Exploring Predictive Uncertainty and Calibration in NLP: A Study on the Impact of Method & Data Scarcity

TL;DR

This paper examines how different methods estimate predictive uncertainty in NLP models, especially under data scarcity, revealing that pre-trained models and ensembles perform best but can be affected by data volume, with uncertainties mainly driven by data rather than model factors.

Contribution

It provides a comprehensive evaluation of uncertainty estimation methods in low-resource NLP settings, highlighting the influence of data scarcity and model choice on uncertainty quality.

Findings

Pre-trained models and ensembles yield the best uncertainty estimates.

More data can sometimes degrade the quality of uncertainty estimates.

Model uncertainty is less influential than data uncertainty in total uncertainty.

Abstract

We investigate the problem of determining the predictive confidence (or, conversely, uncertainty) of a neural classifier through the lens of low-resource languages. By training models on sub-sampled datasets in three different languages, we assess the quality of estimates from a wide array of approaches and their dependence on the amount of available data. We find that while approaches based on pre-trained models and ensembles achieve the best results overall, the quality of uncertainty estimates can surprisingly suffer with more data. We also perform a qualitative analysis of uncertainties on sequences, discovering that a model's total uncertainty seems to be influenced to a large degree by its data uncertainty, not model uncertainty. All model implementations are open-sourced in a software package.