BERT-based Ranking for Biomedical Entity Normalization.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Biomedical entity normalization is crucial for managing terminological variations.
• Existing deep learning methods often rely on context-independent word embeddings.
• Recent advancements include contextualized word representations from models like BERT, BioBERT, and ClinicalBERT.

Purpose of the Study:

• To propose and evaluate an entity normalization architecture using fine-tuned BERT, BioBERT, and ClinicalBERT models.
• To assess the effectiveness of pre-trained contextualized models for biomedical entity normalization.
• To advance the state-of-the-art in handling term variations in biomedical text.

Main Methods:

• Fine-tuning pre-trained Bidirectional Encoder Representations from Transformers (BERT), BERT for Biomedical Text Mining (BioBERT), and BERT for Clinical Text Mining (ClinicalBERT) models.
• Developing a novel entity normalization architecture.
• Conducting extensive experiments on three diverse biomedical datasets.

Main Results:

• The proposed fine-tuned models consistently outperformed previous state-of-the-art methods.
• Accuracy improvements of up to 1.17% were achieved in biomedical entity normalization.
• Demonstrated the superior effectiveness of contextualized embeddings over traditional methods.

Conclusions:

• Fine-tuning pre-trained BERT-based models offers a significant advancement for biomedical entity normalization.
• The developed architecture effectively addresses the term variation problem in biomedical text.
• This approach sets a new benchmark for high-performance biomedical entity normalization.