You Only Prompt Once: On the Capabilities of Prompt Learning on Large Language Models to Tackle Toxic Content
This is the official implementation of our paper You Only Prompt Once: On the Capabilities of Prompt Learning on Large Language Models to Tackle Toxic Content.
conda env create --file environment.yaml &&
conda activate toxic_prompt
Datasets we used in this paper are in parsed_dataset folder.
# Task 1
Dataset: ["HateXplain", "USElectionHate20", "HateCheck", "SBIC.v2", "measuring-hate-speech"]
# Task 2:
Dataset: ["TSD"]
# Task 3:
Dataset: ["Parallel", "Paradetox"]
# Train the prompt tuning model.
python 1_toxicity_classification.py --plm_eval_mode --model t5 --model_name_or_path t5-small --dataset HateXplain
Train the prompt and perform the generation task:
# Example: TSD dataset with t5-small model.
python 2_and_3_toxic_generation.py --plm_eval_mode --model t5 --model_name_or_path t5-small --dataset TSD
Evaluation:
python 2_calculate_span.py --file_path sfs_out/task23/TSD_t5-small_True.txt
For the baseline methods, we follow the implementation from toxic-span.
Note that Task 2 and Task 3 share the same code as they are both generation tasks.
Train the prompt and perform the generation task:
# Example: Parallel dataset with t5-small model.
python 2_and_3_toxic_generation.py --plm_eval_mode --model t5 --model_name_or_path t5-small --dataset Parallel
Evaluation for toxicity change:
python 3_perspective_evaluation.py --file_path sfs_out/task23/Parallel_t5-small_True.txt --key YOUR_PERSPECTIVE_API_KEY
Note that for the baseline methods and evaluation regarding other metrics (e.g., BLEU, SIM, PPL), we follow the implementation of paradetox.
You can download the model weights and replace the saved_model/ folder from this Link.
Please cite our paper if you use this code in your own work:
@inproceedings{HZSZ224,
author = {Xinlei He and Savvas Zannettou and Yun Shen and Yang Zhang},
title = {{You Only Prompt Once: On the Capabilities of Prompt Learning on Large Language Models to Tackle Toxic Content}},
booktitle = {{IEEE Symposium on Security and Privacy (S\&P)}},
publisher = {IEEE},
year = {2024}
}