Demonstration-based learning for few-shot biomedical named entity recognition under machine reading comprehension.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Deep learning models often require extensive labeled data for effective biomedical named entity recognition (BioNER).
• Few-shot learning scenarios, where labeled data is scarce, pose a significant challenge for current deep learning models in BioNER.
• Existing methods struggle to achieve high accuracy when training data is limited.

Purpose of the Study:

• To develop an effective strategy for improving biomedical entity recognition in few-shot learning scenarios.
• To enhance the performance of models in recognizing biomedical entities with limited annotated data.
• To address the limitations of data-hungry deep learning models in specialized domains like BioNER.

Main Methods:

• Redefined biomedical named entity recognition (BioNER) as a machine reading comprehension (MRC) problem.
• Proposed a demonstration-based learning method utilizing task demonstrations to address few-shot BioNER.
• Evaluated the proposed method against advanced techniques on six benchmark datasets (BC4CHEMD, BC5CDR-Chemical, BC5CDR-Disease, NCBI-Disease, BC2GM, JNLPBA).

Main Results:

• Achieved average F1 score improvements of 1.1% in 25-shot learning and 1.0% in 50-shot learning compared to baseline methods.
• Demonstrated strong performance across six diverse biomedical datasets, with specific F1 scores reported for each.
• Showcased the efficacy of the MRC-based approach in few-shot BioNER tasks.

Conclusions:

• MRC-based language models significantly outperform sequence labeling approaches for few-shot BioNER.
• The proposed MRC models demonstrate competitive performance against fully-supervised methods, even with limited annotated data.
• This research offers promising avenues for advancing few-shot BioNER methodologies, reducing reliance on large labeled datasets.