A well-calibrated neural model produces confidence (probability outputs) closely approximated by the expected accuracy. While prior studies have shown that mixup training as a data augmentation technique can improve model calibration on image classification tasks, little is known about using mixup for model calibration on natural language understanding (NLU) tasks. In this paper, we explore mixup for model calibration on several NLU tasks and propose a novel mixup strategy for pre-trained language models that improves model calibration further. Our proposed mixup is guided by both the Area Under the Margin (AUM) statistic (Pleiss et al., 2020) and the saliency map of each sample (Simonyan et al., 2013). Moreover, we combine our mixup strategy with model miscalibration correction techniques (i.e., label smoothing and temperature scaling) and provide detailed analyses of their impact on our proposed mixup. We focus on systematically designing experiments on three NLU tasks: natural language inference, paraphrase detection, and commonsense reasoning. Our method achieves the lowest expected calibration error compared to strong baselines on both in-domain and out-of-domain test samples while maintaining competitive accuracy.

MixUp is a data augmentation strategy where additional samples are generated during training by combining random pairs of training samples and their labels. However, selecting random pairs is not potentially an optimal choice. In this work, we propose TDMixUp, a novel MixUp strategy that leverages Training Dynamics and allows more informative samples to be combined for generating new data samples. Our proposed TDMixUp first measures confidence, variability, (Swayamdipta et al., 2020), and Area Under the Margin (AUM) (Pleiss et al., 2020) to identify the characteristics of training samples (e.g., as easy-to-learn or ambiguous samples), and then interpolates these characterized samples. We empirically validate that our method not only achieves competitive performance using a smaller subset of the training data compared with strong baselines, but also yields lower expected calibration error on the pre-trained language model, BERT, on both in-domain and out-of-domain settings in a wide range of NLP tasks. We publicly release our code.