Calibration Meets Explanation: A Simple and Effective Approach for Model Confidence Estimates

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2775 - 2784 December 7-11, 2022 ©2022 Association for Computational Linguistics Calibration Meets Explanation: A Simple and Effective Approach for Model Confidence Estimates Dongfang Li1, Baotian Hu1∗, Qingcai Chen1,2∗ 1Harbin Institute of Technology (Shenzhen), Shenzhen, China 2Peng Cheng Laboratory, Shenzhen, China crazyofapple@gmail.com, {hubaotian, qingcai.chen}@hit.edu.cn Abstract Calibration strengthens the trustworthiness of black-box models by producing better accurate confidence estimates on given examples. How- ever, little is known about if model explanations can help confidence calibration. Intuitively, hu- mans look at important features attributions and decide whether the model is trustworthy. Similarly, the explanations can tell us when the model may or may not know. Inspired by this, we propose a method named CME that lever- ages model explanations to make the model less confident with non-inductive attributions. The idea is that when the model is not highly confident, it is difficult to identify strong in- dications of any class, and the tokens accord- ingly do not have high attribution scores for any class and vice versa. We conduct extensive experiments on six datasets with two popular pre-trained language models in the in-domain and out-of-domain settings. The results show that CME improves calibration performance in all settings. The expected calibration errors are further reduced when combined with tem- perature scaling. Our findings highlight that model explanations can help calibrate posterior estimates. 1 Introduction Accurate estimates of posterior probabilities are crucial for neural networks in various Natural Lan- guage Processing (NLP) tasks (Guo et al., 2017; Lakshminarayanan et al., 2017). For example, it would be helpful for humans if the models de- ployed in practice abstain or interact when they cannot make a decision with high confidence (Jiang et al., 2012). While Pre-trained Language Models (PLMs) have improved the performance of many NLP tasks (Devlin et al., 2019; Liu et al., 2019), how to better avoid miscalibration is still an open research problem (Desai and Durrett, 2020; Dan and Roth, 2021). In this paper, we investigate if ∗ Corresponding authors Positive a fast funny highly enjoyable movie. Negative It’s about following your dreams no matter what your parents think. Table 1: Two motivating examples with highlight ex- planations (Socher et al., 2013). The saturation of the colors signifies the magnitude. The confidence of the model should be easily recognized by looking at token attributions. and how model explanations can help calibrate the model. Explanation methods have attracted considerable research interest in recent years for revealing the internal reasoning processes behind models (Sun- dararajan et al., 2017; Heo et al., 2018; Shrikumar et al., 2017). Token attribution scores generated by explanation methods represent the contribution to the prediction (Atanasova et al., 2020). Intu- itively, one can draw some insight for analyzing and debugging neural models from these scores if they are correctly attributed, as shown in Table 1. For example, when the model identifies a highly indicative pattern, the tokens involved would have high attribution scores for the predicted label and …

Content truncated to 500 words