The code is based on https://github.com/ZNLP/NCLS-Corpora.
word-level-supervised: supervised pre-training with word-level knowledge distillation.
sent-level-supervised: supervised pre-training with sentence-level knowledge distillation.
rl-sim: reinforcement learning fine-tuning with cross-lingual similarity and ROUGE scores as rewards.
python train.py -config train.json
python translate.py -config translate.json