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

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.

• Publication year

  2022

• Venue

  Conference on Empirical Methods in Natural Language Processing

• Publication date

  2022-10-20

• Fields of study

  Linguistics, Computer Science

• Identifiers
  DOI 10.48550/arXiv.2210.15452arXiv 2210.15452
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• Feature Collapse

  2023cited by this paper
• Out-of-Domain Evaluation of Finnish Dependency Parsing

  2022influential reference
• deep-significance - Easy and Meaningful Statistical Significance Testing in the Age of Neural Networks

  2022cited by this paper
• Experimental Standards for Deep Learning Research: A Natural Language Processing Perspective

  2022cited by this paper
• Toward More Meaningful Resources for Lower-resourced Languages

  2022cited by this paper
• Explaining Prediction Uncertainty of Pre-trained Language Models by Detecting Uncertain Words in Inputs

  2022cited by this paper
• Einops: Clear and Reliable Tensor Manipulations with Einstein-like Notation

  2022cited by this paper
• A Simple Approach to Improve Single-Model Deep Uncertainty via Distance-Awareness

  2022influential reference
• Local Languages, Third Spaces, and other High-Resource Scenarios

  2022cited by this paper
• Shifts: A Dataset of Real Distributional Shift Across Multiple Large-Scale Tasks

  2021cited by this paper
• Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration

  2021cited by this paper
• Revisiting the Calibration of Modern Neural Networks

  2021cited by this paper
• On Feature Collapse and Deep Kernel Learning for Single Forward Pass Uncertainty

  2021cited by this paper
• An Information-theoretic Approach to Distribution Shifts

  2021cited by this paper
• Knowing More About Questions Can Help: Improving Calibration in Question Answering

  2021cited by this paper
• Uncertainty-Aware Abstractive Summarization

  2021cited by this paper
• On Hallucination and Predictive Uncertainty in Conditional Language Generation

  2021cited by this paper
• Deterministic Neural Networks with Appropriate Inductive Biases Capture Epistemic and Aleatoric Uncertainty

  2021cited by this paper
• How Certain is Your Transformer?

  2021cited by this paper
• On Calibration and Out-of-domain Generalization

  2021cited by this paper
• Calibrate Before Use: Improving Few-Shot Performance of Language Models

  2021cited by this paper
• On the Calibration and Uncertainty of Neural Learning to Rank Models for Conversational Search

  2021cited by this paper
• Uncertainty Estimation in Autoregressive Structured Prediction

  2021cited by this paper
• Rethinking Calibration of Deep Neural Networks: Do Not Be Afraid of Overconfidence

  2021cited by this paper
• On the Effects of Transformer Size on In- and Out-of-Domain Calibration

  2021cited by this paper
• A Geometric Perspective towards Neural Calibration via Sensitivity Decomposition

  2021cited by this paper
• Uncertainty-Aware Machine Translation Evaluation

  2021cited by this paper
• Types of Out-of-Distribution Texts and How to Detect Them

  2021influential reference
• Genre as Weak Supervision for Cross-lingual Dependency Parsing

  2021cited by this paper
• Datasets: A Community Library for Natural Language Processing

  2021cited by this paper
• Soft Calibration Objectives for Neural Networks

  2021cited by this paper
• Infinite attention: NNGP and NTK for deep attention networks

  2020cited by this paper
• Discriminative Jackknife: Quantifying Uncertainty in Deep Learning via Higher-Order Influence Functions

  2020cited by this paper
• Being Bayesian about Categorical Probability

  2020cited by this paper
• Calibrating Deep Neural Networks using Focal Loss

  2020cited by this paper
• Calibration of Pre-trained Transformers

  2020cited by this paper
• Calibrating Structured Output Predictors for Natural Language Processing

  2020cited by this paper
• On the Inference Calibration of Neural Machine Translation

  2020cited by this paper
• DaNE: A Named Entity Resource for Danish

  2020cited by this paper
• Wat zei je? Detecting Out-of-Distribution Translations with Variational Transformers

  2020influential reference
• Selective Question Answering under Domain Shift

  2020cited by this paper
• Infinite attention: NNGP and NTK for deep attention networks

