Data Augmentation for Few-Shot Biomedical NER Using ChatGPT.

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning

Background:

• Data scarcity is a significant challenge in biomedical Named Entity Recognition (NER).
• Few-shot learning scenarios require effective data augmentation (DA) to improve model generalization and reduce overfitting.
• Existing DA methods may not adequately address the complexities of biomedical text.

Purpose of the Study:

• To propose a novel DA method for biomedical NER tasks.
• To leverage large language models (LLMs) like ChatGPT for high-quality data generation.
• To enhance the performance of NER models in low-data and few-shot settings.

Main Methods:

• Utilized ChatGPT and prompt learning to extract high-quality data for NER.
• Employed transfer learning and efficient decoding strategies for entity recognition.
• Conducted experiments on four public biomedical datasets: BC5CDR, NCBI, BioNLP11EPI, and BioNLP13GE.

Main Results:

• The proposed DA method demonstrated strong stability and entity recognition capabilities in extremely limited data scenarios.
• Achieved average F1 scores of 72.96% (5-shot), 75.05% (20-shot), and 77.42% (50-shot) across the four datasets.
• Showcased significant improvements in model generalization ability in few-shot NER tasks.

Conclusions:

• The novel DA method effectively addresses data scarcity in biomedical NER.
• ChatGPT and prompt learning offer a powerful approach for generating high-quality training data.
• The method shows promise for improving NER model performance in challenging, low-resource biomedical domains.