kedz / styleeq

Code for Low-Level Linguistic Controls for Style Transfer and Content Preservation

styleeq

Code for Low-Level Linguistic Controls for Style Transfer and Content Preservation

INSTALL LIBRARY

First install plum:

$ python setup.py install

Then download the spacy model data:

$ python -m spacy download en_core_web_sm

Then setup the eval scripts:

$ cd eval_scripts; ./install.sh; cd ..

DOWNLOAD DATA

./download_data.sh

Data will apear in a directory called literary_style_data.

DOWNLOAD MODELS

Instead of training your own models, you can use our models we used for the paper by downloading them. Simply run:

./download_models.sh

Models will appear in a directory called style_models.

Train

To train the StyleEQ model run:

$ plumr configs/styleeq.jsonnet --proj style_models/styleeq --run train --gpu GPUNUM

where GPUNUM is the number of the gpu you want to run on. -1 will run on cpu but this is not practical.

To evaluate the model with automatic quality metrics on the test, after training, run:

$ plumr configs/styleeq.jsonnet --proj style_models/styleeq --run eval-test --gpu GPUNUM

Generation Example

To see how to generate text/perform style transfer see the example generation script, generation_example.py.

This gives an example of generating from data already in the format of the jsonl data that you can download above, and how to convert a raw string to that format. To convert a raw string to the correct format requires the Stanford Core NLP library. To download run:

$ wget http://nlp.stanford.edu/software/stanford-corenlp-full-2018-10-05.zip

$ unzip stanford-corenlp-full-2018-10-05.zip

To run the generation script, run:

$ python generation_example.py style_models/styleeq/ literary_style_data stanford-corenlp-full-2018-10-05