Few-shot biomedical NER empowered by LLMs-assisted data augmentation and multi-scale feature extraction.

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning
• Computational Biology

Background:

• Named Entity Recognition (NER) is crucial for biomedical text analysis.
• Few-shot learning methods are explored due to limited labeled biomedical data.
• Existing few-shot NER methods struggle to match fully supervised performance and face data augmentation challenges.

Purpose of the Study:

• To enhance few-shot biomedical Named Entity Recognition (NER).
• To address limitations in current data augmentation techniques that may distort semantics.
• To improve feature representation by capturing multi-scale sentence information.

Main Methods:

• Utilized ChatGPT for generating semantically enriched data to augment existing datasets.
• Employed dynamic convolution to capture multi-scale semantic information within sentences.
• Enhanced feature representation using PubMedBERT as a base model.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art models in most few-shot scenarios across four biomedical NER datasets.
• Performance gains were observed even against large language models like ChatGPT.
• The approach effectively improved data augmentation and model generalization capabilities.

Conclusions:

• The novel method significantly advances few-shot biomedical NER by improving data augmentation and feature extraction.
• The combination of ChatGPT-driven data enrichment and dynamic convolution offers a robust solution for low-resource NER tasks.
• This research provides a promising direction for enhancing the analysis of biomedical literature.