Document-level biomedical relation extraction via hierarchical tree graph and relation segmentation module.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical relation extraction at the document level (Bio-DocRE) is crucial for understanding complex biological texts.
• Current methods often overlook the significance of entity pair information in relation prediction.
• Existing approaches primarily rely on graphs or transformers, modeling entity features directly.

Purpose of the Study:

• To propose an innovative framework, HTGRS, to enhance Bio-DocRE.
• To improve relation prediction by incorporating entity pair information as an intermediate state.
• To decouple the Bio-DocRE task into a three-stage process for better information capture.

Main Methods:

• Developed the Hierarchical Tree Graph (HTG) to integrate document information for entity-based relation reasoning.
• Conceptualized Bio-DocRE as a table-filling problem, inspired by semantic segmentation.
• Introduced a Relation Segmentation (RS) module to refine relation reasoning using entity pair information.

Main Results:

• The proposed HTGRS framework demonstrated superior performance compared to state-of-the-art methods.
• Extensive experiments on three benchmark datasets validated the effectiveness of the approach.
• The framework achieved significant improvements in biomedical relation extraction accuracy.

Conclusions:

• The HTGRS framework offers a novel and effective approach to Bio-DocRE.
• Integrating entity pair information and a hierarchical graph structure enhances relation prediction.
• The proposed method advances the field of automated biomedical knowledge discovery.