Geographic Entity Alignment Based a Semi‐Supervised Dual Strategy Iteration

ABSTRACT Geographic entity alignment (GeoEA) is a crucial method for geographic knowledge integration, enabling the construction of large‐scale and multifunctional geographic knowledge graphs (GeoKGs) by matching and integrating identical entities across different GeoKGs. However, the complex relational information within GeoKGs is often overlooked. While most current embedding‐based methods rely on extensive supervised data for model training, labeled data remains scarce in the geographic domain. Although semi‐supervised models can partially address this limitation, they still struggle with the challenge of achieving satisfactory alignment accuracy for candidates. To address these challenges, this paper proposes a semi‐supervised cross‐domain GeoKG entity alignment framework based on a dual‐strategy iteration (named DSGEA). DSGEA constructs entity embeddings by generating context‐specific relationship embeddings and then splitting and aggregating neighborhood information according to relationship categories, retaining the unique characteristics across different categories. Subsequently, the framework optimizes the search criteria for entity pairs through a dual‐strategy iteration, leveraging entity distance and entity similarity as search criteria. Additionally, it employs seed entity pairs neighborhood propagation to identify alternate entities within a limited range, enhancing the accuracy of entity identification. Experiments on the self‐constructed dataset GeoLinkSet and open‐source datasets demonstrate that DSGEA can effectively align geographic entities and exhibits superior generalization ability.