Interpreting the Robustness of Neural NLP Models to Textual Perturbations

Findings of the Association for Computational Linguistics: ACL 2022, pages 3993 - 4007 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Interpreting the Robustness of Neural NLP Models to Textual Perturbations Yunxiang Zhang1, Liangming Pan2, Samson Tan2, Min-Yen Kan2 1Wangxuan Institute of Computer Technology, Peking University 2School of Computing, National University of Singapore yx.zhang@pku.edu.cn, liangmingpan@u.nus.edu, {samson.tmr,kanmy}@comp.nus.edu.sg Abstract Modern Natural Language Processing (NLP) models are known to be sensitive to input per- turbations and their performance can decrease when applied to real-world, noisy data. How- ever, it is still unclear why models are less ro- bust to some perturbations than others. In this work, we test the hypothesis that the extent to which a model is affected by an unseen tex- tual perturbation (robustness) can be explained by the learnability of the perturbation (defined as how well the model learns to identify the perturbation with a small amount of evidence). We further give a causal justification for the learnability metric. We conduct extensive ex- periments with four prominent NLP models — TextRNN, BERT, RoBERTa and XLNet — over eight types of textual perturbations on three datasets. We show that a model which is better at identifying a perturbation (higher learnability) becomes worse at ignoring such a perturbation at test time (lower robustness), providing empirical support for our hypothesis. 1 Introduction Despite the success of deep neural models on many Natural Language Processing (NLP) tasks (Liu et al., 2016; Devlin et al., 2019; Liu et al., 2019b), recent work has discovered that these models are not robust to noisy input from the real world and thus their performance will decrease (Prabhakaran et al., 2019; Niu et al., 2020; Ribeiro et al., 2020; Moradi and Samwald, 2021). A reliable NLP sys- tem should not be easily fooled by slight noise in the text. Although a wide range of evaluation approaches for robust NLP models have been pro- posed (Ribeiro et al., 2020; Morris et al., 2020; Goel et al., 2021; Wang et al., 2021), few attempts have been made to understand these benchmark results. Given the difference of robustness be- tween models and perturbations, it is a natural question why models are more sensitive to some perturbations than others. It is crucial to avoid over-sensitivity to input perturbations, and under- standing why it happens is useful for revealing the weaknesses of current models and designing more robust training methods. To the best of our knowledge, a quantitative measure to interpret the robustness of NLP models to textual perturbations has yet to be proposed. To improve the robust- ness under perturbation, it is common practice to leverage data augmentation (Li and Specia, 2019; Min et al., 2020; Tan and Joty, 2021). Similarly, how much data augmentation through the pertur- bation improves model robustness varies between models and perturbations. In this work, we aim to investigate two Research Questions (RQ): • RQ1: Why are NLP models less robust to some perturbations than others? • RQ2: Why does data augmentation work bet- ter at improving the …

Content truncated to 500 words