WMT2022-Large-Scale-African

This project introduces Tencent’s Multilingual Machine Translation System for the WMT2022 Large-Scale African Translation shared task.

We (team Borderline) won the 1st place 🥇 in the constrained track (i.e., external data and pretrained models are not allowed).

Competition Results on the Blind Test Sets (By 2022/09/29)

Test our Pretrained Model

Config the Local Environment for Evaluation

Let us assume we are provided with a clean machine, then we have the following steps:

• Download Miniconda3 py37_4.10.3, and see Instruction to install.

# make sure your networks are well configured.
wget https://repo.anaconda.com/miniconda/Miniconda3-py37_4.10.3-Linux-x86_64.sh
sh Miniconda3-py37_4.10.3-Linux-x86_64.sh

# three prompts during installing
# 1. Do you accept the license terms? [yes|no] 
#   yes
# 2. Miniconda3 will now be installed into this location: /home/username/miniconda3
#   ENTER
# 3. Do you wish the installer to initialize Miniconda3 by running conda init? [yes|no]
#   yes

source ~/.bashrc

• Install Pytorch 1.7.1 + CUDA 11.0.

# CUDA 11.0
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

• Install fairseq==0.10.2 and dependencies with sub_requirements.txt.

pip install -r sub_requirements.txt

# packages in sub_requirements.txt
tensorboardX
cffi
cython
dataclasses
hydra-core
numpy
regex
sacrebleu
tqdm
sacremoses
nltk>=3.2
matplotlib
absl-py
sentencepiece
setuptools>=18.0
fairseq==0.10.2

Download our Pretrained Model

• Due to the storage limit, we only release the contrastive submission, which include our code, sentencepiece model, and pretrained model.

# decompress and obtain `submission_contrast`
unzip submission_contrast.zip

• Use the inference_afr.sh for inference.

# update the `inference_afr.sh` file with the one provided in this repo
cd submission_contrast
pip install -e ./fairseq

# sh ./inference_afr.sh MODEL_PATH SRC_LANG TGT_LANG BATCH_SIZE SRC_TEXT OUTPUT_DIR
sh ./inference_afr.sh models/trans_deepwide_dp020_african26_7g106_erm3_ft30k fra eng 16 data/test.fr output

Note: Training instruction is coming soon.

Public Impact

Citation

Please kindly cite our paper if you find it helpful:

@inproceedings{jiao2022wmt,
  title={Tencent’s Multilingual Machine Translation System for WMT22 Large-Scale African Languages},
  author={Wenxiang Jiao and Zhaopeng Tu and Jiarui Li and Wenxuan Wang and Jen-tse Huang and Shuming Shi},
  booktitle = {WMT},
  year      = {2022}
}