Introduction

• One implementation of the paper Learning from Bootstrapping and Stepwise Reinforcement Reward: A Semi-Supervised Framework for Text Style Transfer in NAACL 2022.
• This code and data are only for research use. Please cite the papers if they are helpful.
  

Package Requirements

The model training and inference scripts are tested on following libraries and versions:

1. pytorch==1.8.1
2. transformers==4.8.2

Guideline

1. Run the script classification_style_BERT.py to obtain the style classification model, which will be used as the style discriminator when training the semi-supervised style transfer framework. Here data labeled with sentiment polairty are used (e.g., Yelp and Amazon corpora for sentiment style transfer).
2. Run the script find_sentence_Multi_Thread_Yelp.py to construct the pseudo parallel pairs, see the paper for details. We have provided the pseudo parallel pairs in this repo (pseudo_paired_data_Yelp and pseudo_paired_data_Amazon).
3. Run the script main.py to train and evaluate the semi-supervised style transfer framework. Before your run it, please check the training configurations in global_config.py.

Model Generation

See the system outputs of Yelp, Amazon, and GYAFC test sets in ./model_outputs/.

Citation

If the work is helpful, please cite our papers in your publications, reports, slides, and thesis.

@inproceedings{liu-chen-2022-learning,
    title = "Learning from Bootstrapping and Stepwise Reinforcement Reward: A Semi-Supervised Framework for Text Style Transfer",
    author = "Liu, Zhengyuan  and
      Chen, Nancy",
    booktitle = "Findings of the Association for Computational Linguistics: NAACL 2022",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-naacl.201",
    doi = "10.18653/v1/2022.findings-naacl.201",
    pages = "2633--2648",
}