Hybrid LLM and Embedding Ontology Matching with Reference Ontologies

Abstract

Ontology alignment is critical for semantic interoperability, yet it remains difficult due to semantic ambiguity, poor axiomization, and scalability constraints. While newer Large Language Model (LLM)-based techniques increase semantic comprehension, they frequently rely on repeated model invocations, use stochastic alignment decisions, and are primarily concerned with class-level matching. This study provides an ontology matching system that separates semantic comprehension and alignment judgments. Each element is processed independently, using a single LLM call for semantic typing, reducing stochasticity. This typing guides the selection of a reference ontology (e.g., DBpedia, Friend Of A friend (FOAF), SKOS) as a semantic mediator. Alignment is performed using embedding-based similarity with optional OWL validation. The approach supports both Terminological Box (TBox) and RBox alignment, addressing Role Box (RBox) matching while improving scalability, interpretability, and reliability. It also ties to Retrieval-Augmented Generation (RAG) frameworks by using external ontologies as structured knowledge sources, while avoiding generative alignment decisions.