Semi-supervised Methods for Biomedical Hedge Classification

We introduce an EM based approach to biomedical hedge classification that can used as a standalone classifier or as an extension to Medlock and Briscoe´s weakly supervised learning approach. We compare active learning and transductive learning to weakly supervised learning. We also introduce a support vector machine based feature selection method. An effective and computationally efficient stopping criterion and an adaptive batch size adjustment algorithm are our further contributions resulting in significant performance improvements over baseline.