A validation-driven training controller for cross-lingual biomedical NER via reinforcement learning-based adaptive...

Area of Science:

• Computational linguistics
• Bioinformatics
• Natural Language Processing

Background:

• Cross-lingual biomedical named entity recognition (BioNER) faces challenges in low-resource scenarios.
• Scarce annotations, diverse corpora, and imbalanced labels (dominated by the 'O' class) hinder performance.
• Existing methods often require backbone modifications or struggle with instability.

Purpose of the Study:

• To improve the robustness of BioNER fine-tuning in low-resource settings.
• To develop a method that enhances performance without altering the underlying model architecture.
• To address label imbalance and distribution shift issues in BioNER.

Main Methods:

• Propose RouteB-v4, a validation-driven training controller.
• Adaptively reweights token-level loss using entity-level feedback from validation data.
• Employs conservative, bounded updates to loss weights and regulates the 'O' label weight to ensure stability.
• Designed as a drop-in module for standard token-classification pipelines.

Main Results:

• RouteB-v4 demonstrates stable span-level F1 improvements over various baselines in English-Spanish transfer settings and Spanish datasets.
• Achieves clearer gains on minority entity types and enhances external generalization.
• Reaches 0.868 F1 on PharmaCoNER and improves CANTEMIST generalization using XLM-R within a fixed training budget.
• Performance advantage is largely retained even when controller feedback is separated from checkpoint selection.

Conclusions:

• Validation-driven adaptive loss control effectively enhances BioNER robustness.
• The proposed method successfully tackles label imbalance and distribution shift without backbone modification.
• RouteB-v4 offers a practical solution for improving cross-lingual BioNER in challenging low-resource environments.