README.md

Cross-lingual Retrieval for Iterative Self-Supervised Training

https://arxiv.org/pdf/2006.09526.pdf

Introduction

CRISS is a multilingual sequence-to-sequnce pretraining method where mining and training processes are applied iteratively, improving cross-lingual alignment and translation ability at the same time.

Requirements:

• faiss: https://github.com/facebookresearch/faiss
• mosesdecoder: https://github.com/moses-smt/mosesdecoder
• flores: https://github.com/facebookresearch/flores
• LASER: https://github.com/facebookresearch/LASER

Unsupervised Machine Translation

1. Download and decompress CRISS checkpoints

cd examples/criss
wget https://dl.fbaipublicfiles.com/criss/criss_3rd_checkpoints.tar.gz
tar -xf criss_checkpoints.tar.gz

2. Download and preprocess Flores test dataset

Make sure to run all scripts from examples/criss directory

bash download_and_preprocess_flores_test.sh

3. Run Evaluation on Sinhala-English

bash unsupervised_mt/eval.sh

Sentence Retrieval

1. Download and preprocess Tatoeba dataset

bash download_and_preprocess_tatoeba.sh

2. Run Sentence Retrieval on Tatoeba Kazakh-English

bash sentence_retrieval/sentence_retrieval_tatoeba.sh

Mining

1. Install faiss

Follow instructions on https://github.com/facebookresearch/faiss/blob/master/INSTALL.md

2. Mine pseudo-parallel data between Kazakh and English

bash mining/mine_example.sh

Citation

@article{tran2020cross,
  title={Cross-lingual retrieval for iterative self-supervised training},
  author={Tran, Chau and Tang, Yuqing and Li, Xian and Gu, Jiatao},
  journal={arXiv preprint arXiv:2006.09526},
  year={2020}
}