Improving the performance of protein-protein interaction article selection using domain specific features

Abstract

We investigate the use of a combination of features including term weights schemes and features generated from the biomedical domain for the protein interaction document classification task. In particular we propose to use output scores of Bayesian based classifiers which capture biomedical domain knowledge as features. Combining such features with term weights and number of protein mentions in text proves to be extremely useful. Results show that our best system achieves an Fscore of 81.31% which is 3.31% better than the best performance in BiOCreAtIvE-II Interaction Article Selection Task. Copyright© (2009) by the International Society for Research in Science and Technology.