ICB-UMA/HERBERT-GP-30

HERBERT: Leveraging UMLS Hierarchical Knowledge to Enhance Clinical Entity Normalization in Spanish

HERBERT-GP is a contrastive-learning-based bi-encoder for medical entity normalization in Spanish.
It leverages hierarchical relationships from UMLS (parents and grandparents) to enhance the candidate retrieval step for entity linking in Spanish clinical texts.

Key features:

• Base model: PlanTL-GOB-ES/roberta-base-biomedical-clinical-es
• Trained with 30 positive pairs per anchor using synonyms, parents, and grandparents from UMLS/SNOMED-CT.
• Task: Normalization of disease, procedure, and symptom mentions to SNOMED-CT/UMLS codes.
• Domain: Spanish biomedical/clinical texts.
• Corpora: DisTEMIST, MedProcNER, SympTEMIST.

---

Evaluation (top-k accuracy):

Corpus	Top-1	Top-5	Top-25	Top-200
DisTEMIST	0.585	0.727	0.808	0.871
SympTEMIST	0.632	0.783	0.884	0.948
MedProcNER	0.655	0.770	0.840	0.891