Bias-Corrected Ceilings of Emotion Predictability from Human Label Variation Based on Instance-Level Fano Bounds

A recent preprint on arXiv (2608.15619) presents a novel framework called Bias-corrected Affective Ceiling Estimation (BACE), designed to assess the maximum potential accuracy of emotion recognition from text. The authors highlight that, despite ongoing improvements in benchmarks, the critical inquiry into whether an accuracy ceiling has been achieved is often overlooked. BACE evaluates the influence of limited annotations, choice of estimator, annotation noise, and evaluation methods on the accuracy ceiling, thus refining saturation claims. It distinguishes between irreducible and reducible errors, employing an anchored Dirichlet-mixture empirical Bayes estimator, which lies between plug-in and NSB, to approximate the human-consensus distribution. The methodology reveals that point estimates suggest reachability ranges from 0.38 to 1.03, emphasizing the need for meticulous bias correction in saturation assertions. This work is classified as a new announcement and can be accessed on arXiv.