Uncertainty Estimation and Reduction of Pre-trained Models for Text Regression

Abstract State-of-the-art classification and regression models are often not well calibrated, cannot reliably provide uncertainty estimates, limiting their utility in safety-critical applications such as clinical decision-making. While recent work has focused on calibration of classifiers, there is almost no NLP a setting. In this paper, we quantify the pre- trained language for text regression, both intrinsically extrinsically. We further apply estimates to augment training data low-resource domains. Our experiments three tasks self-training active-learning settings show that estimation can be used increase overall performance enhance model generalization.