> ## Documentation Index
> Fetch the complete documentation index at: https://mehen.ophi.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Japanese script composition

> Tier 0 Japanese register signal: kanji, hiragana, katakana, latin, and digit ratios.

Japanese is unusual among major languages: script composition alone carries enough information to
produce defensible readability scores without a tokenizer. This is the foundational insight of
Tateishi, Ono & Yamada (1988) and remains the basis for mehen's Tier-0 Japanese layer.

## Unicode script classification

Each grapheme cluster classifies into:

* Hiragana
* Katakana
* Kanji (Han + Extensions A/B + Compatibility)
* CJK punctuation
* Latin (+ Fullwidth)
* Digit

## Primary ratios

`kanji_ratio`, `hiragana_ratio`, `katakana_ratio`, `latin_ratio`, `digit_ratio`, `script_entropy`
(Shannon entropy over the five classes).

## Register bands

| Kanji ratio | Register                                   |
| ----------- | ------------------------------------------ |
| \< 20 %     | Children's writing, conversation.          |
| 20–30 %     | Casual prose, novels, user-facing content. |
| 30–40 %     | Newspaper, business writing, non-fiction.  |
| 40–50 %     | Technical, legal, academic.                |
| > 50 %      | Classical / literary, specialist text.     |

Katakana > 15 % typically signals software documentation (loanwords like `データベース`) or marketing
copy. Hiragana > 75 % indicates text aimed at small children or machine-translated output.

## Script-run features

A "run" is a maximal substring of same-script characters. Per document:

* Mean chars per alphabet run (`la`).
* Mean chars per hiragana run (`lh`).
* Mean chars per kanji run (`lc`).
* Mean chars per katakana run (`lk`).
* Percentages of each run type (`pa`, `ph`, `pc`, `pk`).
* Mean chars per sentence (`ls`).
* `、` per `。` (`cp`).

These are the exact inputs the Tateishi formula needs.

## Sentence segmentation

Primary terminators `。`, `！`, `？` plus half-width equivalents. Do not split inside `「…」`, `『…』`,
`（…）`. Treat blank-line paragraph boundaries and Markdown block boundaries as hard terminators.
Ellipsis `…` / `‥` / `...` is not a terminator.

## Sentence-length thresholds

* Warning: > 60 chars.
* Hard-to-read: > 90.
* Error: > 120.
* Mean sentence length > 60 triggers a document-level warning.

## References

* Tateishi, K., Ono, Y. & Yamada, H. (1988). *A Computer Readability Formula of Japanese Texts for
  Machine Scoring.* Proceedings of COLING-1988: 649–654.
  [ACL Anthology](https://aclanthology.org/C88-2132/).
* Lee, J. & Hasebe, Y. (2017). *jReadability — a web-based Japanese text-readability indexing
  system.* (Foundation for register-based Japanese readability work.)
  [jReadability](https://jreadability.net/sys/en).

## See also

* [Tateishi RS + Jōyō grade](/metrics/markdown/prose/tateishi-and-jouyou) — readability scores built on
  these inputs.
* [JTF rules](/metrics/markdown/prose/jtf-rules) — Japan Translation Federation conformance.