Document-Level Chemical-Induced Disease Relation Extraction via Hierarchical Representation Learning.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Chemical-induced Disease (CID) relations are crucial in biomedical research and healthcare.
• Existing methods do not fully leverage local and global contextual information in documents, limiting performance.
• Improved CID relation extraction is needed for enhanced biomedical applications.

Purpose of the Study:

• To propose a novel framework for document-level Chemical-induced Disease (CID) relation extraction.
• To enhance the modeling of interactions between local and global contexts within biomedical documents.
• To improve the accuracy of CID relation classification.

Main Methods:

• A stacked Hypergraph Aggregation Neural Network (HANN) is introduced to model complex contextual interactions.
• HANN layers generate improved contextualized representations for CID relation extraction.
• A CID Relation Heterogeneous Graph is constructed to incorporate multi-granularity information.

Main Results:

• The proposed framework effectively models local and global contextual interactions.
• The Hypergraph Aggregation Neural Network (HANN) yields better contextualized representations.
• The CID Relation Heterogeneous Graph enhances CID relation classification performance.
• Experimental results on a real-world dataset validate the framework's effectiveness.

Conclusions:

• The novel framework significantly improves document-level CID relation extraction.
• The integration of HANN and heterogeneous graphs offers a powerful approach for biomedical text mining.
• This work contributes to advancing the accuracy and applicability of CID relation identification.