Will Overly Polite Sentences Harm Model Performance? Adversarial Pragmatic Perturbation for NLP

Course project in KAIST Bias and Ethics in Natural Language Processing course (AI620)


Project Summary

Non-native English speakers often struggle with controlling tones and nuances, resulting in non-canonical texts that are excessively polite or informal. In this paper, we address this issue by investigating the impact of tone perturbations on NLP models' performance. To generate tone-perturbed adversarial samples, we employ prompt engineering and in-context learning approaches, producing semantically similar yet overly polite paraphrases of original sentences. Through empirical evaluation, we demonstrate that current NLP models are susceptible to such tone-level perturbations, highlighting potential biases and challenges. Furthermore, we propose a simple augmentation-based method to enhance model robustness against adversarial samples. Our work contributes to the development of more inclusive and user-centric NLP systems by shedding light on the impact of tone variations and addressing the needs of non-native English speakers. By uncovering vulnerabilities and offering practical solutions, we strive to improve the accessibility and user experience of AI technologies in linguistic diversity.


Libraries & Frameworks

  • PyTorch
Master’s student @ KAIST EE

Seungjoo Lee