Violet-yo/Chinese-Braille-Dataset-Full-Tone

Chinese Braille Sentence Corpus (Full Tone)

📃 [Paper] • 💻 [Code] • 📖 [Passage corpus] • 🎬 [Demo]

The sentence-level half of the Braille–Chinese parallel corpus used in "Vision-Braille: A Curriculum Learning Toolkit and Braille–Chinese Corpus for Braille Translation" (EMNLP 2026 Main Conference).

Every Braille sequence here retains all tone markers (retention rate r = 100). This is the source corpus: the tone-omission variants used for curriculum training are derived from it by the tone-deletion operator, not stored separately. See Deriving tone-omission variants.

Dataset details

Direction	Chinese Braille → written Chinese (BR2ZH)
Granularity	Sentence
Tone retention	100% (all tone markers present)
Standard	GF 0019-2018 (国家通用盲文方案)
Source text	Leipzig Corpora Collection, Mandarin news 2007–2009

Splits

Split	Samples	Braille len. (avg/med)	Chinese len. (avg/med)
Train	376,846	116 / 103	31 / 28
Validation	56,258	114 / 101	31 / 28
Test	59,908	113 / 100	31 / 27

Lengths are token counts after prompt formatting. Split ratio is 8:1:1.

Fields

Field	Type	Description
braille_text	string	Chinese Braille, Unicode U+2800–U+28FF, all tone markers retained
chinese_text	string	The parallel written Chinese

Example

{
  "braille_text": "⠎⠕⠄⠊⠄⠐⠕⠄⠵⠖⠄⠛⠳⠆⠛⠾⠂⠙⠡⠆⠃⠔⠄⠌⠢⠆⠉⠆⠃⠊⠄⠎⠪⠆⠛⠊⠁⠓⠺⠆⠇⠳⠂⠛⠮⠄⠇⠮⠂⠝⠔⠆⠓⠢⠂⠞⠔⠁⠐⠆⠘",
  "chinese_text": "所以，我早就决定把这次比赛机会留给雷纳和他。”"
}

Usage

from datasets import load_dataset

ds = load_dataset("Violet-yo/Chinese-Braille-Dataset-Full-Tone")
print(ds["train"][0])

How it was built

Braille–Chinese parallel data is scarce and, where it exists, private: aligning it requires specialists fluent in both Braille and Chinese. Conversion in the other direction is rule-governed and can be automated, so the corpus is built synthetically. Mandarin news sentences from the Leipzig Corpora Collection were converted to Braille with the conversion tool of the Digital Platform for Chinese Braille, which annotates character pronunciations and emits fully toned Braille.

Braille encoding notes

• Braille occupies Unicode U+2800–U+28FF.
• This corpus carries no word division. Braille cells run unbroken, with no U+2800 BRAILLE PATTERN BLANK between words. Chinese is written without inter-word spaces, and in an OCR pipeline space tokens are unreliable and frequently dropped, so the models are trained to recover segmentation from local cell cues instead.
• Tone markers are the four cells ⠁ (tone 1), ⠂ (tone 2), ⠄ (tone 3), ⠆ (tone 4).
• Number sequences are introduced by the number sign ⠼.

Note on word division. An earlier revision of this dataset (before commit 9fa3ec6, 2025-11-26) carried U+2800 word division. It was removed to match the training setup described above. If you need the word-divided variant, it is preserved as the cleaned_stage2_sentence config of Violet-yo/vision-braille-dataset.

Deriving tone-omission variants

Real Braille writers omit tone markers, so the models are trained on a ladder of tone retention rates. Given a retention rate r ∈ [0, 100], each tone marker is deleted independently at random with probability 1 - r/100; every non-tone cell is left alone. A fixed seed (42) makes this deterministic.

Deletion is deliberately unstructured rather than a hand-coded implementation of the 省写 conventions in GF 0019-2018. Random deletion drops tones indiscriminately and does not expose the position-specific regularities an experienced writer follows, so the model has to recover tone from context. That is the harder setting, and it transfers: the model translates expert-produced Braille written under standard 省写 conventions accurately despite never having been trained on rule-based omission.

Scripts are in data/ of the code repository:

python data/removeToneTrainSentence_10pc.py <input.json> <output_dir>      # r = 10
python data/removeToneSeperate_100pc_10pc.py <input.json> <output_dir>     # r = 100 … 10

Two classes of tone marker are always preserved: those following sentence-final punctuation (。？！) and every cell inside a number sequence.

Related resources

Passage-level corpus	Violet-yo/Passage-Chinese-Braille-Dataset-Full-Tone
Training-ready ShareGPT corpora	Violet-yo/vision-braille-dataset
Code	github.com/AlanYWu/VisionBraille
Demo	visionbraille.org · xinshiye.cc

Limitations

The Braille side is machine-generated from Chinese text by a rule-based converter, not transcribed by a human Braille writer. It is therefore free of the transcription slips, contraction choices, and layout conventions found in Braille produced by people. Models trained on it face a domain shift on real, human-produced Braille, and a further shift on Braille recovered by OCR from photographed pages.

License

Released under CC BY-NC-SA 4.0 for non-commercial research use. The source Chinese text remains the property of its original owners; sentence text derives from the Leipzig Corpora Collection under that collection's terms.

Citation

@inproceedings{wu2026visionbraille,
  title     = {Vision-Braille: A Curriculum Learning Toolkit and Braille--Chinese Corpus for Braille Translation},
  author    = {Wu, Alan Yo and Yuan, Ye and Xiao, Zhiping and Zhang, Ming},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}