Biomedical relation extraction with knowledge base-refined weak supervision.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Machine Learning

Background:

• Biomedical relation extraction (BioRE) is crucial for understanding biomedical literature.
• Supervised learning and large language models (LLMs) dominate BioRE, but require extensive human-annotated data.
• Data scarcity and annotation costs limit the performance and coverage of current BioRE systems.

Purpose of the Study:

• To develop a BioRE system for the BioCreative VII challenge-DrugProt track.
• To address the limitations of human-annotated data in BioRE by leveraging weak supervision.
• To improve the performance of BioRE systems through a novel weak supervision strategy.

Main Methods:

• Developed a BioRE system combining a language model structure with weak supervision.
• Created a weakly labeled dataset by training a model on existing data to predict labels on external literature (model-labeled dataset).
• Refined the model-labeled dataset using an external knowledge base before fine-tuning with human-labeled data.

Main Results:

• The approach using refined weak supervision demonstrated a significant performance gain compared to models trained solely on human-labeled data.
• The developed system achieved 3rd place in the BioCreative VII challenge-DrugProt track.
• Weak supervision, when refined with external knowledge, proves effective in enhancing BioRE performance.

Conclusions:

• Refined weak supervision is a viable and effective strategy to overcome data limitations in biomedical relation extraction.
• The proposed system offers a promising direction for developing high-performing BioRE tools.
• This work highlights the potential of combining model-generated labels with external knowledge bases for BioRE.