structured output prediction problems (e.g., sequential tagging, hierarchical
multi-class classification) often involve constraints over the output label
space. These constraints interact with the learned models to filter infeasible
solutions and facilitate in building an accountable s