calfa-ai/chiknowpo

CHI-KNOW-PO — Line-Level HTR Ground-Truth for Chinese Historical Texts

Dataset Description

The CHI-KNOW-PO (Chinese Knowledge and Poetry) project aims to digitize and publish an online, searchable corpus of approximately 60,000 pages of xylographed documents from the Chinese medieval period (ca. 200–1000 CE). The corpus covers poetic anthologies, commentaries, dictionaries, encyclopedias, and technical treatises — processed using Handwritten Text Recognition (HTR).

This HuggingFace dataset provides cropped line-level images paired with their transcriptions and rich metadata for 13 historical Chinese documents. It is designed as a ready-to-use benchmark for Chinese historical HTR.

The full page-level dataset (PageXML + full-page images) is available on GitHub.

Key Features

• 13,634 cropped text-line images from 325 annotated pages
• 104,769 transcribed characters covering 5,589 unique sinograms
• 13 documents spanning 9 genres: anthologies, encyclopedias, dictionaries, commentaries, collections, essays, and technical treatises
• Semantic metadata per line: document ID, title (Chinese/English), type, author, edition, library, call number
• Stratified train/val/test splits (80/10/10) by document

Dataset Composition

N°	Abbreviation	Title (Chinese)	Type	Pages	Lines	Characters
A-1	Li Wenxuan	李善注文選	Belles Lettres	56	1,654	15,650
A-3	Yutai	玉臺新詠	Belles Lettres	10	590	4,818
A-4	Tangshi	全唐詩	Belles Lettres	10	492	3,592
S-1	Beitang	北堂書鈔	Encyclopedia	35	1,508	12,940
S-2	Bowu zhi	博物志	Essay	23	302	3,786
S-3	Chuxue	初學記	Encyclopedia	20	1,267	11,572
S-4	Erya	影宋鈔繪圖爾雅	Dictionary	38	2,532	8,568
S-5	Maoshi shu	毛詩草木鳥獸蟲魚疏	Commentary	10	397	3,980
S-6	Yiwen	藝文類聚	Encyclopedia	11	356	3,524
S-7	Zhibuzu	知不足齋叢書	Collection	49	1,766	13,810
T-1	Shiwen leiju	古今事文類聚	Practical Encyclopedia	20	1,053	7,285
T-2	Qimin yaoshu	齊民要術	Agricultural treatise	20	885	8,711
T-3	Xinzhai	心齋十種	Practical collection	23	832	6,533
	TOTAL			325	13,634	104,769

Data Splits

Splits are stratified by doc_id to ensure proportional representation of each document in every split.

Split	Lines	Percentage
Train	10,907	80.0%
Validation	1,363	10.0%
Test	1,364	10.0%

Data Fields

Field	Type	Description
image	Image	Cropped line-level image (JPEG)
file_name	string	Original image filename
transcription	string	Ground-truth transcription of the text line
source_page	string	Filename of the source full-page image
doc_id	string	Document identifier (e.g. A-1, S-3, T-2)
title_zh	string	Document title in Chinese
title_en	string	Document title in romanized form
title_abbr	string	Abbreviated title used in the paper
author	string	Author of the work
compiler	string	Compiler of the work (if applicable)
main_text_creation	string	Date or period of original text composition
edition	string	Edition date of the physical copy
type	string	Genre/type of the document
library	string	Holding library
call_number	string	Library call number

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("calfa-ai/chiknowpo")

# Access a sample
sample = ds["train"][0]
sample["image"].show()
print(sample["transcription"])

# Filter by document
beitang = ds["train"].filter(lambda x: x["doc_id"] == "S-1")

Source Documents

The documents are xylographed (woodblock-printed) editions from late imperial China, preserved in three French libraries:

• BNU — Bibliothèque nationale et universitaire de Strasbourg
• BULAC — Bibliothèque universitaire des langues et civilisations (Paris)
• BIHEC — Bibliothèque de l'Institut des hautes études chinoises, Collège de France (Paris)

Each page typically features vertical columns read from top to bottom and right to left. Commentaries are embedded in double columns with smaller font, interleaved within the main text — a layout characteristic of Chinese xylographic editions.

