A Controlled Study of Auxiliary Supervision for Low-Resource Emotion Classification
Abstract
AbstractWe present a controlled empirical study of auxiliary supervision for low-resource emotion classification. Using a shared-encoder multi-task learning framework, we compare three auxiliary objectives - coarse affective valence, valence-activation structure, and positive-affect detection - across three supervision sources: aligned labels deterministically derived from gold emotion annotations, rule-based weak labels generated with VADER, and synthetic labels generated by Qwen2.5-14B. We evaluate on MELD, GoEmotions, and DAIR-AI Emotion under matched training procedures and multiple data regimes. Auxiliary supervision yields its clearest, though modest, gains in low-resource settings, especially on MELD and GoEmotions, while improvements on DAIR-AI Emotion are smaller and less consistent. Error and per-class analyses suggest that gains concentrate in semantically overlapping or low-frequency categories such as anger/disgust, fear/sadness, remorse, and embarrassment. We further find that agreement with aligned auxiliary labels is only weakly associated with downstream transfer, suggesting that useful auxiliary supervision need not exactly reproduce gold-derived affective mappings. Overall, our results support auxiliary supervision as a lightweight inductive bias for low-resource emotion classification, while also showing that its benefits are dataset-dependent and sensitive to auxiliary-task design.