pangjh3/plm4mt

PLM4MT

This is the code for our ACL 2022 work MSP: Multi-Stage Prompting for Making Pre-trained Language Models Better Translators. The implementation is on top of the open-source NMT toolkit THUMT.

Contents

• Prerequisites
• mGPT
• Format
• Training
• Decoding
• Postprocessing
• License
• Citation

Prerequisites

• Python >= 3.7
• tensorflow-cpu >= 2.0
• torch >= 1.7
• transformers

Please read the document of THUMT before using this Repository.

mGPT

You can download the mGPT checkpoint at this url.

Format

We use <extra_id_0> to separate a source and a target sentence. For the WMT14 En-De dataset, the training file contains lines with the following format:

<extra_id_5> Graphical artwork, corporate identity and corporate design. <extra_id_0> Grafische Gestaltung, Layout, Corporate Identity und Corporate Design.

Here <extra_id_5> is a tag to indicate the source language, which can be omitted.

For inference, the test set contains lines like:

<extra_id_5> Gutach: Increased safety for pedestrians <extra_id_0>

Training

Using the following command to train a prompt for translation:

CODES=<path/to/this-repository>
CKPT=<path/to/mGPT-checkpoint>
export PYTHONPATH=$CODES:$PYTHONPATH

export USE_TF=0
export USE_TORCH=1

python $CODES/thumt/bin/trainer.py \
    --half \
    --input <path/to/traininig-file> \
    --model <model_name> \
    --ptm $CKPT \
    --parameters=device_list=[0,1,2,3,4,5,6,7],\
                 train_steps=40000,update_cycle=16,batch_size=256,\
                 save_checkpoint_steps=2000,max_length=256 \
    --hparam_set base

Here model_name has the following three options:

• mgpt_prompt: mGPT with Prompt tuning
• mgpt_prefix: mGPT with Prefix-tuning
• mgpt_msp: mGPT with multi-stage prompting

Decoding

The following command decodes an input file:

CODES=<path/to/this-repository>
export PYTHONPATH=<path/to/this-repository>:$PYTHONPATH

python $CODES/thumt/bin/translator.py \
  --input <path/to/test-file> \
  --ptm <path/to/mgpt> \
  --output <path/to/output-file> \
  --model <model-name> \
  --half --prefix <path/to/trained-prompt> \
  --parameters=device_list=[0,1,2,3],\
               decode_alpha=0.0,\
               decode_batch_size=4,\
               prompt_length=128

Postprocessing

We use tools/punc.cpp to replace punctuations for Chinese. Use the following command to compile the code:

g++ -std=c++11 -o punc tools/punc.cpp

Then use the following command to replace punctuations

cat <path/to/input-file> | ./punc | <path/to/output-file>

License

Open source licensing is under the BSD-3-Clause, which allows free use for research purposes.

Citation

@article{tan2021msp,
  title={{MSP}: Multi-stage prompting for making pre-trained language models better translators},
  author={Tan, Zhixing and Zhang, Xiangwen and Wang, Shuo and Liu, Yang},
  journal={arXiv preprint arXiv:2110.06609},
  year={2021}
}