Cross-model Back-translated Distillation for Unsupervised Machine Translation

Accepted as conference paper at 38th International Conference on Machine Learning (ICML 2021).

Authors: Xuan-Phi Nguyen, Shafiq Joty, Thanh-Tung Nguyen, Wu Kui, Ai Ti Aw

Github

Paper link: https://arxiv.org/abs/1911.01986

Citation

Please cite as:

@incollection{nguyen2021cbd,
title = {Cross-model Back-translated Distillation for Unsupervised Machine Translation},
author = {Xuan-Phi Nguyen and Shafiq Joty and Thanh-Tung Nguyen and Wu Kui and Ai Ti Aw},
booktitle = {38th International Conference on Machine Learning},
year = {2021},
}

These guidelines demonstrate the steps to run CBD on the WMT En-De

Finetuned model

Model Train Dataset Finetuned model
WMT En-Fr WMT English-French model: download
WMT En-De WMT English-German model: download

0. Installation

./install.sh
pip install fairseq==0.8.0 --progress-bar off

1. Prepare data

Following instructions from MASS-paper to create WMT En-De dataset.

2. Prepare pretrained model

Download XLM finetuned model (theta_1): here, save it to bash variable export xlm_path=...

Download MASS finetuned model (theta_2): here, save it to export mass_path=....

Download XLM pretrained model (theta): here, save it to export pretrain_path...

3. Run CBD model


# you may change the inputs in the file according to your context
bash run_ende.sh