Extracting biomedical relation from cross-sentence text using syntactic dependency graph attention network.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Graph Neural Networks

Background:

• Extracting relations from biomedical texts is crucial for research.
• Cross-sentence text presents challenges due to longer sequences and complex dependencies.
• Existing methods struggle with effectively handling global dependencies and structural information in long sequences.

Purpose of the Study:

• To propose a novel graph attention network guided by syntactic dependency relationship (SR-GAT) for biomedical relation extraction.
• To address the challenge of learning context dependency structural information in cross-sentence texts.
• To improve the efficiency and accuracy of biomedical relation extraction.

Main Methods:

• Developed a Syntactic Relation Graph Attention Network (SR-GAT) incorporating a Syntactic Relation Graph Probability Network (SR-GPR).
• SR-GPR encodes syntactic dependencies to guide the graph attention mechanism.
• The model learns node-to-node syntactic dependencies to discover global dependencies.

Main Results:

• Achieved state-of-the-art performance on a public biomedical dataset with significantly less computational resources.
• Demonstrated high accuracy in "drug-mutation" (93.78% binary, 92.14% multi-class) and "drug-gene-mutation" (93.22% binary, 92.28% multi-class) relation extraction.
• Showcased robust generalization with 69.5% accuracy in text classification without fine-tuning.

Conclusions:

• The proposed SR-GAT method effectively extracts biomedical relations from cross-sentence texts by leveraging syntactic dependency information.
• The model offers improved accuracy and efficiency compared to existing methods.
• SR-GAT demonstrates strong performance and generalization capabilities across different biomedical NLP tasks.