Biomedical ontology normalization maps free‑text expressions to standardized concepts, enabling consistent integration and analysis of biomedical data. Lexical variation and subtle distinctions among hierarchically related concepts often blur concept boundaries. To address this, we introduce OntologyAligner, a three‑stage framework:
- Ontology‑aligned retrieval: constrains the search space using ontology structure to quickly generate a candidate set.
- Large language model (LLM) reranking: feeds the candidate set into an LLM, which scores and reorders candidates based on contextual semantics.
- Hierarchy‑guided selective refinement: performs a second‑level check on the ontology hierarchy, applying hierarchical constraints only to high‑confidence candidates.
We also build PhenoNormBench, a unified benchmark that aggregates seven Human Phenotype Ontology (HPO) datasets into 13,390 samples. OntologyAligner achieves state‑of‑the‑art results on HPO normalization with 88.78% Macro Top‑1 and 86.75% Micro Top‑1 accuracy, surpassing the strongest baseline by 4.85 and 5.07 percentage points respectively. Ablation studies confirm that all three stages contribute complementary gains, while sensitivity analyses show stability across different candidate‑set sizes and model backbones. Applying the same pipeline to MONDO, MEDIC, and NCBITaxon further demonstrates its portability.
The code and benchmark are publicly available at https://github.com/zhelishisongjie/OntologyAligner.
Review