Benchmarking Scalable Predictive Uncertainty in Text Classification

This paper explores the question of how predictive uncertainty methods perform in practice in Natural Language Processing, specifically multi-class and multi-label text classification. We conduct benchmarking experiments with 1-D convolutional neural networks and pre-trained transformers on six real-world text classification datasets in which we empirically investigate why popular scalable uncertainty estimation strategies (Monte-Carlo Dropout, Deep Ensemble) and notable extensions (Heteroscedastic, Concrete Dropout) underestimate uncertainty. We motivate that uncertainty estimation benefits from combining posterior approximation procedures, linking it to recent research on how ensembles and variational Bayesian methods navigate the loss landscape. We find that our proposed method combination of Deep Ensemble with Concrete Dropout, by analysis of in- domain calibration, cross-domain classification, and novel class robustness, demonstrates superior performance, even at a smaller ensemble size. Our results corroborate the importance of fine-tuning dropout rate to the text classification task at hand, which individually and as an ensemble impacts model robustness. We observe in ablation that pre-trained transformers severely underperform in novelty detection, limiting the applicability of transfer learning when distribution shift from novel classes can be expected.