Calibrating Expressions of Certainty

2024-10-06

arXiv.org

AI Summary

This paper addresses the semantic ambiguity and calibration difficulty of deterministic expressions (e.g., “possible”, “very likely”) in natural language by modeling their semantics as probability distributions over the probability simplex—replacing conventional scalar confidence scores. Methodologically, it (i) formally represents deterministic expressions as simplex-valued distributions for the first time; (ii) generalizes the notion of calibration error to distributional settings; (iii) proposes a distribution-mapping-based post-hoc calibration algorithm; and (iv) establishes a human–AI collaborative calibration analysis framework. Experiments involve radiologists and large language models, enabling quantitative, interpretable cross-subject calibration assessment and generating actionable calibration improvement recommendations. Results demonstrate significant gains in semantic consistency of uncertainty expressions and inter-subject reliability, establishing a novel paradigm for trustworthy human–AI collaborative decision-making.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI

Abstract

We present a novel approach to calibrating linguistic expressions of certainty, e.g.,"Maybe"and"Likely". Unlike prior work that assigns a single score to each certainty phrase, we model uncertainty as distributions over the simplex to capture their semantics more accurately. To accommodate this new representation of certainty, we generalize existing measures of miscalibration and introduce a novel post-hoc calibration method. Leveraging these tools, we analyze the calibration of both humans (e.g., radiologists) and computational models (e.g., language models) and provide interpretable suggestions to improve their calibration.

Problem

• Calibrating linguistic expressions of certainty accurately
• Modeling uncertainty as distributions over the simplex
• Improving calibration of humans and computational models

Innovation

• Model uncertainty as simplex distributions
• Generalize miscalibration measures
• Introduce post-hoc calibration method