Learning Adaptive Value of Information for Structured Prediction