srSergio/mindo-hanzi

Mindo — Optimal Chinese Character Learning Order

20,857 Chinese characters ranked by a topologically-sorted, word-propagation-optimized algorithm based on Loach & Wang (2016), validated against the New HSK 2025 standard.

What is this?

A rigorously computed optimal order to learn Chinese characters that balances three objectives:

1. Habilitación — how many useful words does this character unlock?
2. Topological constraints — primitives before compounds (no learning 的 before 白 and 勺)
3. Backwards propagation — high-frequency words pull their prerequisite characters forward

The order is computed over a DAG of 30,252 nodes (20,857 characters + 9,395 HSK words) with 24,634 dependency edges, using Kahn's algorithm with a priority queue and calibrated backwards priority propagation.

Key Features

• 100% stroke coverage (20,857/20,857) from Unihan database
• 0 topological violations guaranteed by Kahn's algorithm
• SUBTLEX-CH frequency data from 46.8M character film subtitle corpus
• HSK 2025 comparison included for all 1,505 official characters
• 4,541 annotated component relationships with semantic/phonetic roles
• Grid-search validated parameter γ = 0.1

Files

File	Description	Format
data/train-00000-of-00001.parquet	20,857 characters with ranks, habilitación, POS, HSK level	Parquet
data/components.parquet	7,266 component relationships (4,541 direct + 2,725 indirect)	Parquet
data/words.parquet	11,908 multi-character words with frequency, POS, definitions	Parquet
data/hsk_comparison.json	HSK 2025 official order vs Mindo optimal order	JSON
data/summary.json	Methodology overview and statistics	JSON

Character Fields

Each entry in characters.jsonl:

Field	Type	Description
character	string	The Chinese character
learning_rank	int	Position in the optimal learning order (1 = learn first)
habilitacion	float	Word-enabling power from SUBTLEX-CH top 10k words
propagation_boost	float	Priority propagated backwards from descendant words
priority	float	Final η = (habilitacion + boost) / learning_cost
learning_cost	float	Cost model: 1 + 0.1×strokes (primitives) or n_comps−1 (compounds)
strokes	int	Number of strokes (Unihan kTotalStrokes)
subtlex_word_count	int	How many SUBTLEX-CH words contain this character
is_standalone_word	bool	Has SUBTLEX-CH frequency as independent word
pos	string	Dominant part of speech (SUBTLEX-CH tag)
frequency_subtlex	float	SUBTLEX-CH normalized frequency (W.million/1e6)
hsk_level	int	HSK level (1-9) from New HSK 2025
hsk_rank	int	Position within the HSK official order
pinyin	string	Pronunciation
definition	string	Dictionary definition
traditional	string	Traditional character variant
card_type	string	Kanji, Radical, or Vocabulary
components	array	Sub-components with roles: [{component, role, is_direct}]

Component roles: semantic (provides meaning), phonetic (provides sound), both, or null.

Example: 妈 = [{component: "女", role: "semantic"}, {component: "马", role: "phonetic"}] — woman radical gives meaning, horse gives sound (mǎ → mā).

Component Fields

Each entry in components.jsonl:

Field	Type	Description
compound	string	The compound character
component	string	A sub-component of the compound
role	string	semantic, phonetic, both, or null

Top 10 Characters

Rank	Char	Pinyin	Meaning (approx.)	Habilitación	Boost
1	子	zǐ	child, son	61.1	1558.8
2	人	rén	person	60.2	1421.4
3	一	yī	one	37.9	882.5
4	大	dà	big	39.2	1017.5
5	不	bù	not	48.2	924.5
6	口	kǒu	mouth	11.9	889.2
7	上	shàng	up, above	35.6	852.6
8	心	xīn	heart	27.6	840.8
9	手	shǒu	hand	31.0	815.1
10	事	shì	thing, matter	24.0	552.2

Methodology

The algorithm is a reimplementation and extension of Loach & Wang (2016), "Optimizing the Learning Order of Chinese Characters Using a Novel Topological Sort Algorithm", PLoS ONE.

Priority Formula

η(c) = (hab(c) + boost(c)) / cost(c)

where:
  hab(c) = Σ (zipf_word / max_zipf)²  for top 10k SUBTLEX-CH words containing c
  boost(c) = Σ η_word × 0.3^d₁ × 0.07^d₂  for all descendant words
  cost(c) = 1 + 0.1 × strokes  (primitives)
          = n_components − 1   (compounds)

Improvements over the original paper

1. Kahn's algorithm instead of post-hoc sorting — guarantees 0 topological violations
2. Backwards priority propagation — words pull their character prerequisites forward
3. Top 10k SUBTLEX-CH words — original paper's word filtering method
4. 100% stroke coverage — 46 manual corrections completing Unihan data
5. HSK 2025 validation — empirical comparison against official standard

Citation

@misc{mindo2025,
  title={Mindo: Optimal Chinese Character Learning Order},
  author={Sergio Rodríguez},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/srsergi0/mindo-hanzi}},
}

License

This dataset is released under CC-BY-4.0. The underlying SUBTLEX-CH data is from Cai & Brysbaert (2010). HSK 2025 data is from the Anki Xiehanzi deck.