A Hybrid System for Extracting Chemical-Disease Relationships from Scientific Literature

We propose a hybrid system for extracting chemical-disease relationships from Medline abstracts. At the core of our approach is a general, rule-based system that extracts causal relations from text, using a combination of trigger lists and syntactic dependencies. We augmented this system with supervised learning. We trained two binary classifiers: one extracts intra-sentential relationships between chemical-disease mention pairs, and the other attempts to extract relationships across sentences. Our hybrid system yielded an F1 score of 36.49. Our results on the development corpus reveal that chemical and disease named entity recognition are still major problems, and that improvements made in this area are likely to have a significant impact in chemical-disease relationship extraction.