An multi-embedding entity alignment approach based on matching information interaction over atomically divisible entities

Entity alignment, as a fundamental technique in knowledge graph integration, aims to identify the equivalent entity pairs that refer to the same real-world entity across different knowledge graphs. In domain-specific knowledge graphs, based on the atomic divisibility characteristics of entity categories, these categories can be classified into atomically divisible and atomically indivisible types. However, traditional methods for entity alignment in atomically indivisible knowledge graphs often fall short when dealing with tasks that involve more than one atomically divisible entity category. To address such a challenge, we propose a novel multi-embedding entity alignment approach based on matching information interaction over atomically divisible entities (MiAD). MiAD effectively exploits the matching information interaction between low-level entities and high-level entities to enrich the semantic representation of high-level entities and to enhance the reliability of the alignment results. The high-level entities are atomically divisible, and the matching result among the low-level entities can provide a significant hint for the entity alignment of high-level entities by analyzing their inherent interrelationships. To tackle the insufficient utilization of these matching information interactions, we propose a multiple hierarchical embedding strategy, which consists of a structural semantic aggregator, an attribute semantic aggregator, a corroborative aggregator, and a joint aggregator. This method leverages both the intra-level interactions and cross-level interactions to generate the hierarchical embeddings of high-level and low-level entities, respectively, from multiple perspectives (i.e., relations, attributes, and corroborative textural descriptions). As a result, MiAD significantly improves the accuracy of alignment outcomes and strengthens the overall structural and semantic understanding of higher-level entities. Comprehensive experimental results demonstrate that MiAD achieves substantial performance improvements compared with seven baseline methods on the Chinese recipe dataset Ta-da.