Information Extraction from Medical Texts with BERT Using Human-in-the-Loop Labeling.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Neural network language models like BERT excel at information extraction from unstructured medical text.
• Pre-training on large corpora allows models to learn domain-specific language.
• Fine-tuning with labeled data enables task-specific performance.

Purpose of the Study:

• To propose a human-in-the-loop labeling pipeline for creating annotated data.
• To facilitate Estonian healthcare information extraction.
• To provide an accessible method for low-resource languages.

Main Methods:

• Utilizing BERT, a powerful neural network language model.
• Implementing a human-in-the-loop labeling strategy for data annotation.
• Focusing on Estonian healthcare information extraction.

Main Results:

• The proposed pipeline effectively generates annotated data for Estonian medical texts.
• Human-in-the-loop labeling proves more accessible than rule-based methods for medical professionals.
• The approach is suitable for low-resource language scenarios.

Conclusions:

• The human-in-the-loop pipeline is a viable solution for Estonian healthcare information extraction.
• This method enhances the utility of neural network models in specialized, low-resource domains.
• It offers a practical alternative to complex rule-based systems in medical informatics.