Applying Feature Coupling Generalization for Protein-Protein Interaction Extraction

We present the application of a recently proposed semi-supervised learning strategy - feature coupling generalization (FCG) - in the task of protein-protein interaction extraction from biomedical literatures. FCG is a framework that generates new features from relatedness of two special types of old features: example-distinguishing features (EDFs) and class-distinguishing features (CDFs). Their relatedness estimated from unlabeled data tends to capture indicative information not available in labeled data. For this task, we designed several EDFs and CDFs derived from the text patterns surrounding the co-occurrence proteins, and combined the new features generated by FCG with Boolean lexical features. The experimental results on AIMED corpus show that the new features yield significant improvement over a strong baseline, and the combined method achieves state-of-the-art performance without using any syntactic information.