Spectral Regularization for Max-Margin Sequence Tagging

We frame max-margin learning of latent variable structured prediction models as a convex opti-mization problem, making use of scoring func-tions computed by input-output observable oper-ator models. This learning problem can be ex-pressed as an optimization problem involving a low-rank Hankel matrix that represents the input-output operator model. The direct outcome of our work is a new spectral regularization method for max-margin structured prediction. Our exper-iments confirm that our proposed regularization framework leads to an effective way of control-ling the capacity of structured prediction models. 1