Subsequence and distant supervision based active learning for relation extraction of Chinese medical texts.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Relation extraction from unstructured medical texts is crucial but requires extensive manual annotation.
• Existing methods face challenges due to the time-consuming and costly nature of creating labeled datasets.
• Efficient and economical annotation strategies are needed to advance medical relation extraction.

Purpose of the Study:

• To propose an efficient and economical method for annotating sequences for medical relation extraction.
• To reduce the burden of manual data labeling in medical text analysis.
• To improve the performance of relation extraction models in the medical domain.

Main Methods:

• A novel active learning approach utilizing information-rich subsequences as sampling units.
• Integration of distant supervision for pre-labeling unlabeled data via text matching.
• Development of a continuously updated dictionary to store labeled subsequences and their labels.
• Application of a Chinese-RoBERTa-CRF model for relation extraction on Chinese medical texts.

Main Results:

• The proposed method achieved the best performance on the CMeIE dataset compared to existing approaches.
• The best F1 score obtained using different sampling strategies reached 55.96%.
• Demonstrated the effectiveness of subsequence-based active learning and distant supervision in medical relation extraction.

Conclusions:

• The subsequence and distant supervision-based active learning method offers an efficient and economical solution for medical relation extraction.
• This approach significantly reduces manual annotation efforts while maintaining high performance.
• The study highlights the potential of integrating advanced NLP models with novel learning strategies for medical text analysis.