Direct-Style-Transfer

Code for "On Learning Text Style Transfer with Direct Rewards", NAACL 2021.

Libraries

We use PyTorch 1.2.0 for our experiments. The dependencies are specified in requirements.txt

Data

Yelp and Amazon

Please download the data at Yelp and Amazon and put the downloaded data at ./yelp and ./amazon.

IMDb

Please download the data at IMDb and put the downloaded data at ./imdb.

GYAMC

Please download the data at GYAMC and put the downloaded data at ./formality_family. (We use Family & Relationships category for our experiments)

Preprocessing

Please rename the downloaded files following the format {sentiment, formality}.{train, dev, test}.{0, 1}.txt.

For example, sentiment.test.0.txt contains the negative samples in the test set.

Code

We use huggingface Transformers in our experiments.

bootstrap.py - First Stage Training

To run the experiment, you may specify the hyperparameters in config dict, and run

python bootstrap.py --cuda --gpuid [GPUID] -l 

main.py - Second Stage Training

To run the experiment, you may specify the hyperparameters in config dict, and run

python main.py --cuda --gpuid [GPUID] -l -s -r -p -u

evaluate.py - Evaluation Functions

To run the experiment, please run

python evaluate.py --cuda --gpuid [GPUID] --file [OUTPUT_FILE_NAME] --dataset [DATASET_NAME] --model_pt [MODEL_CHECKPOINT]

classifier.py - classifiers for training

Dataloader.py - dataloaders

gpt_utils.py - modified transformers function to enable approximate word indexes

ref_sim.py, sim_models.py, sim_utils.py - code for SIM model

utils.py - utility functions

Results

The results of our models can be found in the ./output directory.

Each line of the files contains the source, reference and model output, seperated by \t (source and model output only for IMDb dataset).

Due to licensing restrictions, we only provide the model outputs for GYAMC dataset.