Entity relation extraction in the medical domain: based on data augmentation.

Area of Science:

• Medical informatics
• Natural Language Processing
• Machine Learning

Background:

• Medical knowledge graphs require accurate entity relation extraction.
• Medical datasets are often unbalanced, posing challenges for model training.
• Existing methods struggle with the skewed distribution of entity relations in medical texts.

Purpose of the Study:

• To develop and evaluate a novel data augmentation method for medical entity relation extraction.
• To address the challenge of unbalanced datasets in the medical domain.
• To compare the effectiveness of traditional versus reinforcement learning-based data augmentation.

Main Methods:

• A probability-based approach to data augmentation was proposed, guided by class distribution.
• Reinforcement learning was employed to dynamically select optimal data augmentation strategies.
• The methods were applied to the entity relation extraction of the "Pharmacopoeia of the People's Republic of China" using a consistent neural network model.

Main Results:

• Data augmentation improved deep learning model performance, particularly for low-volume relation classes.
• Reinforcement learning-based data augmentation outperformed traditional methods.
• Significant improvements in evaluation metrics were observed across multiple relation classes after applying reinforcement learning-based augmentation.

Conclusions:

• Data augmentation is effective for enhancing deep learning on unbalanced medical datasets.
• Reinforcement learning-based data augmentation offers superior performance compared to traditional approaches.
• The proposed method demonstrates a robust strategy for improving medical entity relation extraction.