Modeling Fluency and Faithfulness for Diverse Neural Machine Translation

PyTorch implementation of the models described in the AAAI2020 long paper Modeling Fluency and Faithfulness for Diverse Neural Machine Translation.

Related code

Implemented based on Fairseq-py, an open-source toolkit released by Facebook which was implemented strictly referring to Vaswani et al. (2017).

DiverseNMT

Download the preprocessed WMT'16 EN-DE data provided by Google and extract it and preprocess the dataset with a joined dictionary.

Train a base model.

$ sh run_ende.sh

Inference the test set.

$ python generate.py $DATA --path $SMODEL \
    --gen-subset test --beam 4 --batch-size 128 \
    --remove-bpe --lenpen 0.6 > pred.de \
# because fairseq's output is unordered, we need to recover its order
$ grep ^H pred.de | cut -f1,3- | cut -c3- | sort -k1n | cut -f2- > pred.de

Citation

If you find the resources in this repository useful, please consider citing:

@article{feng2019modeling,
  title={Modeling Fluency and Faithfulness for Diverse Neural Machine Translation},
  author={Feng, Yang and Xie, Wanying and Gu, Shuhao and Shao, Chenze and Zhang, Wen and Yang, Zhengxin and Yu, Dong},
  journal={arXiv preprint arXiv:1912.00178},
  year={2019}
}