Evaluating the Robustness of Discrete Prompts

Yoichi Ishibashi, Danushka Bollegala, Katsuhito Sudoh, Satoshi Nakamura: Evaluating the Robustness of Discrete Prompts (EACL 2023)

Setup

Install the required packages.

pip install -r requirements.txt

Usage

Our experiment is divided into two phases (1) prompt learning (2) analyzing the robustness of the learned prompts.

1. Learning prompt tokens by AutoPrompt (AP).

cd ap
sh ap_label-token-search.sh
sh ap_trigger-token-search.sh

2. Fine-tuning PLM by Manually-written Prompts (MP).

cd mp
sh mp_finetuning.sh

3. Evaluating the robustness of LM prompt The following scripts perform the four robustness evaluations of LM prompts.

AutoPrompt (AP)

cd ap
sh ap_run-all-robust-eval.sh 

Manually-written Prompts (MP)

cd mp
sh mp_run-all-robust-eval.sh 

The adversarial NLI dataset

We created the adversarial NLI dataset (see Sec 3.5 Adversarial Perturbations in our paper). These datasets were used for the prompt robustness evaluation described above.

data/superglue/cb/perturbation-label-change.tsv
data/superglue/cb/perturbation-label-no-change.tsv
data/superglue/mnli/perturbation-label-change.tsv
data/superglue/mnli/perturbation-label-no-change.tsv

External Libraries

• OpenPrompt
• autoprompt

Citation

@inproceedings{Ishibashi:EACL:2023,
  author = {Yoichi Ishibashi and Danushka Bollegala and Katsuhito Sudoh and Satoshi Nakamura},  
  title = {Evaluating the Robustness of Discrete Prompts},
  booktitle = {Proc. of  the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2023)},
  year = {2023}
}