Learning from the Best: Rationalizing Prediction by Adversarial Information Calibration