The authors assembled a modular biomedical named-entity normalization workflow that layers several retrieval and selection techniques. Exact matching sits at the front of the pipeline. When exact matches fail, the system falls back to lexical methods — including fuzzy matching and character n-gram similarity — and to semantic retrieval powered by SapBERT. The information-retrieval backbone incorporates BM25 for ranked lexical retrieval alongside these complementary approaches. A reranking stage follows candidate retrieval to select the final ontology term. The team emphasized that the workflow operates without requiring ontology-specific training, keeping the pipeline broadly applicable across different biological ontologies.
Investigators evaluated the workflow across four distinct ontologies: Human Phenotype Ontology (HPO), Cell Ontology (CL), Uberon, and Gene Ontology (GO). They used multiple benchmark datasets to measure performance at three separable points: candidate retrieval recall, reranking accuracy, and overall normalization accuracy. The study team reported variability in retrieval performance across the ontologies and across different benchmarks, indicating that no single retrieval strategy dominated in every setting. The authors made their code and supporting materials available in a public repository (footnote link), allowing reproducibility and additional testing.
The workflow analysis explored how the number of retrieved candidates influences downstream outcomes. Increasing the candidate-set size improved retrieval recall, as expected; more candidates raised the chance the correct ontology entry was present among them. However, the authors observed that larger candidate pools did not always translate into higher end-to-end normalization accuracy. In some configurations, expanding the candidate list degraded reranker performance, reducing final selection accuracy. Thus, the data indicated a trade-off: broader candidate coverage versus effective selection by reranking. The study reported these dynamics without prescribing a single optimal candidate-set size, underscoring that the balance may depend on ontology and benchmark characteristics.
Data indicated that retrieval difficulty was associated with both mention characteristics and ontology categories. Short, ambiguous, or variably phrased mentions tended to be harder to retrieve reliably. Certain ontology branches and term naming conventions also created systematic challenges for the retrieval component. The authors separated retrieval performance from reranking performance to pinpoint where failures originate — whether the correct candidate never entered the pool, or whether the reranker failed to choose it. This diagnostic approach allowed more granular interpretation of error modes across HPO, CL, Uberon, and GO.
The team compared their modular workflow with CellExLink, a specialized system focused on cell-type normalization. The abstract reports that such a comparison was performed; it also indicates that retrieval methods (lexical and semantic) provided complementary candidate coverage in some settings. The source did not include full quantitative comparative metrics in the abstract. Readers are referred to the repository and full preprint for specific numerical results and benchmark-by-benchmark comparisons.
The authors identified several performance characteristics and limitations of applying a common normalization workflow across heterogeneous ontologies. First, retrieval performance varied by ontology and benchmark; therefore a one-size-fits-all configuration may not be optimal. Second, larger candidate sets improved recall but could harm reranking, indicating a need for careful tuning of candidate-size parameters or for improved reranking approaches. Third, mention phrasing and ontology naming conventions remain important determinants of difficulty. The abstract states the workflow provides a methodological foundation for extracting and integrating biological information from rare-disease literature and suggests potential applicability to other biomedical information-extraction tasks that require normalization to these ontologies. The source did not report detailed methodology parameters or full comparative metrics in the abstract; those specifics must be obtained from the full preprint and linked code repository.
Acknowledged open questions include which candidate-set sizes and reranking strategies generalize best across ontologies, how to mitigate mention-level ambiguity systematically, and how the workflow performs on additional, unseen benchmarks. The authors have published their code to permit community-driven testing and further development. The preprint has not been peer reviewed, and full evaluation details are provided in the main manuscript and supplementary materials referenced in the repository.