An open-set semi-supervised multi-task learning framework for context classification in biomedical texts.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Biomedical relation extraction (RE) is crucial for understanding biological processes.
• Current NLP methods for RE often lack essential contextual information like cell type and location.
• Previous approaches to context-relation association are limited by data scarcity and error propagation.

Purpose of the Study:

• To propose CELESTA (Context Extraction through LEarning with Semi-supervised multi-Task Architecture), an open-set semi-supervised multi-task learning (OSSL-MTL) framework.
• To improve biomedical context classification accuracy and extract implicit contextual information.
• To address limitations in existing RE methods by integrating context extraction directly into the learning framework.

Main Methods:

• Developed an MTL architecture combined with SSL strategies to utilize unlabeled data (ID and OOD).
• Created a large-scale dataset for five context classification tasks using BEL corpora and a new annotation method.
• Implemented an OOD detector to differentiate ID and OOD instances and used data augmentation.

Main Results:

• CELESTA significantly improved context classification performance across tasks.
• Achieved high F1 scores: 77.75% for location and 82.87% for disease classification.
• Demonstrated effective OOD detection and improved extraction of implicit contexts compared to baselines.

Conclusions:

• CELESTA framework effectively enhances context classification and information extraction in biomedical research.
• The developed framework and dataset contribute to advancing NLP applications in biology.
• Publicly available code and dataset facilitate further research and development.