. Specifically, during the feature learning phase, the semantic features of entity names and attributes were first extracted using pre-trained language models to mitigate structural sparsity and heterogeneity across knowledge graphs. Subsequently, the alignment performance was further enhanced by calculating string similarity between entity names. A relation-aware graph attention network was then introduced to capture multi-relational interactions between entities, thereby obtaining cross-graph structural features. In the feature fusion phase, the Wasserstein distance was employed to integrate multi-view features for alignment computation. The fused Gromov-Wasserstein distance was adopted to combine multi-view feature cost matrices with intra-graph structural cost matrices, generating final alignment results. Experimental evaluations on publicly available cross-lingual datasets (ZH-EN, JA-EN, FR-EN) demonstrate that the proposed method achieves Hits@1 scores of 0.989, 0.992, and 0.999, respectively. Compared to other models, the proposed method demonstrates superior performance in alignment.