Paper | Video | Slides | Poster
This repository contains information about the Context Aware Polite Paraphrase (CAPP) dataset - a dialogue-style corpus of rude utterances and corresponding polite paraphrases, with samples accompanied by additional context in the form of prior turns from the dialogue. The original paper Demonstrations Are All You Need: Advancing Offensive Content Paraphrasing using In-Context Learning was accepted in ACL 2024 Findings. We also provide the generated paraphrases for the different methods described in the paper.
-
The train split has 7939 samples and the test split has 1120 samples. Each sample will always have a rude utterance and a polite paraphrase. About 55% of the train data and 53% of the test data also has prior dialogue turn information to provide additional context.
-
The following notebook loads and describes the training and test data splits: notebook
-
For samples with multiple dialogue turns, each dialogue turn will be separated using the [SEP] separator. For example:
How long have we been here? [SEP] It's been 2 days and 7 hours. [SEP] What will happen to us?
We provide the inoffensive paraphrases generated using the below listed models for each dataset explored in the paper.
- text-davinci-003
- gpt-3.5-turbo-instruct
- gpt-3.5-turbo-0613
- gpt-3.5-turbo-1106
- Vicuna-13b
@article{som2023demonstrations,
title={Demonstrations Are All You Need: Advancing Offensive Content Paraphrasing using In-Context Learning},
author={Som, Anirudh and Sikka, Karan and Gent, Helen and Divakaran, Ajay and Kathol, Andreas and Vergyri, Dimitra},
journal={arXiv preprint arXiv:2310.10707},
year={2023}
}
Please report any issues to Github Issues.
For any questions, please contact: Anirudh Som ([email protected])
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001122C0032. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views or policies of DARPA, the Department of Defense or the U.S. Government.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
