NatLogAttack: A Framework for Attacking Natural Language Inference Models with Natural Logic

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9960–9976 July 9-14, 2023 ©2023 Association for Computational Linguistics NatLogAttack: A Framework for Attacking Natural Language Inference Models with Natural Logic Zi’ou Zheng & Xiaodan Zhu Department of Electrical and Computer Engineering & Ingenuity Labs Research Institute Queen’s University {ziou.zheng,xiaodan.zhu}@queensu.ca Abstract Reasoning has been a central topic in artifi- cial intelligence from the beginning. The re- cent progress made on distributed representa- tion and neural networks continues to improve the state-of-the-art performance of natural lan- guage inference. However, it remains an open question whether the models perform real rea- soning to reach their conclusions or rely on spurious correlations. Adversarial attacks have proven to be an important tool to help evalu- ate the Achilles’ heel of the victim models. In this study, we explore the fundamental prob- lem of developing attack models based on logic formalism. We propose NatLogAttack to per- form systematic attacks centring around nat- ural logic, a classical logic formalism that is traceable back to Aristotle’s syllogism and has been closely developed for natural language inference. The proposed framework renders both label-preserving and label-flipping attacks. We show that compared to the existing attack models, NatLogAttack generates better adver- sarial examples with fewer visits to the victim models. The victim models are found to be more vulnerable under the label-flipping set- ting. NatLogAttack provides a tool to probe the existing and future NLI models’ capacity from a key viewpoint and we hope more logic- based attacks will be further explored for un- derstanding the desired property of reasoning. 1 1 Introduction While deep neural networks have achieved the state- of-the-art performance on a wide range of tasks, the models are often vulnerable and easily deceived by imposing perturbations to the original input (Good- fellow et al., 2014; Kurakin et al., 2018), which seriously hurts the accountability of the systems. In depth, this pertains to model robustness, capac- ity, and the development of models with more ad- vanced intelligence. 1The code of NatLogAttack is available at https:// github.com/orianna-zzo/NatLogAttack. Natural language inference (NLI), also known as textual entailment (Dagan et al., 2005; Iftene and Balahur-Dobrescu, 2007; MacCartney, 2009; Bowman et al., 2015), is a fundamental problem that models the inferential relationships between a premise and hypothesis sentence. The models built on distributed representation have significantly im- proved the performance on different benchmarks (Bowman et al., 2015; Chen et al., 2017; Williams et al., 2018; Chen et al., 2018; Devlin et al., 2019; Liu et al., 2019; Zhang et al., 2020; Pilault et al., 2021). However, it is still highly desirable to con- duct research to probe if the models possess the desired reasoning ability rather than rely on spuri- ous correlation to reach their conclusions (Glock- ner et al., 2018; Poliak et al., 2018; Belinkov et al., 2019; McCoy et al., 2019; Richardson et al., 2020). Adversarial attacks have proven to be an impor- tant tool to reveal the Achilles’ heel of victim …

Content truncated to 500 words