Boosting performance of gene mention tagging system by classifiers ensemble

Abstract

To further improve the tagging performance of single classifiers, a classifiers ensemble experimental framework is presented for gene mention tagging. In the framework, six classifiers are constructed by four toolkits (CRF++, YamCha, Maximum Entropy (ME) and MALLET) with different training methods and feature sets and then combined with a two-layer stacking algorithm. The recognition results of different classifiers are regarded as input feature vectors to be incorporated, and then a high-powered model is obtained. Experiments carried out on the corpus of BioCreative II GM task show that the classifiers ensemble method is effective and our best combination method achieves an F-score of 88.09%, which outperforms most of the top-ranked Bio-NER systems in the BioCreAtIvE II GM challenge.