Learning to explain is a good biomedical few-shot learner.

Area of Science:

• Biomedical Natural Language Processing (NLP)
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning for biomedical text mining requires extensive expert-annotated data, which is often scarce, costly, or privacy-sensitive.
• Existing methods primarily focus on prediction without providing explanations, limiting their real-world applicability.
• Biomedical few-shot learning presents a realistic challenge due to data limitations, necessitating methods that can learn effectively from minimal data.

Purpose of the Study:

• To investigate the impact of interpretability on biomedical few-shot learning.
• To develop a novel approach that enhances inductive reasoning in low-resource biomedical NLP tasks.
• To address the challenge of data scarcity by leveraging explanations from large language models (LLMs).

Main Methods:

• Introduced LetEx-Learning, a multi-task generative approach utilizing LLM-generated reasoning explanations.
• Developed a workflow using Chain-of-Thought (CoT) prompting and self-training to collect high-quality explanations.
• Unified diverse biomedical NLP tasks into a text-to-text generation framework, using explanations as additional supervision via multi-task training.

Main Results:

• Learning to explain significantly improved performance across various biomedical NLP tasks in few-shot settings.
• The proposed method outperformed strong baseline models by up to 6.41% in low-resource scenarios.
• The 220M LetEx model demonstrated superior reasoning explanation capabilities compared to larger LLMs.

Conclusions:

• Interpretability, specifically through learning to explain, is crucial for advancing biomedical few-shot learning.
• The LetEx-Learning approach offers a promising solution for data-scarce biomedical NLP challenges.
• Leveraging LLM-generated explanations can enhance model performance and reasoning abilities in specialized domains.