Learning to Reject with a Fixed Predictor: Application to Decontextualization