Learning Structured Classifiers with Dual Coordinate Descent

We present a unified framework for online learning of structured classifiers. This framework handles a wide family of convex loss functions that includes as particular cases CRFs, structured SVMs, and the structured perceptron. We introduce a new aggressive online algorithm that optimizes any loss in this family; for the structured hinge loss, this algorithm reduces to 1-best MIRA; in general, it can be regarded as a dual coordinate ascent algorithm. No learning rate parameter is required. Our experiments show that the technique is faster to converge to an accurate model than stochastic gradient descent, on two NLP problems, at least when inference is exact.