Sieve-based coreference resolution enhances semi-supervised learning model for chemical-induced disease relation...

Area of Science:

• Biomedical text mining
• Computational biology
• Bioinformatics

Background:

• The BioCreative V chemical-disease relation (CDR) challenge aimed to advance text mining for understanding chemical-disease interactions.
• Previous systems, including UET-CAM, showed competitive but improvable performance.

Purpose of the Study:

• To enhance the UET-CAM system for improved chemical-disease relation extraction.
• To evaluate the impact of a novel 'silver' corpus and advanced coreference resolution on system performance.

Main Methods:

• Employed joint inference with a perceptron-based named entity recognizer and a back-off model for Disease Named Entity Recognition and Normalization (DNER).
• Utilized a pipeline for chemical-induced disease (CID) relation extraction, incorporating multi-pass sieve coreference resolution and a Support Vector Machine model.
• Developed a large 'silver' CID corpus (over 50,000 sentences) from the Comparative Toxicogenomics Database (CTD) for model training.

Main Results:

• Achieved state-of-the-art F1 scores: 82.44% for DNER and 58.90% for CID relation extraction on the CDR test set.
• Demonstrated significant performance gains: +4.13% F1 from multi-pass sieve coreference resolution and +7.3% F1 from the silver CID corpus.

Conclusions:

• The enhanced UET-CAM system, leveraging a silver CID corpus and multi-pass sieve coreference resolution, significantly improves chemical-disease relation extraction.
• The developed silver corpus and coreference method offer valuable resources and techniques for advancing biomedical text mining in this domain.