constraint-based learning reduces the burden of collecting labels by having
users specify general properties of structured outputs, such as constraints
imposed by physical laws. We propose a novel framework for simultaneously
learning these constraints and using them for supervision, b