Discriminative Modeling of Extraction Sets for Machine Translation