Distantly Supervised Biomedical Relation Extraction via Negative Learning and Noisy Student Self-Training.

Area of Science:

• Bioinformatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical relation extraction identifies relationships (e.g., gene associations, drug interactions) in texts.
• Scarcity of labeled training data hinders biomedical relation extraction performance.
• Existing methods struggle with noisy and insufficient data.

Purpose of the Study:

• To develop a novel approach for biomedical relation extraction addressing data insufficiency.
• To leverage noisy student self-training and negative learning for improved performance.
• To enhance noise reduction and relabeling capabilities in relation extraction models.

Main Methods:

• Utilized a noisy student self-training strategy combined with negative learning.
• Employed distantly supervised data to generate high-quality labeled samples.
• Implemented negative learning for robust discernment and relabeling of noisy samples to prevent overfitting.

Main Results:

• The proposed framework effectively mitigates the impact of noisy data.
• Demonstrated superior performance compared to existing benchmarks on noisy datasets.
• Achieved enhanced noise reduction and relabeling capabilities.

Conclusions:

• The novel framework significantly improves biomedical relation extraction with limited and noisy data.
• Noisy student self-training and negative learning are effective strategies for this task.
• The approach offers a robust solution for data scarcity challenges in biomedical NLP.