LoGo-GR: A Local to Global Graphical Reasoning Framework for Extracting Structured Information From Biomedical...

Area of Science:

• Computational Biology
• Natural Language Processing
• Bioinformatics

Background:

• Biomedical literature contains complex entity interactions distributed across multiple sentences.
• Manual extraction of biomedical knowledge is costly and impractical due to the increasing volume of literature.
• Document-level relation extraction is crucial for efficiently mining structured knowledge from vast medical texts.

Purpose of the Study:

• To address the challenge of synthesizing global context and local dependencies in document-level relation extraction.
• To propose a novel framework, Local to Global Graphical Reasoning (LoGo-GR), for capturing complex entity interactions.
• To enhance the accuracy and efficiency of extracting structured biomedical knowledge from unstructured text.

Main Methods:

• Developed a Local to Global Graphical Reasoning (LoGo-GR) framework utilizing a novel Biased Graph Attention mechanism (B-GAT).
• LoGo-GR learns global context features and local relation path dependencies using mention-level interaction and entity-level path graphs.
• B-GAT integrates structural dependencies as attention biases into the Graph Attention mechanism (GAT) to guide information aggregation.

Main Results:

• Achieved state-of-the-art performance on the Drug-Mutation Interaction (DV) dataset (96.14%-97.39% F1).
• Demonstrated advanced performance on the Chemical-induced Disease (CDR) dataset (68.89% F1) and Gene-Disease Association (GDA) dataset (84.22% F1).
• Showcased robust performance on the general-domain document-level relation extraction dataset, DocRED.

Conclusions:

• The LoGo-GR framework effectively captures complex interactions between entities in document-level biomedical texts.
• The proposed B-GAT mechanism enhances graphical reasoning by adaptively guiding information aggregation.
• LoGo-GR is an effective and robust framework for document-level relation extraction, applicable to both biomedical and general domains.