JTIS: enhancing biomedical document-level relation extraction through joint training with intermediate steps

Area of Science:

• Biomedical Natural Language Processing
• Bioinformatics
• Computational Biology

Background:

• Biomedical Relation Extraction (RE) is vital for understanding biological information but traditionally focuses on single sentences.
• Increasing literature volume necessitates multi-sentence analysis for complex biomedical relationships.
• Existing methods struggle with semantic structures, impacting extraction accuracy.

Purpose of the Study:

• To develop a novel method for biomedical RE that analyzes titles and abstracts beyond single sentences.
• To accurately classify novel findings among extracted entity relationships.
• To improve the overall quality and accuracy of biomedical relation extraction.

Main Methods:

• Utilized a multitask training approach to fine-tune a Pre-trained Language Model specialized for biology.
• Designed a broad spectrum of tasks to provide more effective supervision for relation extraction and novel finding classification.
• Employed a model ensemble technique to further enhance performance.

Main Results:

• Achieved significant performance improvements on the BioRED dataset.
• Surpassed existing baselines by 3.94% in Relation Extraction F1 score.
• Improved Triplet Novel Typing by 3.27% in F1 score.

Conclusions:

• The proposed multitask learning method effectively handles complex biomedical literature RE tasks.
• The approach enhances both the quality of relation extraction and the accuracy of novel finding classification.
• This work offers a more robust solution for extracting and classifying biomedical relationships from scientific texts.