  2020cited by this paper
• Contrastive Training for Improved Out-of-Distribution Detection

  2020cited by this paper
• Repulsive Attention: Rethinking Multi-head Attention as Bayesian Inference

  2020cited by this paper
• Empirical Frequentist Coverage of Deep Learning Uncertainty Quantification Procedures

  2020cited by this paper
• Uncertainty-Aware Semantic Augmentation for Neural Machine Translation

  2020cited by this paper
• Understanding Neural Abstractive Summarization Models via Uncertainty

  2020cited by this paper
• Formalizing Trust in Artificial Intelligence: Prerequisites, Causes and Goals of Human Trust in AI

  2020cited by this paper
• Modeling Token-level Uncertainty to Learn Unknown Concepts in SLU via Calibrated Dirichlet Prior RNN

  2020cited by this paper
• A Survey on Recent Approaches for Natural Language Processing in Low-Resource Scenarios

  2020cited by this paper
• Computational Methods in Legal Analysis

  2020cited by this paper
• Underspecification Presents Challenges for Credibility in Modern Machine Learning

  2020cited by this paper
• Trust Issues: Uncertainty Estimation Does Not Enable Reliable OOD Detection On Medical Tabular Data

  2020cited by this paper
• Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty

  2020cited by this paper
• Uncertainty Estimation and Calibration with Finite-State Probabilistic RNNs

  2020cited by this paper
• DaN+: Danish Nested Named Entities and Lexical Normalization

  2020cited by this paper
• How Can We Know When Language Models Know? On the Calibration of Language Models for Question Answering

  2020cited by this paper
• Know your limits: Uncertainty estimation with ReLU classifiers fails at reliable OOD detection

  2020cited by this paper
• Quantifying the Carbon Emissions of Machine Learning

  2019cited by this paper
• Measuring Calibration in Deep Learning

  2019influential reference
• On Mixup Training: Improved Calibration and Predictive Uncertainty for Deep Neural Networks

  2019cited by this paper
• Bayesian Modelling in Practice: Using Uncertainty to Improve Trustworthiness in Medical Applications

  2019cited by this paper
• Deep Dominance - How to Properly Compare Deep Neural Models

  2019cited by this paper
• An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction

  2019influential reference
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

  2019cited by this paper
• Multilingual is not enough: BERT for Finnish

  2019cited by this paper
• Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods

  2019cited by this paper
• Out-of-Domain Detection for Low-Resource Text Classification Tasks

  2019cited by this paper
• Energy Usage Reports: Environmental awareness as part of algorithmic accountability

  2019cited by this paper
• PyTorch: An Imperative Style, High-Performance Deep Learning Library

  2019cited by this paper
• A guide to deep learning in healthcare

  2019cited by this paper
• Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift

  2019cited by this paper
• Predictive Uncertainty Estimation via Prior Networks

  2018cited by this paper
• Understanding Measures of Uncertainty for Adversarial Example Detection

  2018cited by this paper
• Why ReLU Networks Yield High-Confidence Predictions Far Away From the Training Data and How to Mitigate the Problem

  2018cited by this paper
• Do Deep Generative Models Know What They Don't Know?

  2018cited by this paper
• GPyTorch: Blackbox Matrix-Matrix Gaussian Process Inference with GPU Acceleration

  2018cited by this paper
• Deep Bayesian Active Learning for Natural Language Processing: Results of a Large-Scale Empirical Study

  2018cited by this paper
• Evidential Deep Learning to Quantify Classification Uncertainty

  2018cited by this paper
• Single-Model Uncertainties for Deep Learning

  2018cited by this paper
• Attention is All you Need

  2017cited by this paper
• On Calibration of Modern Neural Networks

  2017cited by this paper
• An Optimal Transportation Approach for Assessing Almost Stochastic Order

  2017cited by this paper
• Bayesian Recurrent Neural Networks

  2017influential reference
• Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization

  2016cited by this paper
• A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks

  2016cited by this paper
• Scalable Bayesian Learning of Recurrent Neural Networks for Language Modeling

  2016cited by this paper
• Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles

  2016cited by this paper
• Data-driven HR - Résumé Analysis Based on Natural Language Processing and Machine Learning

