Uncovering and Categorizing Social Biases in Text-to-SQL

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1343–1367 July 9-14, 2023 ©2023 Association for Computational Linguistics A Close Look into the Calibration of Pre-trained Language Models Yangyi Chen∗ UIUC Lifan Yuan∗ HUST Ganqu Cui, Zhiyuan Liu Tsinghua University Heng Ji UIUC yangyic3@illinois.edu lievanyuan173@gmail.com Abstract Pre-trained language models (PLMs) may fail in giving reliable estimates of their predic- tive uncertainty. We take a close look into this problem, aiming to answer two ques- tions: (1) Do PLMs learn to become cali- brated in the training process? (2) How effec- tive are existing calibration methods? For the first question, we conduct fine-grained con- trol experiments to study the dynamic change in PLMs’ calibration performance in training. We consider six factors as control variables, including dataset difficulty, available training samples, training steps, the number of tun- able parameters, model scale, and pretrain- ing. We observe a consistent change in cal- ibration performance across six factors. We find that PLMs don’t learn to become cali- brated in training, evidenced by the contin- ual increase in confidence, no matter whether the predictions are correct or not. We high- light that our finding somewhat contradicts two established conclusions: (a) Larger PLMs are more calibrated; (b) Pretraining improves model calibration. Next, we study the ef- fectiveness of existing calibration methods in mitigating the overconfidence issue. Besides unlearnable calibration methods (e.g., label smoothing), we adapt and extend two recently proposed learnable methods that directly col- lect data to train models to have reasonable confidence estimations. Experimental results show that learnable methods significantly re- duce PLMs’ confidence in wrong predictions. The code is available at https://github. com/lifan-yuan/PLMCalibration. 1 Introduction Pre-trained language models (PLMs) are success- ful in many downstream tasks regarding perfor- mance (Wang et al., 2019). In high-stake appli- cations, it’s equally essential for PLMs to pos- sess a sense of calibration (Vaicenavicius et al., ∗Equal contribution 0.0 0.8 1.6 2.4 3.2 4.0 Steps (×103) 0 20 40 60 80 100 Acc Conf ECE CErrpos CErrneg Under-fitted Over-fitted Figure 1: The demonstration of the under-fitted and over-fitted states in the training process with RoBERTa on SST-2. 2019). However, the confidence scores (a.k.a, pre- dictive probability) of existing deep neural net- works cannot serve as reliable estimates of their uncertainty (Guo et al., 2017), and a deep under- standing of PLMs calibration is lacking. In this paper, we give a systematical analysis of PLMs calibration. We consider two questions about PLMs calibration: (1) Do PLMs learn to become calibrated in the training process? (2) How effective are existing calibration methods? We first introduce the metrics we adopt for cali- bration performance evaluation. The most widely used calibration metric ECE (Expected Calibra- tion Error (Naeini et al., 2015)) is considered. It measures the difference between confidence and accuracy by portioning samples into various con- fidence zones. To give a more comprehensive and practical calibration evaluation, we provide an application-driven perspective, describing two un- desirable situations in practice: (1) Correct predic- …

Content truncated to 500 words