Target-Side Data Augmentation

This code is for ACL 2023 long paper "Target-Side Augmentation for Document-Level Machine Translation".

Paper | Poster | Slides | Video

Brief Intro

Environment

• Python3.6
• PyTorch1.9.0
• Setup the environment:

    git clone --recursive https://github.com/baoguangsheng/target-side-augmentation.git
    cd target-side-augmentation
    bash setup.sh
  

(Notes: our experiments are run on 4 GPUs of Tesla V100.)

Description of Codes

• ./G-Trans -> submodule reference to G-Transformer.
• ./scripts_gtrans -> updated G-Transformer scripts with new settings.
• ./scripts_tgtaug -> scripts for target-side augmentation.
• ./scripts_srcaug -> scripts for source-side augmentation.
• ./scripts_bothaug -> integrated scripts for both-side augmentation.

Experiment Workspace

Following folders are created for our experiments:

• ./exp_main -> main experiments for target-side augmentation.
• ./exp_main/subexp_da -> for DA model.
• ./exp_main/subexp_mt -> for MT model.
• ./exp_backtrans -> back-translation plus target-side augmentation.
• ./exp_backtrans/exp_srcaug -> source-side aug with back-translation.
• ./exp_backtrans/exp_bothaug -> further target-side aug with DA model.

(Notes: we put recent experiment logs in ./logs/*.zip for reference.)

How to Run

Baselines + Target-side augmentation

The scripts will run four steps:

1. Prepare data;
2. Sent-level augmentation on Transformer;
3. Doc-level augmentation on G-Transformer;
4. Report s/d-BLEU scores.

  CUDA_VISIBLE_DEVICES=0,1,2,3 bash scripts_tgtaug/run-all.sh nc2016 exp_main

The "nc2016" is the data name for News, which could be replaced with "iwslt17" for TED and "europarl7" for Europarl.

We place the configuration scripts for the experiment at exp_main/scripts. After running, two sub-folders will be created. One for the DA model (exp_main/subexp_da), and another for the MT model (exp_main/subexp_mt).

Back-translation + Targets-side Augmentation

The scripts will run three steps: 1) Run source-side augmentation with back-translation; 2) Initialize the data with back-translated sources for target-side augmentation; 3) Run target-side augmentation with DA model.

CUDA_VISIBLE_DEVICES=0,1,2,3 bash scripts_bothaug/run-all.sh nc2016 exp_backtrans

We treat back-translation as a special DA model with specific configuration (exp_backtrans/exp_srcaug/scripts). We put experiments with back-translation under exp_backtrans/exp_srcaug and experiments with further target-side augmentation under exp_backtrans/exp_bothaug.

Transformer and G-Transformer baselines

CUDA_VISIBLE_DEVICES=0,1,2,3 bash scripts_gtrans/run-baseline.sh nc2016 exp_main

We put the experiments under exp_main/subexp_baseline.

Citation

@inproceedings{bao2023target,
  title={Target-Side Augmentation for Document-Level Machine Translation},
  author={Bao, Guangsheng and Teng, Zhiyang and Zhang, Yue},
  booktitle={Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={10725--10742},
  year={2023}
}