AgentMap: Joint Equivalence and Subsumption Discove... | AI Research

AgentMap is a multi-agent framework designed to solve Hybrid Ontology Matching (HOM), a task that requires identifying both equivalence and subsumption relationships between concepts in different ontologies. While traditional systems typically focus on only one of these relationships, AgentMap uses a series of interdependent, LLM-based decisions to determine if a concept has an exact equivalent or, if not, the most specific subsumer in the target ontology.

The Hybrid Ontology Matching Task

Ontology matching is essential for integrating knowledge across different domains, such as healthcare or food science. However, existing systems often struggle because they treat equivalence and subsumption as separate problems. HOM unifies these by requiring a system to first prioritize finding an equivalent concept. If no equivalent exists, the system must identify the most specific subsumer—a concept that is more general but remains the closest possible match in the hierarchy.

How AgentMap Works

AgentMap functions through three primary modules that orchestrate the matching process:

• Data Preprocessing and Candidate Retrieval: The system encodes concepts into semantic embeddings and uses cosine similarity to generate two sets of candidates. A smaller set is reserved for intensive LLM reasoning, while a larger set is used for lexical matching.

• Agent-Based Reasoning: This module employs three specialized agents. An equivalence screening agent checks for exact matches, a verification agent confirms these matches by comparing them against parent and child concepts, and a subsumption discovery agent performs an iterative, hierarchy-aware search if no equivalent is found.

• Lexical Matching and Conflict Resolution: This module runs in parallel to identify matches based on labels and synonyms. If the lexical and agent-based results conflict, an LLM-based resolution step reconciles the findings to produce a final, unified mapping.

Performance and Evaluation

The researchers evaluated AgentMap using a new benchmark constructed from four existing medical and food-related datasets. By reorganizing these datasets to include both equivalence and subsumption ground truths, they created a controlled environment to test the framework.
Experimental results indicate that AgentMap consistently outperforms baseline LLM strategies in the hybrid setting. Notably, it achieved superior subsumption accuracy across all four datasets, with improvements in some cases reaching 30.8% over the best baseline. The framework also performed effectively when tested in isolation, outperforming existing equivalence-only and subsumption-only systems.

Considerations for Implementation

The effectiveness of AgentMap relies on its ability to decompose complex matching tasks into smaller, manageable steps. By using an "equivalence-first" strategy and incorporating the target ontology's structure into the search process, the framework avoids the limitations of single-step reasoning. The researchers note that the system's performance is consistent across different datasets, though it is designed to work within a specific candidate-based search range, with a maximum upward traversal limit (d_max) set to two levels in the hierarchy to maintain efficiency.