Over the last five years, we have been able to extend the theory of linear classifiers to structure prediction problems, combining the benefits of discriminative learning and of structured probabilist
