Uncertainty Decomposition via Cyclical SG-MCMC and Soft-label Learning for Subjective NLP
The paper discusses a novel approach to uncertainty decomposition in subjective natural language processing (NLP). It integrates soft-label learning with Bayesian deep learning to assess annotator disagreement in emotion classification. The proposed method demonstrates improved performance on the GoEmotions benchmark compared to existing techniques.
- ▪The study focuses on annotator disagreement in emotion classification, which reflects intrinsic ambiguity in emotion concepts.
- ▪The authors propose a method that combines cyclical stochastic gradient Markov chain Monte Carlo with soft-label learning.
- ▪Results show that the new method outperforms Monte Carlo Dropout and Deep Ensemble on multiple evaluation axes.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.24773 (cs) [Submitted on 23 May 2026] Title:Uncertainty Decomposition via Cyclical SG-MCMC and Soft-label Learning for Subjective NLP Authors:Keito Inoshita, Takato Ueno View a PDF of the paper titled Uncertainty Decomposition via Cyclical SG-MCMC and Soft-label Learning for Subjective NLP, by Keito Inoshita and 1 other authors View PDF HTML (experimental) Abstract:Annotator disagreement in emotion classification reflects ambiguity intrinsic to emotion concepts and is essential for predictor-quality assessment in subjective NLP. Yet no prior work integrates soft-label learning with Bayesian deep learning to evaluate uncertainty along axes including annotator-distribution fidelity.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