Corpus Description

The corpus was designed to represent a literate library of the Chinese first millennium (ca. 200–1000 CE), excluding Buddhist texts. It favors thematic coherence around plants and co-occurrences across genres: knowledge texts (lexicons, encyclopedias), poetry, treatises on materia medica, and agricultural treatises.

Three categories of texts are represented:

• Anthologies (A): Collections of poetry and belles-lettres with scholarly commentaries (Wenxuan, Yutai xinyong, Quan Tang shi).
• Scholarship (S): Encyclopedias, dictionaries, commentaries, and reference compilations (Beitang shuchao, Chuxue ji, Erya, Yiwen leiju, etc.).
• Technical and practical knowledge (T): Agricultural treatises and practical encyclopedias (Qimin yaoshu, Shiwen leiju, Xinzhai).

Annotation Process

Annotations were made on the Calfa Vision platform, a free web-based annotation tool for documents and images designed for Oriental scripts. The platform incorporates active learning strategies, automatically generating and refining layout and text predictions as annotations progress.

Benchmark Results

HTR results using a CRNN architecture enhanced with GAN-augmented data (CycleGAN for rare character augmentation). A generic model was first trained on all documents, then fine-tuned per target manuscript:

Document	N°	Accuracy (%)
Li Wenxuan	A-1	99.38 (± 1.2)
Yutai	A-3	98.52 (± 1.2)
Tangshi	A-4	99.25 (± 1.8)
Beitang	S-1	98.76 (± 1.8)
Bowu zhi	S-2	99.18 (± 1.8)
Chuxue	S-3	97.57 (± 1.7)
Erya	S-4	96.57 (± 0.4)
Maoshi shu	S-5	98.42 (± 1.8)
Yiwen	S-6	98.72 (± 1.7)
Zhibuzu	S-7	98.70 (± 1.8)
Shiwen leiju	T-1	97.47 (± 4.5)
Qimin yaoshu	T-2	99.35 (± 2.8)
Xinzhai	T-3	97.61 (± 3.2)

Average accuracy: 98.45% (± 1.9%)

The accuracy figures (±) account for predictions following uncontrolled layout detection. Recognition accuracy for unknown characters (including GAN-generated ones) reaches 86.21%.

Challenges

Working with this corpus presents several specific challenges:

• Character diversity: 5,589 unique sinograms, with 30.46% appearing only once in the dataset — a significant few-shot learning challenge.
• Character imbalance: Most frequent characters (之: 2,239 samples; 也: 1,552; 曰: 1,549) vs. rare characters with a single occurrence.
• Graphic variants (yi ti zi 異體字): Xylographic printing preserves handwriting diversity, leading to graphic variants not standardized in Unicode.
• Complex layouts: Main text and commentaries interleaved in single/double columns with different font sizes.
• Reading order: Right-to-left, top-to-bottom hierarchical reading complicated by embedded commentaries.

Related Resources

• Paper: Optimizing HTR and Reading Order Strategies for Chinese Imperial Editions with Few-Shot Learning (ICDAR 2024 Workshops)
• PageXML dataset (GitHub): calfa-co/chi-know-po
• Project page: CHI-KNOW-PO Corpus (CollEx-Persée)
• GitLab (research): gitlab.huma-num.fr/chi-know-po
• Annotation platform: Calfa Vision

Citation

@InProceedings{10.1007/978-3-031-70642-4_3,
  author    = {Bizais-Lillig, Marie and Vidal-Gor{\`e}ne, Chahan and Dupin, Boris},
  editor    = {Mouch{\`e}re, Harold and Zhu, Anna},
  title     = {Optimizing HTR and Reading Order Strategies for Chinese Imperial Editions with Few-Shot Learning},
  booktitle = {Document Analysis and Recognition -- ICDAR 2024 Workshops},
  year      = {2024},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {37--56}
}

License

This dataset is released under the Apache License 2.0.

Acknowledgements

The CHI-KNOW-PO project was funded by the University of Strasbourg Institute for Advanced Studies (USIAS) and CollEx-Persée. It was conducted in collaboration with three libraries in France: the BULAC (Paris), the BNU (Strasbourg), and the BIHEC at the Collège de France (Paris). The Calfa start-up was in charge of developing HTR models.