PLACES: Prompting Language Models for Social Conversation Synthesis

Citation:

@inproceedings{chen-etal-2023-places,
    title = "{PLACES}: Prompting Language Models for Social Conversation Synthesis",
    author = "Chen, Maximillian  and
      Papangelis, Alexandros  and
      Tao, Chenyang  and
      Kim, Seokhwan  and
      Rosenbaum, Andy  and
      Liu, Yang  and
      Yu, Zhou  and
      Hakkani-Tur, Dilek",
    booktitle = "Findings of the Association for Computational Linguistics: EACL 2023",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-eacl.63",
    pages = "844--868",
}

PLACES-GPT3.5 Quick Info

Below is a version of the dyadic data generated using PLACES with GPT 3.5-Turbo as the backbone:

https://raw.githubusercontent.com/maxlchen/PLACES-GPT3.5/main/PLACES-GPT3.5-Dyadic.jsonlist

A multiparty version of the data is coming shortly!

PLACES-GPT3.5 is also featured in DialogStudio: https://github.com/salesforce/DialogStudio#loading-data

Quick Start

This code can be used to recreate the conversations from the paper, which used a list of reference topics from FITS. Feel free to try out different topic and prompt inputs!

Follow these steps to generate synthetic conversations with PLACES.

1. Download Data

PLACES can use existing conversations as examples in the prompt, or hand-crafted ones. In the paper, we use 3 datasets: TopicalChat [1], DailyDialog [2], and FITS [3].

DailyDialog

Download from here.

Topical Chat

Download from here.

FITS

Download FITS data from ParlAI.

First, you need to install ParlAI: pip install parlai

Using: parlai display_data -t fits will tell you where the FITS data is stored on your local machine. It may take some time to download the data the first time you call it.

2. Set up your environment

We've tested our code with Python 3.8 and transformers 4.26.0 but it should work with earlier versions of transformers too. After creating a virtual environment, you can install the requirements:

pip install transformers

3. Run PLACES

If you want to use Topical Chat conversations as prompts, you need to first parse Topical-Chat into a simpler format:

python parse_topical_chat.py --tc_path <PATH_TO_TOPICAL_CHAT>

This will produce a .jsonlist file into the prompts directory.

The general command to run PLACES is:

python conversation_synthesis.py <ARGUMENTS>

For example:

python conversation_synthesis.py --fits_path <PATH_TO_FITS>                                 

Or:

python conversation_synthesis.py --fits_path <PATH_TO_FITS>
                                 --in_context_dataset "daily_dialog"
                                 --in_context_dataset_path <PATH_TO_DAILY_DIALOG>                                 

If you want to run triadic conversations, use the --triadic flag:

For example:

python conversation_synthesis.py --fits_path <PATH_TO_FITS>       
                                 --triadic                          

Or:

python conversation_synthesis.py --fits_path <PATH_TO_FITS>
                                 --in_context_dataset "daily_dialog"
                                 --in_context_dataset_path <PATH_TO_DAILY_DIALOG>
                                 --triadic                                 

While we haven't tested multi-party conversations with more than 3 participants, it should be possible to do so by creating the appropriate prompts in utils.py.

References

1. Karthik Gopalakrishnan, Behnam Hedayatnia, Qinlang Chen, Anna Gottardi, Sanjeev Kwatra, Anushree Venkatesh, Raefer Gabriel, Dilek Hakkani-Tür, Topical-Chat: Towards knowledge-grounded open-domain conversations, Interspeech 2019
2. Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset. IJCNLP 2017.
3. Xu J, Ung M, Komeili M, Arora K, Boureau YL, Weston J. Learning New Skills after Deployment: Improving open-domain internet-driven dialogue with human feedback. arXiv preprint 2022.