Offline Metathesaurus Translation for Cross-Lingual Medical Entity Linking
(2026) ECML PKDD — Location: Naples, Italy (7.September.2026)
Files
Paper.pdf
Author postprint
Embargo Access until 2027-10-04 - Adobe PDF
- 544.14 KB
Details
- Authors
- Abstract
- Cross-lingual medical entity linking (MEL) is essential for extracting structured information from clinical text. Yet, progress remains largely English-centric because biomedical terminologies and annotated corpora are far more abundant in English. To address this resource gap, we investigate a translation-based supervised MEL setting for SNOMED CT code retrieval. English UMLS concepts are translated of-fline to construct positive synonym pairs used to train SapBERT-style bi-encoders, without requiring online translation, reranking, or ensembling at inference time. We evaluate the approach on French, Spanish, and German using public benchmarks, and quantify how the choice of machine translation model affects downstream entity linking performance. Rank-1 metrics (F1@1 and P @1) are reported for each benchmark, as well as top-100 candidate recall (R@100) to disentangle candidate coverage from ranking errors. We further introduce an ontology-aware metric, MGD@100, which summarises how far retrieved candidates lie from the ground-truth concept in the SNOMED CT graph, enabling the analysis of semantically close concepts when the exact concept is not retrieved. In macro-averaged comparisons across languages, the Helsinki-NLP OPUS-MT translation models provide the best average candidate coverage (mean R@100 = 0.880) while maintaining competitive ontol-ogy proximity (mean MGD@100 = 7.016, where lower values indicate that candidates are closer to the ground-truth concept). Compared with prior systems, the proposed retrieval-only approach remains competitive across languages despite the absence of reranking or ensembling, establishes new best results on all Spanish benchmarks and on several French and German ones. It is also shown that biomedical encoder backbones yield consistent but modest gains on rank-1 metrics over general-purpose encoders. Finally, correlation and information-theoretic analyses show that MGD@100 is not a surrogate for standard quality metrics, but a complementary ontology-aware evaluation signal.
- Affiliations
Citations
Fierens, A., Cousin, T., Roucoux, F., & Jodogne, S. (2026). Offline Metathesaurus Translation for Cross-Lingual Medical Entity Linking. Lecture Notes in Computer Science, 16949, 582-599. https://doi.org/10.1007/978-3-032-37682-4_34 (Original work published 2026)