Fine-Tuning Bert Untuk Named Entity Recognition Pada Dokumen Klinis Gizi

Abstract

This study focuses on the development of an information extraction system for clinical nutrition documents using a Named Entity Recognition (NER) approach based on Bidirectional Encoder Representations from Transformers (BERT). The system is designed to identify and classify important entities in clinical nutrition documents, including health conditions, food ingredients, nutritional components, and other supporting information. This approach aims to transform unstructured textual data into structured information that can be further utilized. Several Indonesian-language BERT models were evaluated to determine the most suitable model for the clinical nutrition domain. The evaluation process involved performance comparison across models, followed by fine-tuning of the selected model to improve its performance in the Named Entity Recognition task. Model performance was assessed using precision, recall, and F1-score metrics to measure the accuracy of entity recognition. The results indicate that BERT-based models provide stable and effective performance for NER tasks in clinical nutrition texts, particularly for entities with larger data distributions. However, entities with limited data remain challenging, especially in terms of precision. These findings demonstrate the strong potential of BERT-based approaches for processing clinical nutrition documents, while also highlighting opportunities for further improvement through fine-tuning optimization and data imbalance handling.