PRAL

Code for the paper: A Tailored Pre-Training Model for Task-Oriented Dialog Generation

Pretrain Dataset

For the pretrain dataset, first download the repository.

# download
git clone https://github.com/qywu/DialogCorpus.git
cd DialogCorpus

You can manually download and process the dataset.

# download data for daily_dialog
python daily_dialog/download_data.py
# process the data
python daily_dialog/process_data.py
# the processed data is stored as the {folder_name}.json
vi daily_dialog/data/daily_dialog.json

Or you can just use one command.

python prepare_all_data.py \
       --download \
       --process \
       --join

Or you can just download our processed version: https://drive.google.com/file/d/1VS9GndEAsrdiyIzlyhy2LAKyu_bR2Lpz/view?usp=sharing

Detailed Dialog Processing for each dataset:

• Daily Dialog

  • Removed tokenization space for punctuations

• Persona Chat

  • Used huggingface's version [link]
  • Recovered lower cased utterances
  • Removed tokenization space for punctuations

• Cornell Movie Corpus

  • Ignored UTF-8 Errors
  • Extracted Names

• Task Master

  • Nothing specific

• CCPE

  • Nothing specific

• Frames

  • Nothing specific

• Chit-Chat Challenge

  • Nothing specific

• Self-dialogue

  • Nothing specific

• Schema Dialog

  • Nothing specific

Links

• Daily Dialog [link]

• Conversational flow in Oxford-style debates [link]

• Persona-chat [Google Drive]

Training

After process or download the data, put dialog_corpus.json in the current directory and train the model with the following:

python main.py

Evaluation

You can refer to ARDM's evaluation code https://github.com/qywu/ARDM. For the chatbot demo, you can checkout the colab example and load the pretrained weights: https://colab.research.google.com/drive/1ib7YCeNhkIDAzuOKotSlw1CfIBP_zE4r

Pretrained Weights

We provide the download option to our pretrained weights: https://drive.google.com/file/d/17S0TYjbUQmMzsNvfgZwY2DFULYlPQZ7h/view?usp=sharing

Citation

You can cite the paper with:

@article{PRAL,
  author    = {Jing Gu and
               Qingyang Wu and
               Chongruo Wu and
               Weiyan Shi and
               Zhou Yu},
  title     = {A Tailored Pre-Training Model for Task-Oriented Dialog Generation},
  journal   = {CoRR},
  volume    = {abs/2004.13835},
  year      = {2020},
  url       = {https://arxiv.org/abs/2004.13835},
  archivePrefix = {arXiv},
  eprint    = {2004.13835},
  timestamp = {Sat, 02 May 2020 19:17:26 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2004-13835.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}