Huyisbeee/ViKm-Translation-Task

ViKm-Trans

A high-quality synthetic Vietnamese–Khmer parallel corpus.

<p align="center"> <img src="data-generation-pipeline.png" width="900"> </p>

Overview

ViKm-Trans is a synthetic parallel corpus for Vietnamese ↔ Khmer machine translation.

Due to the scarcity of publicly available Vietnamese–Khmer parallel data, we propose a synthetic data generation framework that leverages abundant Vietnamese monolingual corpora together with large language models to construct high-quality parallel sentence pairs.

The dataset was introduced in our PACLIC 2025 paper:

Towards Improving Low-Resource Machine Translation with Lightweight Training and Synthetic Data: Case Study of Vietnamese-Khmer

Data Construction Pipeline

Our synthetic data generation framework consists of four main stages.

1. Monolingual Data Selection

We begin with the TED2020 Vietnamese corpus, which contains approximately 350K Vietnamese sentences.

To maximize translation quality while maintaining data diversity, we select 30K high-quality Vietnamese sentences using a TF-IDF-based sentence selection strategy.

2. Synthetic Translation via Knowledge Distillation

The selected Vietnamese sentences are translated into Khmer using GPT-4o as a black-box teacher model.

Instead of manually annotating bilingual data, we leverage knowledge distillation to efficiently generate synthetic Khmer translations.

3. Independent Back Translation

To verify translation quality, each generated Khmer sentence is translated back into Vietnamese.

We employ Google Translate as an independent translation system to avoid circular reasoning and reduce potential biases introduced by using the same model for both forward and backward translation.

4. Quality Filtering

The quality of each synthetic sentence pair is evaluated by comparing the original Vietnamese sentence with its back-translated version.

Following our paper, we compute a hybrid quality score:

$$ h(s)=2\cdot\frac{\overline{\mathrm{BLEU}}s\cdot\overline{\mathrm{METEOR}}s} {\overline{\mathrm{BLEU}}s+\overline{\mathrm{METEOR}}s} $$

BLEU and METEOR scores are first normalized to the range [0,1], and the final quality score is computed as their harmonic mean. Sentence pairs with low quality scores are discarded.

Dataset Statistics

File Format

Data is in CSV format.

Citation

If you find this dataset useful for your research, please consider giving this repository a ⭐ and citing our paper.

@inproceedings{nguyen-etal-2025-towards,
    title = {Towards Improving Low-Resource Machine Translation with Lightweight Training and Synthetic Data: Case Study of Vietnamese-Khmer},
    author = {Nguyen, Trong Huy and
              Le, Thanh Huong and
              Tran, Que Nhu},
    booktitle = {Proceedings of the 39th Pacific Asia Conference on Language, Information and Computation},
    pages = {38--50},
    year = {2025},
    address = {Hanoi, Vietnam},
    publisher = {Association for Computational Linguistics},
    url = {https://aclanthology.org/2025.paclic-1.4/}
}

Acknowledgements

This dataset was developed as part of our research on low-resource Vietnamese–Khmer machine translation.

We sincerely thank the creators of:

• —TED2020

whose resources made this work possible.

License

Please refer to the LICENSE file for usage terms.