A span-graph neural model for overlapping entity relation extraction in biomedical texts.

Area of Science:

• Biomedical text mining
• Natural Language Processing
• Bioinformatics

Background:

• Entity relation extraction is crucial for biomedical text mining.
• Joint methods outperform traditional pipeline methods by preventing error propagation.
• Existing joint models struggle with overlapping entities and relations due to limitations in capturing long-range dependencies.

Purpose of the Study:

• To propose a novel span-graph neural model for jointly extracting overlapping entity relations in biomedical texts.
• To address the limitations of sequential models in handling long-range dependencies and overlapping information.
• To improve the accuracy and robustness of biomedical relation extraction.

Main Methods:

• A novel span-graph neural model is proposed, treating entity relation extraction as relation triplets prediction.
• The model constructs an entity-graph by enumerating candidate entity spans.
• It utilizes a span scorer and a relation scorer to capture relationships between correlated entities.

Main Results:

• The proposed model demonstrates superior performance on drug-drug and protein-protein interaction detection tasks.
• It significantly outperforms previous models in extracting overlapping relations.
• The span-graph approach shows enhanced capability in capturing long-range dependencies compared to sequential models.

Conclusions:

• The span-graph neural model is effective for jointly extracting overlapping entity relations in biomedical texts.
• The method offers a significant improvement over existing approaches, particularly for complex sentences with long-range dependencies.
• The publicly available code facilitates further research and application in the field.