  2016cited by this paper
• Obtaining Well Calibrated Probabilities Using Bayesian Binning

  2015cited by this paper
• A Theoretically Grounded Application of Dropout in Recurrent Neural Networks

  2015influential reference
• Universal Dependencies for Finnish

  2015cited by this paper
• Weight Uncertainty in Neural Network

  2015cited by this paper
• Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

  2015influential reference
• Adam: A Method for Stochastic Optimization

  2014cited by this paper
• Recurrent Neural Network Regularization

  2014cited by this paper
• Building the essential resources for Finnish: the Turku Dependency Treebank

  2013cited by this paper
• Processing

  2013cited by this paper
• A unifying view on dataset shift in classification

  2012cited by this paper
• Random Search for Hyper-Parameter Optimization

  2012cited by this paper

• On the Fallacy of Global Token Perplexity in Spoken Language Model Evaluation

  2026cites this paper
• Making Foundation Models Probabilistic via Singular Value Ensembles

  2026cites this paper
• Show me the evidence: Evaluating the role of evidence and natural language explanations in AI-supported fact-checking

  2026cites this paper
• Enhancing Causal Text Detection Using Uncertainty-Weighted Machine Learning Ensembles

  2026cites this paper
• Calibration Is Not Enough: Evaluating Confidence Estimation Under Language Variations

  2026cites this paper
• The Capabilities and Limitations of Weak-to-Strong Generalization: Generalization and Calibration

  2025cites this paper
• Interplay between Bayesian neural networks and deep learning: A survey

  2025cites this paper
• Understanding the Capabilities and Limitations of Weak-to-Strong Generalization

  2025cites this paper
• Uncertainty in Causality: A New Frontier

  2025cites this paper
• MutantPrompt: Prompt Optimization via Mutation Under a Budget on Modest-sized LMs

  2025cites this paper
• Do LLMs Know When to NOT Answer? Investigating Abstention Abilities of Large Language Models

  2024cites this paper
• Crowd-Calibrator: Can Annotator Disagreement Inform Calibration in Subjective Tasks?

  2024cites this paper
• Beyond Simple Averaging: Improving NLP Ensemble Performance with Topological-Data-Analysis-Based Weighting

  2024cites this paper
• Graph-based Confidence Calibration for Large Language Models

  2024cites this paper
• Can LLMs Learn Uncertainty on Their Own? Expressing Uncertainty Effectively in A Self-Training Manner

  2024cites this paper
• Understanding the Impact of Confidence in Retrieval Augmented Generation: A Case Study in the Medical Domain

  2024cites this paper
• On the State of NLP Approaches to Modeling Depression in Social Media: A Post-COVID-19 Outlook

  2024cites this paper
• Non-Exchangeable Conformal Language Generation with Nearest Neighbors

  2024cites this paper
• Diversity-Aware Ensembling of Language Models Based on Topological Data Analysis

  2024cites this paper
• Interpreting Predictive Probabilities: Model Confidence or Human Label Variation?

  2024cites this paper
• API Is Enough: Conformal Prediction for Large Language Models Without Logit-Access

  2024cites this paper
• Calibrating Large Language Models Using Their Generations Only

  2024cites this paper
• How much reliable is ChatGPT's prediction on Information Extraction under Input Perturbations?

  2024cites this paper
• When Quantization Affects Confidence of Large Language Models?

  2024cites this paper
• Are Data Augmentation Methods in Named Entity Recognition Applicable for Uncertainty Estimation?

  2024cites this paper
• Efficient Nearest Neighbor based Uncertainty Estimation for Natural Language Processing Tasks

  2024cites this paper
• Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models

  2023cites this paper
• Uncertainty in Natural Language Generation: From Theory to Applications

  2023cites this paper
• Shifting Attention to Relevance: Towards the Uncertainty Estimation of Large Language Models

  2023cites this paper
• A Close Look into the Calibration of Pre-trained Language Models

  2022cites this paper
• Prior and Posterior Networks: A Survey on Evidential Deep Learning Methods For Uncertainty Estimation

  2021cites this paper
• S HIFTING A TTENTION TO R ELEVANCE : T OWARDS THE U NCERTAINTY

  year unknowncites this paper