Local Structure Matters Most: Perturbation Study in NLU

Findings of the Association for Computational Linguistics: ACL 2022, pages 3712 - 3731 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Local Structure Matters Most: Perturbation Study in NLU Louis Clouâtre1,3 Prasanna Parthasarathi2,3 Amal Zouaq1 and Sarath Chandar 1,3,4 1 Polytechnique Montréal 2 School of Computer Science, McGill University 3 Quebec Artificial Intelligence Institute (Mila) 4 Canada CIFAR AI Chair Abstract Recent research analyzing the sensitivity of nat- ural language understanding models to word- order perturbations has shown that neural mod- els are surprisingly insensitive to the order of words. In this paper, we investigate this phe- nomenon by developing order-altering pertur- bations on the order of words, subwords, and characters to analyze their effect on neural mod- els’ performance on language understanding tasks. We experiment with measuring the im- pact of perturbations to the local neighborhood of characters and global position of characters in the perturbed texts and observe that pertur- bation functions found in prior literature only affect the global ordering while the local or- dering remains relatively unperturbed. We em- pirically show that neural models, invariant of their inductive biases, pretraining scheme, or the choice of tokenization, mostly rely on the local structure of text to build understanding and make limited use of the global structure. 1 Introduction Recent research has shown that neural language models have an understanding of well-formed En- glish syntax in recurrent neural networks, convo- lutional neural networks, and in large pretrained (PT) Transformers (Gulordava et al., 2018; Zhang and Bowman, 2018; Chrupała and Alishahi, 2019; Lin et al., 2019a; Belinkov and Glass, 2019; Liu et al., 2019a; Jawahar et al., 2019; Rogers et al., 2020). Other studies, however, take a critical stance with experiments suggesting that models may be in- sensitive to word-order perturbations (Pham et al., 2021; Sinha et al., 2021, 2020; Gupta et al., 2021; O’Connor and Andreas, 2021), showing that shuf- fled word-order has little to no impact during training or inference with neural language mod- els. While some research show that models learn some abstract notion of syntax, further probing into their insensitivity to the perturbation of syntax is necessary. Specifically, What are the underlying mechanisms causing those unintuitive, or unnat- ural, results from neural models is still a largely unanswered question. Recent research exploring the sensitivity to syn- tax of pretrained models has primarily been apply- ing perturbations to text through perturbing the or- der of words (Pham et al., 2021; Sinha et al., 2021, 2020; Gupta et al., 2021; O’Connor and Andreas, 2021). Perturbations applied and quantified at this granularity of text offer only a limited understand- ing of the learning dynamics of the neural language models. Analyzing perturbations at a finer granular- ity such as subwords (Bojanowski et al., 2017) or characters (Gao et al., 2018; Ebrahimi et al., 2018), may provide a deeper insight into the insensitivity to word-order of neural models. In this paper, we define two types of structure1 in text, global which relates to the absolute posi- tion of characters, and …

Content truncated to 500 words