Work

OntoAlign-RAG: Traceable and Uncertainty-Calibrated Ontology Alignment with Selective Prediction

Year: 2026
Type: conference-paper
Abstract: Heterogeneous industrial ontologies often describe overlapping concepts through incompatible labels, hierarchies, and relation vocabularies. This paper presents OntoAlign-RAG, a traceable retrieve-augment-verify framework that combines lexical, sparse, dense, and structural retrieval with ontology-bound evidence, a calibrated relation verifier, conformal abstention, and a deterministic evidence-consistency gate. A target-purged evaluation protocol separates training, calibration, and test roles and supports reproducible result reconstruction. Experiments on 29 public OAEI tasks show that retrieval and relation quality remain dataset dependent, while probability calibration improves consistently. Conformal selection reduces accepted-set risk at the cost of coverage. The evidence gate removes proxy-inconsistent outputs without materially improving correctness. Borderline dense routing reduces computation with negligible loss in relation F1. The results establish an auditable evaluation workflow and quantify the trade-offs among alignment quality, coverage, uncertainty, evidence consistency, and computational cost.
Institution: Shaoguan University
Language: English
Open Access status: hybrid
OpenAlex ID: W7222760211
FWCI:
Cites: 17
Cited by:
Related to: 10