ICB-UMA at BioCreative VIII @ AMIA 2023 Task 2 SYMPTEMIST (Symptom TExt Mining Shared Task)

Abstract

These working notes summarize the contribution of the ICB research group from the University of Malaga to the BioCreative VIII Workshop @AMIA 2023, from our participation in Task 2 - SympTEMIST. Engaged in both subtasks, our approaches tackled symptom, sign, and clinical finding entities recognition (subtask 1 - SymptomNER) and their normalization to the corresponding SNOMED CT concepts (subtask 2 - SymptomNorm). For subtask 1, we analyzed the performance of some BERT-based models tailored for the nuances of Spanish clinical data. These models, specifically fine-tuned on the SymptomNER corpus, showed remarkable precision (0.804), recall (0.699), and F1-score (0.748) for the test set. For SymtomNorm subtask, we incorporated recent strategies using bi-encoder and cross-encoder models, especially SapBERT models enhanced with FAISS methods for similarity search. Finally, the model's predictions were further refined by leveraging a gazetteer with more than 150,000 concepts. Our strategy achieved 0.58 accuracy for the test set.

This article is part of the Proceedings of the BioCreative VIII Challenge and Workshop: Curation and Evaluation in the era of Generative Models.