On the Calibration of Pre-trained Language Models using Mixup Guided by Area Under the Margin and Saliency

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5364 - 5374 May 22-27, 2022 c⃝2022 Association for Computational Linguistics On the Calibration of Pre-trained Language Models using Mixup Guided by Area Under the Margin and Saliency Seo Yeon Park and Cornelia Caragea Computer Science University of Illinois Chicago spark313@uic.edu cornelia@uic.edu Abstract A well-calibrated neural model produces con- fidence (probability outputs) closely approx- imated by the expected accuracy. While prior studies have shown that mixup train- ing as a data augmentation technique can im- prove model calibration on image classifica- tion tasks, little is known about using mixup for model calibration on natural language un- derstanding (NLU) tasks. In this paper, we explore mixup for model calibration on sev- eral NLU tasks and propose a novel mixup strategy for pre-trained language models that improves model calibration further. Our pro- posed mixup is guided by both the Area Un- der the Margin (AUM) statistic (Pleiss et al., 2020) and the saliency map of each sample (Si- monyan et al., 2013). Moreover, we combine our mixup strategy with model miscalibra- tion correction techniques (i.e., label smooth- ing and temperature scaling) and provide de- tailed analyses of their impact on our proposed mixup. We focus on systematically design- ing experiments on three NLU tasks: natu- ral language inference, paraphrase detection, and commonsense reasoning. Our method achieves the lowest expected calibration er- ror compared to strong baselines on both in- domain and out-of-domain test samples while maintaining competitive accuracy. 1 Introduction Training a well-calibrated classifier that produces a match between confidence (the probability output that a model assigns to a prediction) and correct- ness (accuracy), is important in modern neural net- works. As an example, if an AI-based application knows what it does not know, or in other words, the chance that the current prediction is wrong, a human is more helpful to correct the error. How- ever, many works reveal that current deep neural networks are prone to over-confidence, which im- plies that the models’ confidence is not reliable (Guo et al., 2017). This is a critical issue on the deployment of AI-based user applications such as the healthcare domain (Zhu et al., 2018; Li et al., 2019) or safety-critical domain (Sarabadani, 2019) due to the problem of prediction trustworthiness. Recently, the study of calibration on neural net- work models especially on natural language pro- cessing tasks has started to receive attention. To overcome the problem of miscalibration, numerous suggestions on how to address it have been pro- posed. For example, Guo et al. (2017) revealed that using temperature scaling before the final soft- max layer reduces calibration errors. Müller et al. (2019), Kumar and Sarawagi (2019), and Wang et al. (2020a) found that label smoothing and its variants yield better calibration for neural machine translation. Desai and Durrett (2020) also reported that the aforementioned miscalibration correction methods can be applied to calibrate pre-trained language models which are often miscalibrated po- tentially due to over-parameterization. …

Content truncated to 500 words