Glossary

BLEU; Bilingual Evaluation Understudy

An n-gram-based indicator of surface similarity between a generated text and a reference (with a focus on precision). Useful for comparing outputs, but insufficient on its own to assess practical utility.

Statistical uncertainty calibration; Conformal Prediction (CP)

A threshold is set based on a small calibration set (in this case, 30 questions). With α = 5%, the maximum rate of risky answers is controlled. Above the threshold: abstention.

International Electrotechnical Commission (IEC)

Standards organization (electricity and electronics).

Operational Design Domain (ODD)

Scope of application where the AI is valid (device, task, language, versions, profiles). Uncertainty thresholds...