The maps-mt from zwhe99

🗺️ MAPS: Multi-Aspect Prompting and Selection

Implementaion of our paper:

Exploring Human-Like Translation Strategy with Large Language Models

🔥 Update

[March 19, 2024]: Accepted to TACL 2024!
[June 14, 2023]: This work now has a . Try it!
[June 10, 2023]: interactive.py now enables running of MAPS-mt in an interactive mode.
[June 10, 2023]: We now support translation between any pair of these languages: English, Chinese, Japanese, French, and German.

MAPS

Motivation

The difference between machine and human translation in an English-Chinese example. Typical neural machine translation is a source-target mapping process, while human translators can take complex steps to ensure the quality and accuracy of the translation.

Framework

MAPS aims to enable LLMs to mimic the human translation process by multi-aspect prompting and selection.

Dependencies

Download COMET and BLEURT checkpoints:

wget https://unbabel-experimental-models.s3.amazonaws.com/comet/wmt21/wmt21-comet-qe-da.tar.gz
tar -xf wmt21-comet-qe-da.tar.gz -C eval_ckpt/

wget https://storage.googleapis.com/bleurt-oss-21/BLEURT-20.zip .
unzip -d eval_ckpt/ BLEURT-20.zip

Create conda env

conda create -n maps -c pytorch -c nvidia python==3.8.13 krb5 git pytorch==2.0.0 pytorch-cuda=11.7

Install other python packages
```
pip3 install -r requirements.txt
```

Reproduce the main results

Run

Preparation

Set your openai API_KEY in model/openai/translate.py
Set Alpaca checkpoint file in run-maps-alpaca.sh and run-translation-alpaca.sh

Run MAPS

text-davinci-003: sh run-maps.sh
Alpaca: sh run-maps-alpaca.sh

Run other methods

text-davinci-003: sh run-translation-003.sh
Alpaca: sh run-translation-alpaca.sh

Note: The translation results have already been generated and saved in the output directory. Therefore, the scripts won't repeat the inference. If you want to regenerate the results, simply delete the contents within the output directory.

Evaluation

sh run-evaluation.sh > evaluation.log

Interactive

If you just want to have a try, you can try the interactive script like this without need of GPU or CUDA (only text-davinci-003 now):

# Preparation
wget https://unbabel-experimental-models.s3.amazonaws.com/comet/wmt21/wmt21-comet-qe-da.tar.gz
tar -xf wmt21-comet-qe-da.tar.gz -C eval_ckpt/   
conda create -n maps -c pytorch python==3.8.13 pytorch==2.0.0  
conda activate maps
pip3 install -r requirements.txt

# Interactive
(maps) zwhe@zhiweideMacBook-Pro MAPS-mt % python3 interactive.py --lang-pair en-zh

Enter source English sentence: Joint Aid for Dogs is a high specification joint and muscle supplement with glucosamine for dogs, designed to aid freedom of movement.

Output:

Remember to set your openai API_KEY in model/openai/translate.py. You can also take a look at the demo website.

Citation

@article{he2023exploring,
    author = {He, Zhiwei and Liang, Tian and Jiao, Wenxiang and Zhang, Zhuosheng and Yang, Yujiu and Wang, Rui and Tu, Zhaopeng and Shi, Shuming and Wang, Xing},
    title = "{Exploring Human-Like Translation Strategy with Large Language
                    Models}",
    journal = {Transactions of the Association for Computational Linguistics},
    volume = {12},
    pages = {229-246},
    year = {2024},
    month = {03},
    issn = {2307-387X},
    doi = {10.1162/tacl_a_00642},
    url = {https://doi.org/10.1162/tacl\_a\_00642},
    eprint = {https://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl\_a\_00642/2346100/tacl\_a\_00642.pdf},
}

zwhe99 / maps-mt Goto Github PK

maps-mt's Introduction

🗺️ MAPS: Multi-Aspect Prompting and Selection

MAPS

Dependencies

Reproduce the main results

Run

Evaluation

Interactive

Citation

maps-mt's People

Contributors

Stargazers

Watchers

Forkers

maps-mt's Issues

Recommend Projects

Recommend Topics

Recommend Org