CoolFace
Datasetpublic

raptorkwok/cantonese-chinese-parallel-corpus

Cantonese-Written Chinese Parallel Corpus (CCPC) About the Dataset Data Splits Training Data (train): 160,000 Sentence Pairs Validation Data (validation): 20,000 Sentence Pairs Test Data (test): 5,461 Sentence Pairs Languages Cantonese (yue) Traditional Chinese (zh-TW) Original Data Structure JSON lines consisting of yue, zh and ref fields. Data Source Apart from the data from our first generation… See the full description on the dataset page: https://huggingface.co/datasets/raptorkwok/cantonese-chinese-parallel-corpus.

sourceHugging Facecc-by-nc-sa-4.0updated 5mo agoView on Hugging Face
4likes71downloads
Dataset Card

Cantonese-Written Chinese Parallel Corpus (CCPC)

About the Dataset

Data Splits

  • —Training Data (train): 160,000 Sentence Pairs
  • —Validation Data (validation): 20,000 Sentence Pairs
  • —Test Data (test): 5,461 Sentence Pairs

Languages

  • —Cantonese (yue)
  • —Traditional Chinese (zh-TW)

Original Data Structure

  • —JSON lines consisting of yue, zh and ref fields.

Data Source

Apart from the data from our first generation of CCPC, there are additional data sources used by this dataset:

Written Chinese is translated by Microsoft Bing and DeepSeek V3, then translated to Traditional Chinese. Test Set's reference data are translated by a group of professional human translators.

Topic Classification
Topic CategoryDescriptionSentence Count (%)
LawCrime, court, judgment, legal matters54,947 (29.63%)
CultureTraditions, festivals, customs, social norms34,066 (18.37%)
SportsGames, athletes, competitions, exercises24,907 (13.43%)
EconomyFinance, business, markets, money, trade19,245 (10.38%)
PoliticsGovernment, elections, policy, international relations16,557 (8.93%)
ArtsLiterature, painting, film, theater, design13,299 (7.17%)
TechnologyGadgets, Internet, AI, software, innovation7,488 (4.04%)
NewsCurrent events, reports, headlines5,519 (2.98%)
MusicSongs, instruments, concerts, genres, performers3,955 (2.12%)
ScienceGeneral science, research, physics, chemistry2,545 (1.37%)
Biology & MedicineHealth, diseases, body, medical treatment1,920 (1.04%)
Daily conversationsCasual talk, greetings, small talk, everyday life1,033 (0.56%)

Considerations for Using the Data

Social Impact of Dataset

This dataset helps preserve the Cantonese language.

Discussion of Biases

As the data was mostly scraped from LIHKG and Telegram, there may be biases in the contents.

Known Limitations

Most Chinese translations in the training data are machine-translated by Microsoft Bing and DeepSeek V3; the translation may not be 100% accurate.

Other Information

Dataset Curator

License

CC BY-NC-SA 4.0

Citation

@inproceedings{10.1145/3639233.3639332,
    author = {Kwok, Raptor Yick-Kan and Au Yeung, Siu-Kei and Li, Zongxi and Hung, Kevin},
    title = {Cantonese to Written Chinese Translation via HuggingFace Translation Pipeline},
    year = {2024},
    isbn = {9798400709227},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3639233.3639332},
    doi = {10.1145/3639233.3639332},
    abstract = {Cantonese, a low-resource language [5] that has been used in Southeastern China for hundreds of years, with over 85 million native speakers worldwide, is poorly supported in the mainstream language model for existing translation platforms such as Baidu, Google and Bing. This paper presents a large parallel corpus of 130 thousand Cantonese and Written Chinese pairs. The data are used to train a translation model using the translation pipeline of the Hugging Face Transformers architecture, a dominant architecture for natural language processing nowadays [18]. The BLEU score and manual assessment evaluate the performance. The translation results achieve a BLEU score of 41.35and chrF++ score of 44.88on the entire validation set. The model also works reasonably well with long sentences of over 20 Chinese characters. It achieves a BLEU score of 48.61and chrF++ score of 39.87on long sentences. Those results are comparable with the existing Baidu Fanyi and Bing Translate. We also establish a Cantonese sentence evaluation metric to classify the quality of the source Cantonese sentence by professional translators. We then compare the BLEU and chrF++ scores with the corresponding evaluation score and found that the better the quality of the source sentence, the higher the BLEU and chrF++ scores. Last, we proved that our corpus enabled the Cantonese translation capability of the Chinese BART pre-trained model.},
    booktitle = {Proceedings of the 2023 7th International Conference on Natural Language Processing and Information Retrieval},
    pages = {77–84},
    numpages = {8},
    keywords = {Cantonese, Written Chinese, neural networks, translation},
    location = {Seoul, Republic of Korea},
    series = {NLPIR '23}
}
@misc {cantoneseWrittenChineseTranslationsGen3,
    author = {Kwok, R.},
    title = {Cantonese-Written Chinese Parallel Corpus (CCPC)},
    year = {2024},
    url = {https://huggingface.co/datasets/raptorkwok/cantonese-chinese-parallel-corpus},
}