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

TL;DR

This paper introduces a novel mixup strategy guided by AUM and saliency maps to improve calibration of pre-trained language models on NLU tasks, achieving lower calibration error while maintaining accuracy.

Contribution

It proposes a new mixup method guided by AUM and saliency, specifically tailored for NLP models, and combines it with calibration techniques for better model confidence estimation.

Findings

Achieves lower expected calibration error on multiple NLU tasks.

Maintains competitive accuracy while improving calibration.

Effective on both in-domain and out-of-domain data.

Abstract

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.