Biomedical relation extraction with knowledge base–refined weak supervision

Biomedical relation extraction (BioRE) is essential for constructing knowledge graphs and understanding drug-protein interactions. However, supervised learning for BioRE is hindered by the high cost and complexity of creating large-scale, human-annotated datasets. To address this, the authors propose a hybrid approach that combines weak supervision with knowledge base (KB) refinement to augment training data.

The system operates in three phases. First, it generates a large-scale augmented dataset by using a model trained on a small, high-quality human-annotated set (DrugProt) to predict relations across a vast corpus of biomedical literature. Second, it filters these predictions using an external knowledge base (the Comparative Toxicogenomics Database, or CTD); only predictions that align with known drug-protein interactions are retained. Finally, a transformer-based sequence classification model is pretrained on this augmented dataset and then fine-tuned on the original human-annotated data.

This method produced a training dataset approximately 13.5 times larger than the original DrugProt dataset. The authors demonstrate that pretraining on this refined, weakly labeled data provides a significant performance boost compared to models trained solely on human-annotated data. The system achieved third place in the BioCreative VII DrugProt challenge, proving that leveraging large-scale, automatically generated, and filtered data is an effective strategy for overcoming the scarcity of high-quality manual annotations in the biomedical domain.

This research provides a scalable framework for BioRE that reduces the reliance on expensive, manual annotation. By demonstrating that noisy, model-generated labels can be effectively cleaned using existing knowledge bases, the authors offer a practical pathway for researchers to build more robust NLP models for specialized domains where data is limited but external knowledge is abundant.