paper link

Since the paper wasn’t accepted and I was quite busy that year, I ended up abandoning this work. Also, this was one of my earlier projects, so the code isn’t very well organized — I apologize for that.

• To do style transfer for yelp dataset

  cd method/yelp and run python main.py --attack --not_nll --pos_victim_lr=0.2 --neg_victim_lr=0.2

• The outputs are stored in the outputs folder, the outputs sentences (*_out) are automatically exrtacted from the log files (*_log). To reimplement the extraction process, run python get_i_ter.py --file=FILENAME --pos_c=POS_C --neg_c=NEG_C, where:

  • FILENAME is the log file to extract from
  • POS_C and NEG_C are two hyperparameters, specifying different values results in different output file

  For example, cd outputs and run python get_i_iter.py --file=yelp_log --pos_c=2 --neg_c=2

@misc{fan2022gradientguidedunsupervisedtextstyle,
      title={Gradient-guided Unsupervised Text Style Transfer via Contrastive Learning}, 
      author={Chenghao Fan and Ziao Li and Wei wei},
      year={2022},
      eprint={2202.00469},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2202.00469}, 
}