BioEGRE: a linguistic topology enhanced method for biomedical relation extraction based on BioELECTRA and graph...

Area of Science:

• Bioinformatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical relation extraction (BioRE) is vital for text mining.
• Current methods use pre-trained language models but underutilize linguistic topology.
• Sequence-based models have limitations in processing graphical features.

Purpose of the Study:

• To propose a novel method, BioEGRE, for sentence-level BioRE.
• To leverage linguistic topological features for improved relation extraction.
• To enhance the accuracy and generalization of biomedical relation extraction.

Main Methods:

• BioEGRE preprocesses literature to identify relevant sentences.
• Dependency parsing creates sentence graphs with BioELECTRA token representations.
• A graph pointer neural network layer optimizes node representations.

Main Results:

• BioEGRE achieved F1-scores of 79.97% (CHEMPROT), 83.31% (GAD), and 83.51% (EU-ADR).
• The method surpassed existing state-of-the-art models on three BioRE tasks.
• Results demonstrate superior performance in multi-class and binary relation extraction.

Conclusions:

• BioEGRE proves effective and generalizable across biomedical datasets.
• Linguistic topology and graph pointer networks enhance BioRE performance.
• The proposed method advances automated biomedical knowledge discovery.