Awesome

Domain Adaptive Text Style Transfer

Introduction

This is a tensorflow implementation of Domain Adaptive Text Style Transfer by Dianqi Li, Yizhe Zhang, Zhe Gan, Yu Cheng, Chris Brockett, Ming-Ting Sun and Bill Dolan, EMNLP 2019.

Environment

The code is based on python 3.6 and tensorflow 1.12 version. The code is developed and tested using one NVIDIA GTX 1080Ti.

Data Format

• Data should save in: /data/${dataset}/train/*.txt, /data/${dataset}/valid/*.txt, /data/${dataset}/test/*.txt.
• Taking yelp dataset as an example, please use the below script to generate corresponding format:

import codecs
import json

# SAVE DATA
write_train_file = codecs.open('/data/yelp/train.txt', "w", "utf-8")
dict = {"review": line.strip(), "score": score, "other_field_you_want": xxx}
string_ = json.dumps(dict)
write_train_file.write(string_ + '\n')

# LOAD DATA
reader = codecs.open('/data/yelp/train.txt', 'r', 'utf-8')
while True:
    string_ = reader.readline()
    if not string_: break
    dict_example = json.loads(string_)
    review = dict_example["review"]
    score = dict_example["score"]

• In each line of train.txt, the format will look like: {"review": "michael is absolutely wonderful .", "score": 1, "something you want"}.

Run

In this repo: TARGET_DATASET={yelp, amazon}; SOURCE_DATASET={filter_imdb}; NETWORK={CrossAlign, ControlGen}; DA_NETWORK={DAST, DASTC}. For more configurations, please see config.py.

1. Train binary style classifier:

python train_classifier.py --dataset ${TARGET_DATASET}

2. Train domain classifier:

python train_domain_classifier.py --domain_adapt --dataset ${TARGET_DATASET} --source_dataset ${SOURCE_DATASET}

3. Train style transfer model on the target domian only:

python train_style_transfer.py --dataset ${TARGET_DATASET} --network ${NETWORK}

4. Train styel transfer model with domain adaptation:

python train_domain_adapt.py --domain_adapt --dataset ${TARGET_DATASET} --source_dataset ${SOURCE_DATASET} --network ${DA_NETWORK} --training_portion ${TARGET_DATASET_PORTION}

All logs, tensorboard, generated texts will appear in logs/

5. Evaluation on generated samples

python evaluation.py --domain_adapt --dataset ${TARGET_DATASET} --source_dataset ${SOURCE_DATASET}

Note: You need to define the evaluation path folder_path by hand in evaluation.py file. Sometimes, the pos/neg samples order may be switched, you need to change line 21-22 in the file. We provide our testing results in samples folder for future comparison.

Citing

if you find our work is useful in your research, please consider citing:

@InProceedings{li2019domian,
  author = {Li Dianqi and Zhang Yizhe and Gan Zhe and Cheng Yu and Brockett Chris and Sun Ming-Ting and Dolan Bill},
  title     = {Domain Adaptive Text Style Transfer},
  booktitle = {In Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  year      = {2019}
}