EMNLP 2018: A Study of Reinforcement Learning for Neural Machine Translation

@inproceedings{wu2018study,
  title={A Study of Reinforcement Learning for Neural Machine Translation},
  author={Wu, Lijun and Tian, Fei and Qin, Tao and Lai, Jianhuang and Liu, Tie-Yan},
  booktitle={Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  pages={3612--3621},
  year={2018}
}

Please first get familar with the basic Tensor2Tensor project: https://github.com/tensorflow/tensor2tensor.

The tesorflow version is 1.4, and the tensor2tensor version is 1.2.9.

Take WMT17 Chinese-English translation as example:

Different training strategies are provided.

• Different RL training strategies for NMT, evaluated on bilingual dataset.
  (1) HPARAMS=zhen_wmt17_transformer_rl_total_setting: terminal reward + beam search
  (2) HPARAMS=zhen_wmt17_transformer_rl_delta_setting: reward shapping + beam search
  (3) HPARAMS=zhen_wmt17_transformer_rl_delta_setting_random: reward shapping + multinomial sampling
  (4) HPARAMS=zhen_wmt17_transformer_rl_total_setting_random: terminal reward + multinomial sampling
  (5) HPARAMS=zhen_wmt17_transformer_rl_delta_setting_random_baseline: reward shaping + multinomial sampling + reward baseline
  (6) HPARAMS=zhen_wmt17_transformer_rl_delta_setting_random_mle: reward shapping + multinomial sampling + objectives combination

• Different monolingual data combination traininig in RL4NMT
  (1) zhen_src_mono: source monolingual data RL training based on bilingual data MLE model
  (2) zhen_tgt_mono: target monolingual data RL training based on bilingual data MLE model
  (3) zhen_src_tgt_mono: sequential mode [target monolingual data RL trianing based on (bilingual + source monolingual data) MLE model]
  (4) zhen_tgt_src_mono: sequential mode [source monolingual data RL training based on (bilinugal + target monolingual data) MLE model]
  (5) zhen_bi_src_tgt_mono: unified model

Supports MRT (minimum risk training) for NMT.

• This file contains the model builder, do sampling and feed the sampled translations for training.
• This file contains the loss function building for RL and regular MLE.
• This file includes the detailed bleu reward calculation.