This repo contains the code and data of the following paper:
"Style Transfer from Non-Parallel Text by Cross-Alignment". Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. NIPS 2017. arXiv
The method learns to perform style transfer between two non-parallel corpora. For example, given positive and negative reviews as two corpora, the model can learn to reverse the sentiment of a sentence.
Please name the corpora of two styles by "x.0" and "x.1" respectively, and use "x" to refer to them in options. Each file should consist of one sentence per line with tokens separated by a space.
data/yelp/ directory contains an example Yelp review dataset.
To train a model, first create a
tmp/ folder (where the model and results will be saved), then go to the
code/ folder and run the following command:
bash python style_transfer.py --train ../data/yelp/sentiment.train --dev ../data/yelp/sentiment.dev --output ../tmp/sentiment.dev --vocab ../tmp/yelp.vocab --model ../tmp/model
To test the model, run the following command:
bash python style_transfer.py --test ../data/yelp/sentiment.test --output ../tmp/sentiment.test --vocab ../tmp/yelp.vocab --model ../tmp/model --load_model true --beam 8
To download a trained model, run
bash download_model.sh, and then run the testing command with
--model options specifying
code/options.py for all running options.
Python >= 2.7, TensorFlow 1.3.0