This repository is contains code for reproducing the results in Fine-grained Text Style Transfer with Diffusion-Based Language Models.

The majority of code from this repository is taken or adapted from DiffuSeq repository. Please see their repository for environment configuration and package installation instructions.

We included the processed data of StylePTB under the dataset folder. Each subfolder is either a single style transfer (such as "VEM" for verb emphasis), the combined dataset for multitask training ("multitask"), or compositional style transfers (such as "Tense + PP Removal").

To train a model on a particular dataset, first go into the scripts directory with cd scripts. Then, in train.sh, edit the field "data_dir" to your desired dataset (for example, to train on the transfer "verb emphiasis", set data_dir to ".../datasets/VEM"). Also edit "dataset" and "notes" to the desired dataset name. Then you can start training by bash train.sh.

When training is done, trained model checkpoints will show up under diffusiom_models/. To run inference, edit run_decode.sh to include your desired dataset as well as the checkpoint you want to use. Then, run bash run_decode.sh. The result will be under generation_outputs/ folder.

To evaluate the results, use tohypref.py to convert an output json under generation_outputs/ into hypothesis+reference. Then use nlgeval to evaluate metrics between ../hyp.txt and ../ref.txt using

nlg-eval --hypothesis ../hyp.txt --references ../ref.txt