Uncertainty Decomposition via Cyclical SG-MCMC and Soft-label Learning for Subjective NLP
Researchers have developed a novel method for decomposing uncertainty in subjective Natural Language Processing tasks, specifically emotion classification. This approach combines cyclical stochastic gradient Markov chain Monte Carlo (cSG-MCMC) with soft-label learning, utilizing a frozen RoBERTa model. The method demonstrated superior performance on the GoEmotions benchmark compared to existing techniques like Monte Carlo Dropout and Deep Ensemble, achieving better results across multiple evaluation axes including divergence to annotator distribution and selective-prediction metrics. AI
IMPACT Introduces a novel technique for improving uncertainty estimation in subjective NLP tasks, potentially leading to more reliable emotion classification models.