On the Copying Behaviors of Pre-Training for Neural Machine Translation (Findings of ACL 2021)

Citation

Please cite as:

@inproceedings{liu2021copying,
  title={On the Copying Behaviors of Pre-Training for Neural Machine Translation},
  author={Liu, Xuebo and Wang, Longyue and Wong, Derek F and Ding, Liang and Chao, Lidia S and Shi, Shuming and Tu, Zhaopeng},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2021},
  year={2021}
}

Requirements and Installation

This implementation is based on fairseq(v0.9.0)

• PyTorch version >= 1.2.0
• Python version >= 3.6

git clone https://github.com/SunbowLiu/CopyingPenalty
cd CopyingPenalty
pip install --editable .

Addtional Parameters

The copying penalty can be applied to both vanilla sequence-to-sequence learning models (--task translation) and (m)BART-initialized models (--task translation_from_pretrained_bart).
Please refer to fairseq and mBART for the training of models.

Features

1. As simple and powerful as the length penalty.
2. Particularly effective in the task of similar input and output domains.
3. Trivial computational cost.

parameter	description
--copypen	Copying penalty: <1.0 favors copying source tokens, >1.0 favors translating (converting) source tokens; Default: 1.0
--target-dictionary-path	Path to target side dictionary for automatically obtaining the punctuation list.

Mainly modified code: fairseq/sequence_generator.py

Tune a copying penalty on the validation set

mkdir $PATH_TO_OUTPUT/validation
for cp in 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5
do
RESULT=$PATH_TO_OUTPUT/validation/$cp.txt
python generate.py \
    $PATH_TO_DATA \
    --gen-subset valid \
    --path $PATH_TO_OUTPUT/checkpoint_best.pt \
    --copypen $cp --target-dictionary-path $PATH_TO_TGT_DICTIONARY \
    > $RESULT
done

Apply the tuned copying penalty to the test set

python generate.py \
    $PATH_TO_DATA \
    --gen-subset test \
    --path $PATH_TO_OUTPUT/checkpoint_best.pt \
    --copypen $cp --target-dictionary-path $PATH_TO_TGT_DICTIONARY