Anchor-Aware Position Enhancement for Weakly Supervised Knowledge Graph Alignment

Entity alignment (EA) models currently rely heavily on large-scale labeled anchor links to succeed. However, these models often face challenges due to the limited availability of anchor links. To enhance the alignment performance of existing models under weakly supervised settings, some researchers have turned to active learning to search for more informative anchor links. Despite these efforts, the fundamental issue remains that existing neural EA models tend to overfit the limited anchor links, resulting in poor alignment performance. To tackle this problem, this paper addresses the challenge of low-resource EA by focusing on anchor entities and proposes a novel approach called Anchor-aware Position-Enhanced Low-resource Entity Alignment (APLEA). In the encoding process, we introduce a simple yet effective position encoding method that captures the long-term relationships between anchor entities and other distant entities using random walk with restart (RWR). This method allows us to extract more advanced positional features through a position attention layer. By incorporating structural information, relational information, and positional information from these three aspects, we obtain the final entity embeddings for alignment. In the decoding process, we address the issue of existing EA models disregarding the matched neighborhood constraints when inferring alignment results. Specifically, neighboring entities of a specific entity $e$ may not be strictly mapped to those of $e$’s counterpart. To overcome this limitation, we propose a simple refinement approach to the final similarity matrix without the need for additional training. Extensive experiments conducted on publicly available datasets demonstrate that APLEA significantly outperforms the state-of-the-art approaches by a large margin in weakly supervised EA tasks.