Learning rules to extract protein interactions from biomedical text

Abstract

We present a method for automatic extraction of protein interactions from scientific abstracts by combing machine learning and knowledge-based strategies. This method uses sample sentences, which are parsed by a link grammar parser, to learn extraction rules automatically. By incorporating heuristic rules based on morphological clues and domain specific knowledge, this method can remove the interactions that are not between proteins and improve the performance of extraction process. We present experimental results for a test set of MEDLINE abstracts. The results are encouraging and demonstrate the feasibility of our method to perform accurate extraction without need of manual rule building.