Document-Level Biomedical Relation Extraction Using Graph Convolutional Network and Multihead Attention: Algorithm...

Area of Science:

• Biomedical text mining
• Natural Language Processing
• Bioinformatics

Background:

• Chemical-disease relation (CDR) extraction is vital for understanding biomedical literature.
• Existing methods often overlook valuable cross-sentence dependency information, hindering accurate relation extraction.
• Accurate identification of intrasentence and intersentence relations is crucial for CDR extraction.

Purpose of the Study:

• To propose a novel end-to-end neural network for chemical-disease relation extraction.
• To leverage document-level dependency graphs and graph convolutional networks (GCNs) for improved CDR extraction.
• To enhance intersentence relation extraction by incorporating cross-sentence syntactic information.

Main Methods:

• Constructed a document-level dependency graph to capture cross-sentence syntactic dependencies.
• Applied Graph Convolutional Networks (GCNs) for feature representation of the dependency graph.
• Utilized multihead attention to learn important contextual features and deep context representation for enhanced input.

Main Results:

• Achieved a superior F-measure of 63.5% on the CDR corpus, outperforming state-of-the-art methods.
• Demonstrated strong performance at the intrasentence level with 59.1% precision, 81.5% recall, and 68.5% F-measure.
• Showcased improved intersentence relation extraction with 47.8% precision, 52.2% recall, and 49.9% F-measure.

Conclusions:

• Graph convolutional networks effectively utilize cross-sentence dependency information to enhance CDR extraction.
• Deep context representation and multihead attention mechanisms significantly contribute to improved CDR extraction performance.
• The proposed model offers a robust approach for accurate chemical-disease relation identification in biomedical texts.