Authors:
Yusuke Sakai*, Zhi Qu*, Hidetaka Kamigaito, Taro Watanabe, Xiaojiang Liu
This repository accompanies our paper:
Domain Translation with Monolingual Lexical Distribution, which has been accepted by TMLR.
You can find our paper by this URL.
This repository serves as the public hub for releasing the code and data used in our work.
The fine-tuning approach proposed in the paper is a variant of energy-based models (EBMs), built upon Conditional Distributional Policy Gradients (CDPG).
The task in our paper is domain adaptation in machine translation.
The English-German data is publicly available.
In this repository, we only provide the subsets that are used in our experiments.
If you need the full dataset, please refer to the original project.
Note:
The test sets have been cleaned and corrected as described in our paper.
Results may therefore differ from those obtained using the original raw data.
The English-Chinese data is subject to licensing restrictions, so, please obtain it from UM-Corpus
As described in the paper, the monolingual data used during training is derived from the WMT Translation Tasks.
The monolingual corpora used in our experiments are provided in this repository.
You can find the raw data from WMT23.
The experiments were conducted with the following core dependencies:
Python 3.10
datasets 3.0.0
torch 2.5.1+cu121
transformers 4.35.0
Our tech is based on the toolkit released by NAVER, named disco. We have a lot of updates, including but not limited to
./disco/extra/
./disco/distributions/lm_distribution.py
...
You can explore the original project disco.
Note:
Please be careful on ./disco/distributions/lm_distribution.py,
the codes related to setting pad_token_id is fixed in original version.
Given there are various models are usable, we follow the original logic to add and fix codes for translation models.
you can install the toolkit by
cd ./disco
pip install -e .
You can run below script as a minimal use case, which shows a CDPG with fixed hyperparameters.
bash run.sh
For Dynamic CDPG, you can run
bash run_dynamic.sh
For test, please refer to ./translation.py
@article{
sakai2026domain,
title={Domain Translation with Monolingual Lexical Distribution},
author={Yusuke Sakai and Zhi Qu and Hidetaka Kamigaito and Taro Watanabe and Xiaojiang Liu},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=UKLBobrFCR},
note={}
}