AgentMap Unifies Equivalence and Subsumption in Ontology Matching

Ontology matching (OM) traditionally focuses on finding either equivalent concepts or subsumption relationships between different ontologies. Existing systems typically handle only one type of semantic correspondence, limiting their utility in complex integration scenarios. This research proposes a new task called Hybrid Ontology Matching (HOM), which aims to discover both equivalence and subsumption mappings simultaneously. To address HOM, the paper introduces AgentMap, a multi-agent framework built upon Large Language Models (LLMs). AgentMap employs a series of interdependent semantic decisions, integrating semantic retrieval, hierarchical search, and collaborative LLM reasoning. For any given concept in a source ontology, AgentMap systematically explores the target ontology to identify either an equivalent concept or the most fine-grained subsumer. Evaluations on extended OM datasets demonstrate that AgentMap achieves promising performance in the hybrid setting. Furthermore, it outperforms existing baselines in both equivalence-only and subsumption-only scenarios, showcasing its versatility and effectiveness in unifying different types of semantic correspondence discovery.