A semiparametric generative model for efficient structured-output supervised learning

We present a semiparametric generative model for supervised learning with structured outputs. The main algorithmic idea is to replace the parameters of an underlying generative model (such as a stochastic grammars) with input-dependent predictions obtained by (kernel) logistic regression. This method avoids the computational burden associated with the comparison between target and predicted structure during the training phase, but requires as an additional input a vector of sufficient statistics for each training example. The resulting training algorithm is asymptotically more efficient than structured output SVM as the size of the output structure grows. At the same time, by computing parameters of a joint distribution as a function of the full input structure, typical expressiveness limitations of related conditional models (such as maximum entropy Markov models) can be potentially avoided. Empirical results on artificial and real data (in the domains of natural language parsing and RNA secondary structure prediction) show that the method works well in practice and scales up with the size of the output structures.