AugAlign: Augmenting Ontology Alignment With LLMs Across the Full Alignment Pipeline

Abstract

Ontology alignment identifies semantic correspondences between concepts in separately developed ontologies. Recent Large Language Model (LLM)-based ontology alignment approaches adopt a retrieve-then-prompt pipeline, in which a sentence encoder first retrieves candidate correspondences from the target ontology, and the LLM judges whether each candidate pair is semantically equivalent. This pipeline faces three problems. First, ontology concepts are not always well-defined, labels can be abbreviated, opaque, or missing contextual descriptions, causing sentence encoders to produce uninformative embeddings and retrieval to miss correct targets. Second, a single dense-encoding pass cannot capture all forms of similarity, such as synonymy or compositional overlap, leaving valid matches outside the candidate set. Third, a single fixed threshold cannot reliably separate correct from incorrect pairs, because retrieval scores and LLM confidence carry complementary but differently distributed evidence. To address these problems, we propose AugAlign, a framework that leverages the LLM beyond its conventional matching role. The central idea is that the same LLM that judges equivalence can also serve as a knowledge-driven preprocessor, rewriting unclear labels and generating contextual descriptions, so that downstream encoders receive meaningful input rather than opaque symbols. Retrieval is further strengthened by combining multiple complementary similarity signals beyond a single encoding pass. A cascaded postprocessing strategy then fuses retrieval and LLM evidence through tiered thresholds that trade off LLM confidence against retrieval evidence, rather than relying on a single rigid cutoff. Experiments show that the framework is competitive with state-of-the-art systems, with the largest gains on ontologies whose concept definitions are incomplete or ambiguous.