Investigating Selective Prediction Approaches Across Several Tasks in IID, OOD, and Adversarial Settings

Findings of the Association for Computational Linguistics: ACL 2022, pages 1995 - 2002 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Investigating Selective Prediction Approaches Across Several Tasks in IID, OOD, and Adversarial Settings Neeraj Varshney, Swaroop Mishra, Chitta Baral Arizona State University {nvarshn2, srmishr1, cbaral}@asu.edu Abstract In order to equip NLP systems with ‘selective prediction’ capability, several task-specific ap- proaches have been proposed. However, which approaches work best across tasks or even if they consistently outperform the simplest base- line MaxProb remains to be explored. To this end, we systematically study selective pre- diction in a large-scale setup of 17 datasets across several NLP tasks. Through comprehen- sive experiments under in-domain (IID), out-of- domain (OOD), and adversarial (ADV) settings, we show that despite leveraging additional re- sources (held-out data/computation), none of the existing approaches consistently and con- siderably outperforms MaxProb in all three set- tings. Furthermore, their performance does not translate well across tasks. For instance, Monte-Carlo Dropout outperforms all other ap- proaches on Duplicate Detection datasets but does not fare well on NLI datasets, especially in the OOD setting. Thus, we recommend that fu- ture selective prediction approaches should be evaluated across tasks and settings for reliable estimation of their capabilities. 1 Introduction Despite impressive progress made in Natural Lan- guage Processing (NLP), it is unreasonable to ex- pect models to be perfect in their predictions. They often make incorrect predictions, especially when inputs tend to diverge from their training data dis- tribution (Elsahar and Gallé, 2019; Miller et al., 2020; Koh et al., 2021). While this is acceptable for tolerant applications like movie recommenda- tions, high risk associated with incorrect predic- tions hinders the adoption of these systems in real- world safety-critical domains like biomedical and autonomous robots. In such scenarios, selective prediction becomes crucial as it allows maintaining high accuracy by abstaining on instances where error is likely. Selective Prediction (SP) has been studied in machine learning (Chow, 1957; El-Yaniv et al., 2010) and computer vision (Geifman and El-Yaniv, 2017, 2019), but has only recently gained atten- tion in NLP. Kamath et al. (2020) proposed a post- hoc calibration-based SP technique for Question- Answering (QA) datasets. Garg and Moschitti (2021) distill the QA model to filter out error-prone questions. Unfortunately, despite the shared goal of making NLP systems robust and reliable for real-world applications, SP has remained underex- plored; the community does not know which tech- niques work best across tasks/settings or even if they consistently outperform the simplest baseline MaxProb (Hendrycks and Gimpel, 2017) (that uses a threshold over the maximum softmax probability for selective prediction). In this work, we address the above point and study selective prediction in a large-scale setup of 17 datasets spanning over Natural Language Infer- ence (NLI), Duplicate Detection, and QA tasks. Our comprehensive experiments under In-Domain (IID), Out-Of-Domain (OOD), and Adversarial (ADV) settings result in the following findings: 1. None of the existing SP approaches consistently and considerably outperforms MaxProb. Slight improvement in IID: Most of the ap- proaches …

Content truncated to 500 words