Contrastive Learning for Many-to-many Multilingual Neural Machine Translation(mCOLT/mRASP2), ACL2021

The code for training mCOLT/mRASP2, a multilingual NMT training framework, implemented based on fairseq.

mRASP2: paper

mRASP: paper, code

News

We have released two versions, this version is the original one. In this implementation:

• You should first merge all data, by pre-pending language token before each sentence to indicate the language.
• AA/RAS muse be done off-line (before binarize), check this toolkit.

New implementation: https://github.com/PANXiao1994/mRASP2/tree/new_impl

• Acknowledgement: This work is supported by Bytedance. We thank Chengqi for uploading all files and checkpoints.

Introduction

mRASP2/mCOLT, representing multilingual Contrastive Learning for Transformer, is a multilingual neural machine translation model that supports complete many-to-many multilingual machine translation. It employs both parallel corpora and multilingual corpora in a unified training framework. For detailed information please refer to the paper.

Pre-requisite

pip install -r requirements.txt

Training Data and Checkpoints

We release our preprocessed training data and checkpoints in the following.

Dataset

We merge 32 English-centric language pairs, resulting in 64 directed translation pairs in total. The original 32 language pairs corpus contains about 197M pairs of sentences. We get about 262M pairs of sentences after applying RAS, since we keep both the original sentences and the substituted sentences. We release both the original dataset and dataset after applying RAS.

Checkpoints & Results

• Please note that the provided checkpoint is sightly different from that in the paper. In the following sections, we report the results of the provided checkpoints.

English-centric Directions

We report tokenized BLEU in the following table. (check eval.sh for details)

Unsupervised Directions

We report tokenized BLEU in the following table. (check eval.sh for details)

Zero-shot Directions

• row: source language
• column: target language We report sacreBLEU in the following table.

Training

export NUM_GPU=4 && bash train_w_mono.sh ${model_config}

• We give example of ${model_config} in ${PROJECT_REPO}/examples/configs/parallel_mono_12e12d_contrastive.yml

Inference

• You must pre-pend the corresponding language token to the source side before binarize the test data.

fairseq-generate ${test_path} \
    --user-dir ${repo_dir}/mcolt \
    -s ${src} \
    -t ${tgt} \
    --skip-invalid-size-inputs-valid-test \
    --path ${ckpts} \
    --max-tokens ${batch_size} \
    --task translation_w_langtok \
    ${options} \
    --lang-prefix-tok "LANG_TOK_"`echo "${tgt} " | tr '[a-z]' '[A-Z]'` \
    --max-source-positions ${max_source_positions} \
    --max-target-positions ${max_target_positions} \
    --nbest 1 | grep -E '[S|H|P|T]-[0-9]+' > ${final_res_file}
python3 ${repo_dir}/scripts/utils.py ${res_file} ${ref_file} || exit 1;

Synonym dictionaries

We use the bilingual synonym dictionaries provised by MUSE.

We generate multilingual synonym dictionaries using this script, and apply RAS using this script.

Contact

Please contact me via e-mail [email protected] or via wechat/zhihu

Citation

Please cite as:

@inproceedings{mrasp2,
  title = {Contrastive Learning for Many-to-many Multilingual Neural Machine Translation},
  author= {Xiao Pan and
           Mingxuan Wang and
           Liwei Wu and
           Lei Li},
  booktitle = {Proceedings of ACL 2021},
  year = {2021},
}