Chemical-induced disease relation extraction via attention-based distant supervision.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Accurate identification of chemical-disease relations (CDRs) is vital for biomedical research and healthcare.
• Supervised machine learning methods for CDR extraction rely heavily on large, manually annotated corpora, which are labor-intensive and costly to create.

Purpose of the Study:

• To develop an automated method for extracting chemical-disease relations (CDRs) from biomedical literature.
• To overcome the limitations of supervised learning by utilizing distant supervision, thereby eliminating the need for manual annotation.

Main Methods:

• An attention-based distant supervision paradigm was employed for the BioCreative-V CDR extraction task.
• Training examples were automatically generated from the Comparative Toxicogenomics Database (CTD) at both intra- and inter-sentence levels.
• Attention-based neural networks and stacked auto-encoder networks were utilized to build learning models and extract relations.

Main Results:

• The proposed method achieved a precision of 60.3%, recall of 73.8%, and F1-score of 66.4% on the BioCreative-V CDR task.
• The system outperformed state-of-the-art supervised learning systems.
• These results were obtained without using any manually annotated corpus.

Conclusions:

• Distant supervision is a promising approach for the automated extraction of chemical-disease relations from biomedical literature.
• Simultaneously capturing local and global attention features is effective in attention-based distantly supervised learning for CDR extraction.