Data selection for discriminative training in statistical machine translation