afcarl / clweadv Goto Github PK

Code for "Towards cross-lingual distributed representations without parallel text trained with adversarial autoencoders"

Antonio Valerio Miceli Barone

arXiv: https://arxiv.org/abs/1608.02996

Requires:

Theano http://deeplearning.net/software/theano/
Lasagne https://lasagne.readthedocs.io/en/latest/
scikit-learn http://scikit-learn.org/stable/
Word embeddings for two languages in word2vec https://code.google.com/archive/p/word2vec/ text format

This package includes code to run various experiments with different variants of the approach presented in the paper. The best looking results were obtained with the adversarial autoencoder with Resnet discriminator, cosine distance reconstruction loss and pairwise cosine distance matching:

English to German: emb_lin_adversarial_resnet_cos_autoenc_cos_en2de.py
German to English: emb_lin_adversarial_resnet_cos_autoenc_cos_de2en.py
English to Italian: emb_lin_adversarial_resnet_cos_autoenc_cos_en2it.py

This code is distributed under the GNU LGPLv3.