LaDer: A Two-Stage Unsupervised Method for Stem Cell Entity Recognition Based on Reinforcement Learning

Extracting stem cell entities from biomedical literature is crucial for developing knowledge bases that can inform clinical treatments and drug discovery. Addressing the challenge of limited supervised data and proprietary dictionaries, we introduce LaDer, an unsupervised, two-stage labeling and distillation method. Initially, LaDer leverages a pre-trained language model to identify candidate stem cell entities without fine-tuning. Subsequently, it refines these labels through a secondary detection process modeled as a reinforcement learning problem, utilizing unsupervised signals to optimize labeling precision. This method demonstrates state-of-the-art precision at 94.67% on the stem cell entity labeling task. Integration with BioBERT further enhances our method, achieving a 91.16% F1 score. LaDer has successfully extracted 1134 unique stem cell entities from over 10,000 articles, contributing to a new “stem cells-disease” knowledge base. We make our codes and database publicly available at:https://github.com/Zxy-MLlab/Stem-cell-NER-based-on-RL.