jiezhanggt / cdalign

Code for AAAI 2021 paper "Lexically Constrained Neural Machine Translation with Explicit Alignment Guidance"

Code for AAAI 2021 paper "Lexically Constrained Neural Machine Translation with Explicit Alignment Guidance". [Paper]

Install and Data preprocess

The code is implemented on fairseq v0.9.0, follow the same steps to install and prepare the processed fairseq dataset with script here. You may need to process other datasets similarly. The python package fastbpe is also needed.

git clone https://github.com/ghchen18/cdalign.git
cd cdalign
pip install --editable ./

Step 1: Train vanilla transformer

See scripts/run_train.sh

Step 2: Extract alignment using Att-Input method and process alignment data

See scripts/extract_alignment.sh

Step 3: Train with EAM-Output method

See scripts/run_train.sh

Step 4: Test on lexically constrained NMT task

See scripts/run_test.sh

Citation

@inproceedings{chen2021lexically,
  title={Lexically Constrained Neural Machine Translation with Explicit Alignment Guidance},
  author={Guanhua, Chen and Yun, Chen and Victor O.K., Li},
  booktitle = {Proceedings of AAAI},
  year      = {2021},
  pages  = {12630--12638},
  volume={35},
  number={14},
}