Total: 1
Text classifiers built on Pre-trained Language Models (PLMs) have achieved remarkable progress in various tasks including sentiment analysis, natural language inference, and question-answering. However, these text classifiers sometimes make uncertain predictions, which challenges their trustworthiness during deployment in practical applications. Much effort has been devoted to designing various probes in order to understand what PLMs capture. But few works have explored factors influencing PLM-based classifiers' predictive uncertainty. In this paper, we propose a novel framework CUE for interpreting uncertainties of PLM-based models' predictions. In particular, we first map PLM-encoded representations to a latent space via a variational auto-encoder. We then generate text representations by perturbing the latent space which causes fluctuation in predictive uncertainty. By comparing the predictive uncertainty difference between the perturbed text representation and the original text representation, we are able to identify the latent dimensions that cause uncertainty and thus trace back to input features that lead to uncertainty. Our extensive experiments on four benchmark datasets for linguistic acceptability classification, emotion classification, and natural language inference show the feasibility of our proposed framework.