Iterative difference-based distance matrix refinement for entity alignment

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Cross-lingual entity alignment is a key task in natural language processing, aiming to identify and align entities that refer to the same real-world objects across knowledge graphs in different languages. While some approaches have been proposed, most rely heavily on labeled data and are sensitive to noise caused by structural inconsistencies in entity neighborhoods. To address these challenges, we propose a cross-lingual entity alignment framework through i terative d ifference-based distance matrix r efinement (IDR). Our model automatically generates initial alignment seeds from entity name similarity, eliminating the need for manually labeled seed alignments. Unlike traditional methods that rely solely on embedding similarity, our approach introduces a difference-based refinement strategy: entity alignment is performed by refining the entity distance matrix with relation-aware neighborhood matching scores, while relation alignment is achieved by refining the relation distance matrix with entity-aware relation matching scores. These two processes are integrated into a unified iterative optimization framework in which entity alignment and relation alignment mutually reinforce one another. Extensive experiments are conducted on widely used cross-lingual knowledge graph datasets. The results demonstrate that our method achieves superior performance, with Hits@1 scores of 91.6% and 94.1% on the Chinese-English and Japanese-English subsets of the DBP 15 K dataset, outperforming the best baseline models by 3.3% and 5.4%, respectively. Meanwhile, on the French-English subset, our method obtains 95.8% Hits@1 and maintains highly competitive performance.