Stop Measuring Calibration When Humans Disagree
This addresses a methodological issue in machine learning evaluation for tasks with inherent human disagreement, such as natural language inference, and is incremental in refining calibration metrics.
The paper tackles the problem of evaluating classifier calibration when human annotators inherently disagree, showing that using the human majority class is theoretically flawed and empirically problematic on ChaosNLI, and proposes new instance-level measures based on statistical properties of human judgments.
Calibration is a popular framework to evaluate whether a classifier knows when it does not know - i.e., its predictive probabilities are a good indication of how likely a prediction is to be correct. Correctness is commonly estimated against the human majority class. Recently, calibration to human majority has been measured on tasks where humans inherently disagree about which class applies. We show that measuring calibration to human majority given inherent disagreements is theoretically problematic, demonstrate this empirically on the ChaosNLI dataset, and derive several instance-level measures of calibration that capture key statistical properties of human judgements - class frequency, ranking and entropy.