Improving chemical disease relation extraction with rich features and weakly labeled data

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Identifying chemical-disease relationships is crucial for drug discovery and chemical safety.
• Biomedical literature contains vast information on these relationships, necessitating automated extraction methods.
• Existing methods for chemical-induced disease (CID) relation extraction require improvement.

Purpose of the Study:

• To enhance the state-of-the-art in biomedical relation extraction, specifically for CID relations.
• To develop an automatic system for extracting CID relations from biomedical literature.
• To leverage advances in named entity recognition and BioCreative efforts.

Main Methods:

• A Support Vector Machine (SVM) based approach utilizing a rich feature set.
• Incorporation of novel statistical features, linguistic knowledge, and domain resources.
• Integration of a rule-based system's output and automatically generated labeled text from knowledge bases.

Main Results:

• The system achieved an F-score of 57.51% on the BioCreative V dataset, outperforming previous methods.
• Performance improved to 61.01% F-score when augmented with automatically generated weakly labeled data.
• The approach combines the strengths of rule-based and machine learning systems.

Conclusions:

• The developed text-mining approach achieves state-of-the-art performance in disease-chemical relation extraction.
• This work highlights the value of utilizing curated document-level annotations from existing biomedical databases.
• The findings suggest a more effective way to develop text-mining systems by leveraging overlooked resources.