This is the code for the following article:
Privacy-preserving Neural Representation of Text
Maximin Coavoux, Shashi Narayan, Shay B. Cohen
EMNLP 2018
[pdf] [preprint] [bib]
See dependencies.txt for a list of dependencies.
Download and preprocess data (might take a long time since it also trains an LDA model on the blog dataset).
cd dataset
sh download_data.sh
cd ..
To launch an experiment:
cd src
## see python main.py --help for full description of options and list of dataset ids
python main.py <modelname> <dataset_id> ( | --atraining | --generator | --ptraining --alpha <float>)
# e.g.
python main.py mymodel tp_fr --atraining
The options to use the defense methods (during the training of the main model) that are used in the article are the following:
--atraining Adversarial classification
--generator Adversarial generation
--ptraining --alpha <float> declustering