Calibrating Structured Output Predictors for Natural Language Processing