wxjiao / fairseq-dro-mnmt

Distributionally Robust Multilingual Machine Translation

This repository contains codes for experiments in the following paper.

Distributionally Robust Multilingual Machine Translation
Chunting Zhou*, Daniel Levy*, Xian Li, Marjan Ghazvininejad, Graham Neubig
EMNLP 2021

Setup

• This repo is based on fairseq (tag v0.10.0) (please follow the instructions in the fairseq repo for requirements on apex, Python and Pytorch version.)

  cd fairseq; pip install --editable ./
  

Training and Inference

Slurm scripts for training and decoding on the TED58 dataset can be found under exps/. The preprocessing scripts of the datasets in the paper can be found under utils/. Please replace the data path to your own MT data if you'd like to train on a different dataset.

sbatch exps/1_baseline_t1_chi_square_0.05_ted.sh

To compute the baseline losses, exps/eval_baselines_ted.sh is an example script of how to load a pretrained ERM model for computing the average loss on the training set.

Other baseline methods

The scripts for running other baseline methods can be found under exps/ as well, similarly, please replace with your own dataset hyperparameters correspondingly.

1. ERM

   sbatch exps/2_erm.sh or bash exps/2_erm.sh (need slight modification on the env variable SLURM_ARRAY_TASK_ID).

2. CVaR group DRO with primal-dual method

   sbatch exps/3_baseline_cvar_pd.sh

3. CVaR group DRO with our iterated best response (IBR)

   sbatch exps/4_baseline_cvar_ibr.sh

4. Chi-square group DRO with primal-dual method

   sbatch exps/5_baseline_chi_square_primal_dual.sh

5. group DRO (Sagawa et al., 2020)
   sbatch exps/6_baseline_eg_reweight.sh

Reference

@inproceedings{zhou21emnlp,
    title = {Distributionally Robust Multilingual Machine Translation},
    author = {Chunting Zhou and Daniel Levy and Xian Li and Marjan Ghazvininejad and Graham Neubig},
    booktitle = {Conference on Empirical Methods in Natural Language Processing (EMNLP)},
    address = {Punta Cana, Dominican Republic},
    month = {November},
    year = {2021}
}