Grammatical Error Correction Systems for Automated Assessment: Are They Susceptible to Universal Adversarial Attacks?

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 158–171 November 20–23, 2022. ©2022 Association for Computational Linguistics 158 Grammatical Error Correction Systems for Automated Assessment: Are They Susceptible to Universal Adversarial Attacks? Vyas Raina Cambridge University vr313@cam.ac.uk Edie Lu Cambridge University ytl28@cam.ac.uk Mark Gales Cambridge University mjfg@cam.ac.uk Abstract Grammatical error correction (GEC) systems are a useful tool for assessing a learner’s writ- ing ability. These systems allow the grammat- ical proficiency of a candidate’s text to be as- sessed without requiring an examiner or teacher to read the text. A simple summary of a can- didate’s ability can be measured by the total number of edits between the input text and the GEC system output: the fewer the edits the better the candidate. With advances in deep learning, GEC systems have become increas- ingly powerful and accurate. However, deep learning systems are susceptible to adversar- ial attacks, in which a small change at the in- put can cause large, undesired changes at the output. In the context of GEC for automated assessment, the aim of an attack can be to de- ceive the system into not correcting (conceal- ing) grammatical errors to create the perception of higher language ability. An interesting as- pect of adversarial attacks in this scenario is that the attack needs to be simple as it must be applied by, for example, a learner of English. The form of realistic attack examined in this work is appending the same phrase to each in- put sentence: a concatenative universal attack. The candidate only needs to learn a single at- tack phrase. State-of-the-art GEC systems are found to be susceptible to this form of simple attack, which transfers to different test sets as well as system architectures 1. 1 Introduction Grammatical Error Correction (GEC) systems can form a part of automated language fluency assess- ment: the number of edits from a candidate’s in- put sentence to a GEC system’s grammatically corrected output sentence is indicative of a can- didate’s language ability, where fewer edits sug- gest better fluency. Early GEC systems were designed using hand-crafted rules (Naber, 2003), 1Code is available at: https://github.com/ rainavyas/gec-universal-attack but since, data driven approaches, such as Sta- tistical Machine Translation (Yuan and Felice, 2013), emerged. With encoder-decoder architec- tures dominating in Neural Machine Translation, Yuan and Briscoe (2016) used Recurrent Neural Networks (Cho et al., 2014) to improve GEC perfor- mance. Now state of the art GEC systems are based on the Transformer (Vaswani et al., 2017) architec- ture (Kaneko et al., 2020; Chen et al., 2020; Malmi et al., 2019; Awasthi et al., 2019; Omelianchuk et al., 2020b; Kiyono et al., 2019; Lichtarge et al., 2020; Stahlberg and Kumar, 2020). Despite the success of Transformer-based deep learning systems, there is a shortcoming: Szegedy et al. (2014) discovered that neural networks are susceptible to adversarial attacks, where a small change at the …

Content truncated to 500 words