Balancing Training for Multilingual Neural Machine Translation

Implementation of the paper

Balancing Training for Multilingual Neural Machine Translation

Xinyi Wang, Yulia Tsvetkov, Graham Neubig

Data:

The preprocessed and binarized data for fairseq can be downloaded here

To process data from scrach, see the script

util_scripts/prepare_multilingual_data.sh

Training Scripts:

The training scripts for many-to-one translation of the related language group (Related M2O) is under the directory job_scripts/related_ted8_m2o/.

Our methods:

MultiDDS-S:

job_scripts/related_ted8_m2o/multidds_s.sh 

MultiDDS:

job_scripts/related_ted8_m2o/multidds.sh 

Baselines:

Proportional:

job_scripts/related_ted8_m2o/proportional.sh 

Temperature:

job_scripts/related_ted8_m2o/temperature.sh 

The scripts for Related O2M is under the directory job_scripts/related_ted8_o2m/

The scripts for Diverse M2O is under the directory job_scripts/diverse_ted8_m2o/

The scripts for Diverse O2M is under the directory job_scripts/diverse_ted8_o2m/

Inference Scripts:

Each of the experiment script directory contains a trans.sh file to translate the test set. To translate the test set for the Related M2O MultiDDS-S

job_scripts/related_ted8_m2o/trans.sh checkpoints/related_ted8_m2o/multidds_s/ 

To translate other experiment, simply replace the argument with the experiment checkpoint directory.

Citation

Please cite as:

@inproceedings{wang2020multiDDS,
  title = {Balancing Training for Multilingual Neural Machine Translation},
  author = {Xinyi Wang, Yulia Tsvetkov, Graham Neubig},
  booktitle = {ACL},
  year = {2020},
}