[Paper Review] Uncertainty Quantification with Pre-trained Language Models: A Large-Scale Empirical Analysis

Yuxin Xiao, Paul Pu Liang+4 more|arXiv (Cornell University)|Oct 10, 2022

Natural Language Processing Techniques4 citations

TL;DR

This paper presents a large-scale empirical study on uncertainty quantification in pre-trained language models (PLMs), evaluating key design choices including PLM architecture, size, uncertainty quantifiers, and fine-tuning losses. It finds that using ELECTRA with Focal Loss fine-tuning, Temp Scaling recalibration, and larger model sizes yields the best calibration across NLP classification tasks under in-domain and out-of-domain settings.

ABSTRACT

Pre-trained language models (PLMs) have gained increasing popularity due to their compelling prediction performance in diverse natural language processing (NLP) tasks. When formulating a PLM-based prediction pipeline for NLP tasks, it is also crucial for the pipeline to minimize the calibration error, especially in safety-critical applications. That is, the pipeline should reliably indicate when we can trust its predictions. In particular, there are various considerations behind the pipeline: (1) the choice and (2) the size of PLM, (3) the choice of uncertainty quantifier, (4) the choice of fine-tuning loss, and many more. Although prior work has looked into some of these considerations, they usually draw conclusions based on a limited scope of empirical studies. There still lacks a holistic analysis on how to compose a well-calibrated PLM-based prediction pipeline. To fill this void, we compare a wide range of popular options for each consideration based on three prevalent NLP classification tasks and the setting of domain shift. In response, we recommend the following: (1) use ELECTRA for PLM encoding, (2) use larger PLMs if possible, (3) use Temp Scaling as the uncertainty quantifier, and (4) use Focal Loss for fine-tuning.

Motivation & Objective

• To identify the most effective configuration for building a well-calibrated PLM-based prediction pipeline in safety-critical NLP applications.
• To investigate how different choices—PLM architecture, size, uncertainty quantifier, and fine-tuning loss—affect calibration error in diverse NLP tasks.
• To evaluate pipeline performance under domain shift, assessing robustness and reliability in out-of-distribution settings.
• To provide actionable recommendations based on large-scale empirical evaluation across three NLP tasks and in- and out-of-domain settings.

Proposed method

• Conducted large-scale empirical evaluation across three NLP classification tasks: Sentiment Analysis, Natural Language Inference, and Commonsense Reasoning.
• Evaluated 120 distinct PLM-based pipelines combining different configurations of PLM architecture, size, uncertainty quantifier, and fine-tuning loss.
• Used in-domain and out-of-domain test sets to assess calibration robustness under distribution shift.
• Measured calibration using Expected Calibration Error (ECE), Maximum Mean Calibration Error (MMCE), and selective prediction performance.
• Evaluated out-of-domain detection capability using FAR95 (false alarm rate at 95% recall) to assess uncertainty reliability on unseen domains.
• Applied post-hoc calibration via Temp Scaling and compared it with MC Dropout and ensemble methods.

Experimental results

Research questions

• RQ1 Which pre-trained language model architecture yields the most consistent calibration performance across diverse NLP tasks and domain shifts?
• RQ2 How does model size influence calibration quality, both in-domain and out-of-domain?
• RQ3 Which uncertainty quantifier (e.g., Temp Scaling, MC Dropout, ensemble) most effectively reduces calibration error in PLM-based pipelines?
• RQ4 Which fine-tuning loss (e.g., Cross Entropy, Focal Loss, Label Smoothing, Brier Loss, MMCE) leads to better-calibrated predictions across tasks and domains?

Key findings

• ELECTRA consistently outperforms other PLMs in calibration quality across all three tasks and both in- and out-of-domain settings.
• Larger PLMs yield significantly better calibration, especially in Commonsense Reasoning, where larger models show superior in-domain calibration compared to smaller variants.
• Focal Loss achieves the lowest Expected Calibration Error (ECE) in both in-domain and out-of-domain evaluations and delivers the best FAR95 score, indicating strong out-of-domain detection capability.
• Temp Scaling is the most effective post-hoc uncertainty quantifier, significantly reducing calibration error, particularly in out-of-domain scenarios.
• Ensemble methods are less effective in reducing calibration error when applied to PLM-based pipelines, contrary to their success with traditional models.
• Label Smoothing and Brier Loss improve out-of-domain calibration but are less effective than Focal Loss in maintaining low ECE and high detection performance.