We present a large-margin formulation and algorithm for structured output prediction that allows the use of latent variables. The paper identifies a particular formulation that covers a large range of
