RL-based sample selection improves transfer learning in low-resource and imbalanced clinical settings

ACL ARR 2025 May Submission5084 Authors

Abstract: A common strategy in transfer learning is few shot fine-tuning, but its success is highly dependent on the quality of samples selected as training examples. Although active learning methods like uncertainty sampling and diversity sampling can pick useful samples, they underperform in low-resource and class-imbalanced conditions. We introduce a more robust sample selection strategy using reinforcement learning (RL) to identify the most informative samples. Combined with back-translation data augmentation, this approach greatly improved model adaptability in low-resource and class-imbalanced settings. Experimental evaluations on two clinical datasets related to invasive fungal infection (IFI) show our RL-based sample selection strategy enhances model transferability and still maintains robust performance under extreme class imbalance compared to traditional methods. An ablation study on data augmentation reveals that this approach can greatly enhance performance when only a few samples are available, but as sample size grows, the quality of back-translation is also crucial for the model's performance.

Paper Type: Long

Research Area: Efficient/Low-Resource Methods for NLP

Research Area Keywords: Transfer Learning, Low Resource, Class Imbalance, Reinforcement Learning, Back Translation

Contribution Types: Approaches to low-resource settings

Languages Studied: English

Submission Number: 5084