Self-Training-Dialogue-Generation

This repository contains the data and code for the paper:

Self-training with Two-phase Self-augmentation for Few-shot Dialogue Generation
Wanyu Du, Hanjie Chen and Yangfeng Ji
Findings of EMNLP 2022

Here are a few data examples generated by our method:

Environment Setup

1. Install and setup the environment:

conda env create -f environment.yml
conda activate sa2st

2. Download the pretrained scgpt checkpoint, put it under src:

wget https://bapengstorage.blob.core.windows.net/fileshare/scgpt.tar.gz
tar xvf scgpt.tar.gz
mv scgpt src/

Datasets

1. Data statistics for FewShotWoz

2. Data statistics for FewShotSGD

Models

Inference Only

1. Download the model checkpoints on 2 datasets across all domains (check here).
2. Put the downloaded checkpoints under models.
3. Run the following script to obtain our models' predictions on dev and test set:

bash sa2_infer.sh {DATASET} {DOMAIN} {CUDA}

{DATASET} indicates the dataset, including FewShotWoz, FewShotSGD.
{DOMAIN} indicates the domain, including attraction, hotel, laptop, ....
{CUDA} indicates the index of the GPU (e.g. 0, 1, 2, 3, ...).

Self-training and Inference

If you want to train the model from scratch, you can run the following script:

bash sa2_woz.sh {DOMAIN} {FLAG} {LR} {EPOCH} {CUDA}

{DOMAIN} indicates the domain, including attraction, hotel, laptop, ...
{FLAG} indicates the model checkpoint name flag.
{LR} indicates the learning rate.
{EPOCH} indicates the training epoch.
{CUDA} indicates the index of the GPU (e.g. 0, 1, 2, 3, ...).

Evaluation

The evaluation srcipt is included in sa2_inference.py and sa2_self_training.py.
The evaluation scores on dev and test set will be saved under the current model checkpoint folder.

Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{du2022sa,
    title = {Self-training with Two-phase Self-augmentation for Few-shot Dialogue Generation},
    author = {Du, Wanyu and Chen, Hanjie and Ji, Yangfeng},
    journal = {arXiv preprint arXiv:2204.01227},
    year = {2022},
    doi = {10.48550/ARXIV.2205.09661}, 
    url = {https://arxiv.org/abs/2205.09661},  
}