Evaluating the Robustness of Discrete Prompts

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2373–2384 May 2-6, 2023 ©2023 Association for Computational Linguistics Evaluating the Robustness of Discrete Prompts Yoichi Ishibashi1 Danushka Bollegala2 Katsuhito Sudoh1 Satoshi Nakamura1 1 Nara Institute of Science and Technology 2 University of Liverpool {ishibashi.yoichi.ir3, sudoh, s-nakamura}@is.naist.jp danushka@liverpool.ac.uk Abstract Discrete prompts have been used for fine- tuning Pre-trained Language Models for di- verse NLP tasks. In particular, automatic meth- ods that generate discrete prompts from a small set of training instances have reported supe- rior performance. However, a closer look at the learnt prompts reveals that they contain noisy and counter-intuitive lexical constructs that would not be encountered in manually- written prompts. This raises an important yet understudied question regarding the robustness of automatically learnt discrete prompts when used in downstream tasks. To address this ques- tion, we conduct a systematic study of the ro- bustness of discrete prompts by applying care- fully designed perturbations into an application using AutoPrompt and then measure their per- formance in two Natural Language Inference (NLI) datasets. Our experimental results show that although the discrete prompt-based method remains relatively robust against perturbations to NLI inputs, they are highly sensitive to other types of perturbations such as shuffling and deletion of prompt tokens. Moreover, they gen- eralize poorly across different NLI datasets. We hope our findings will inspire future work on robust discrete prompt learning.1 1 Introduction Pre-trained Language Models (PLMs) have been successfully adapted to a wide range of Natural Language Processing (NLP) tasks using prompt- based learning (Radford et al., 2018, 2019; Brown et al., 2020; Petroni et al., 2019) such as sentiment classification (Gao et al., 2021), natural language inference (NLI) (Schick and Schütze, 2021, 2022), relation extraction (Shin et al., 2020), cross-lingual inference (Qi et al., 2022). However, manually writing prompts that generalize well is very chal- lenging for several reasons such as (a) it might not 1Our codes and the adversarial NLI dataset are available at https://github.com/LivNLP/prompt-robustness always be possible to recruit domain-expert human annotators, (b) human annotators might not be able to cover all corner cases by writing prompts, and (c) there can be disagreements between human annota- tors regarding the coverage of a particular prompt. To address these challenges, automatic learning of discrete prompts has been proposed such as Ad- vTrigger (Wallace et al., 2019), AutoPrompt (AP; Shin et al., 2020), WARP (Hambardzumyan et al., 2021), and RLPrompt (Deng et al., 2022). Although discrete prompt learning methods have achieved good performance in numerous down- stream tasks by automatically learnt prompts, such automatic prompts seem to be significantly differ- ent from the manually-written ones. For exam- ple, Table 1 shows manually-written and AP-learnt prompts for fact retrieval (Petroni et al., 2019). We see that the AP-learnt prompts for BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019) out- perform the manual prompts in precision1 (P@1) scores. However, the AP-learnt prompts contain various counter-intuitive language constructs such as punctuation (e.g. ‘(’, ‘?’, …

Content truncated to 500 words