Soft-Label Learning

Soft-label learning studies how to train models from targets that represent uncertainty, ambiguity, or disagreement instead of choosing one definitive label. Common approaches minimize divergence from annotator distributions, use label smoothing or teacher predictions, and model imprecise supervision as a set of possible distributions. The methods are used in subjective NLP, emotion and intent classification, calibration, and tasks with mixed or noisy regions; evaluations include distributional metrics such as KL divergence alongside accuracy and F1. A central tension is that matching human disagreement can improve uncertainty estimates even when hard-label metrics do not improve, so annotation budgets and evaluation metrics need to reflect the intended prediction target.