Scribe ✍️

A context-aware kanji/kana checker for Japanese official documents — judging the notation that dictionaries and regular expressions cannot, and citing the normative clause as evidence.

💻 GitHub 🧠 Model 📚 Dataset

The problem

Japan's official-writing guidelines ("Considerations for Creating Official Documents," Council for Cultural Affairs, 7 Jan 2022) require the same word to be written differently by its syntactic role: a formal noun こと in kana, but the substantive 事 ("the gravity of the matter") in kanji. Uniform dictionary replacement (textlint + prh) collapses each word to one spelling, so wherever both usages appear it is structurally forced to get half of them wrong. Scribe judges the context and attaches the clause it relied on.

Worked example — Scribe vs. uniform replacement (prh)

✓ Scribe (context-aware)

申請の内容を確認することが必要である。事故のとき、速やかに御連絡ください。重い物を運搬する場合は、3名で行うこと。住む所を確保しておく。

Keeps 物 / 所 in kanji (substantive), fixes 事→こと, 下さい→ください, 行なう→行う, ，→、, ３→3.

✗ prh (uniform replacement)

申請の内容を確認することが必要である。事故のとき，速やかに御連絡ください。重いものを運搬する場合は，３名で行なうこと。住むところを確保しておく。

⚠ Over-converts substantive nouns: 物→もの, 所→ところ are wrong (should stay kanji).

Why it matters — the evidence

Hard-set accuracy by usage — **Figure 1.** On the hard set, uniform replacement reaches 100% on the kana-correct side but ~0% on the kanji-correct side (substantive nouns, main verbs). Collapsing a word to one spelling must fail one side.

Findings vs over-flag rate — **Figure 2.** prh over-flags ~43% (it "corrects" already-correct text); Scribe stays in single digits at a comparable number of findings.

Accuracy by usage type — **Figure 3.** By usage type: uniform replacement scores 0% on substantive nouns and main verbs — exactly the kanji-correct usages.

System	Overall	Hard	Precision	Recall	Over-flag
(A) prh uniform replacement	0.56	0.51	0.56	0.56	0.43
(B) Jōyō-kanji rule check	0.50	0.50	0.00	0.00	0.00
(D) Scribe (heuristic layer)	0.94	0.94	0.94	0.94	0.06

Diagnostic on a synthetic seed corpus (CC0); demonstrates the structural claim, not absolute accuracy on natural text. The trained neural model reaches 0.82 span accuracy on a held-out (by-document) test set.

Run the interactive app locally

Hosted Gradio Spaces now require a PRO subscription, so this Space is a static showcase. The full interactive demo runs locally:

git clone https://github.com/NagaYu/scribe-koyobun
cd scribe-koyobun
pip install -e . && pip install gradio
python app.py            # or: scribe check doc.txt --profile jichitai

Apache-2.0 · Built with a rule layer + a 0.1–0.2B token-classification model · Normative documents' rights belong to their issuers (clauses referenced, not reproduced).