Uncertainty Estimation of Transformer Predictions for Misclassification Detection

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8237 - 8252 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Uncertainty Estimation of Transformer Predictions for Misclassification Detection Artem Vazhentsev1,2 ♦, Gleb Kuzmin1,6 ♦, Artem Shelmanov1,7 ♦, Akim Tsvigun1,4, Evgenii Tsymbalov2, Kirill Fedyanin2, Maxim Panov2, Alexander Panchenko2, Gleb Gusev1,3,5, Mikhail Burtsev1,3, Manvel Avetisian1,5, and Leonid Zhukov1,4 1AIRI, 2Skoltech, 3MIPT, 4HSE, 5Sber AI Lab, 6FRC CSC RAS, 7ISP RAS Research Center for Trusted Artificial Intelligence {vazhentsev, kuzmin, shelmanov, tsvigun, gusev, burtsev, manvel, zhukov}@airi.net {evgenii.tsymbalov, m.panov, k.fedyanin, a.panchenko}@skoltech.ru Abstract Uncertainty estimation (UE) of model predictions is a crucial step for a variety of tasks such as active learning, misclassification detection, adversarial attack detection, out-of- distribution detection, etc. Most of the works on modeling the uncertainty of deep neural networks evaluate these methods on image classification tasks. Little attention has been paid to UE in natural language processing. To fill this gap, we perform a vast empirical investigation of state-of-the-art UE methods for Transformer models on misclassification detection in named entity recognition and text classification tasks and propose two computationally efficient modifications, one of which approaches or even outperforms computationally intensive methods1. 1 Introduction Machine learning methods are naturally prone to errors as they typically have to deal with ambiguous and incomplete data during both training and inference. Unreliable predictions hinder the application of these methods in domains, where the price of mistakes is very high, such as clinical medicine. Even in more error-tolerant domains and tasks, such as intent recognition in general- purpose chatbots, one would like to achieve a better trade-off between expressiveness of a model and its computational performance during inference. Since mistakes are inevitable, it is crucial to understand whether model predictions can be trusted or not and abstain from unreliable decisions. Uncertainty estimation (UE) of model predictions aims to solve this task. Ideally, uncertain instances should correspond to erroneous 1The code for experiments is available online at https://github.com/AIRI-Institute/ uncertainty_transformers ♦ Equal contribution, corresponding authors objects and help in misclassification detection. Besides misclassification detection, UE is a crucial component for active learning (Settles, 2009), adversarial attack detection (Lee et al., 2018), detection of out-of-distribution (OOD) instances (Van Amersfoort et al., 2020), etc. Some classical machine learning models, e.g. Gaussian processes (Rasmussen, 2003), have built-in UE capabilities. Modern deep neural networks (DNNs) usually take advantage of a softmax layer, which output can be considered as a prediction probability and be used for UE. However, the softmax probabilities are usually unreliable and produce overconfident predictions (Guo et al., 2017). Some previously proposed techniques such as deep ensemble (Lakshminarayanan et al., 2017) are known for producing good UE scores but require a large additional memory footprint for storing several versions of weights and multiply an amount of computation for conducting several forward passes. Reliable UE of DNN predictions that does not introduce high computational overhead is an open research question (Van Amersfoort et al., 2020). In this work, we investigate methods for UE of …

Content truncated to 500 words