A metric learning-based method for biomedical entity linking.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Machine Learning

Background:

• Biomedical entity linking connects text mentions to knowledge base concepts (e.g., UMLS).
• Challenges include mention ambiguity and diverse entity expressions.
• Existing BERT-based methods struggle with imbalanced biomedical datasets, where down-sampling hinders learning.
• Down-sampling reduces the model's ability to learn contextual mention representations.

Purpose of the Study:

• To propose a metric-based learning method for biomedical entity linking that addresses data imbalance.
• To develop a technique that treats entities and their mentions holistically, irrespective of mention frequency.
• To improve the learning of mention and entity representations in imbalanced datasets.

Main Methods:

• A metric-based learning approach using triplet loss.
• Integration of a clustering technique to learn representations.
• Treating entities and their mentions as a unified whole during training.
• Evaluation on MedMentions and BC5CDR datasets.

Main Results:

• Successfully addresses the challenge of imbalanced data in biomedical entity linking.
• Achieves competitive performance compared to state-of-the-art models.
• Significantly reduces computational costs during training and inference.
• Demonstrates effective learning of mention and entity representations.

Conclusions:

• The proposed metric-based learning method offers an effective solution for imbalanced biomedical entity linking datasets.
• This approach enhances model performance and efficiency.
• The method provides a robust alternative to traditional down-sampling techniques.