NagaYu/scribe-usage-classifier

scribe-usage-classifier

The model layer of Scribe, a context-aware kanji/kana checker for Japanese official documents.

Official-writing norms require the same word to be written differently depending on its syntactic role — a formal noun こと in kana, the substantive 事 in kanji. Uniform dictionary replacement collapses each word to one spelling and is structurally forced to get the other usage wrong. This model judges each occurrence of a monitored word fromits context and decides kana or kanji; the Scribe pipeline then attaches the normative clause as the rationale. A dictionary or regex cannot make this call, because the input surface may itself be the error.

Measured span accuracy on the held-out test split (split by document): 0.818 overall, 0.800 on the hard set of words that appear in both usages. Prooff-of-concept scale: finetuned on a small synthetic corpus (CC0), so read these as evidence the distinction is learnable from context — not as production accuracy on natural text. Full project, benchmarks against uniform replacement (prh) and a jōyō-kanji checker, and usage: https://github.com/NagaYu/scribekoyobun

• Part of the Scribe project · 💻 GitHub · 🤗 Space · 📚 Dataset
• Base model: google-bert/bert-base-multilingual-cased (~178M params, within the 0.1–0.2B target). Chosen because its fast tokenizer returns offset_mapping, which the span-labeling pipeline needs.

Labels

Per subtoken: O / TARGET-KANA / TARGET-KANJI. Subtokens overlapping a monitored word's span receive the usage label; the input surface may be correct or wrong (the training data includes flipped errors), so the model must learn from context, not surface.

How to use

The model is meant to be driven through the Scribe pipeline, which finds the monitored words and attaches the normative clause as rationale:

from scribe import ScribeChecker
checker = ScribeChecker(model_dir="NagaYu/scribe-usage-classifier")
for f in checker.check("確認する事がある。重い物を運ぶ。"):
    print(f.surface, "→", f.recommended_surface, "|", f.message)

Raw transformers use is also possible (AutoModelForTokenClassification), but you then need to locate the target spans yourself; the Scribe repo does this with a morphological analyzer.

Training data

Fine-tuned on the Scribe usage-judgment dataset (NagaYu/scribe-koyobun-usage), a synthetic seed corpus (CC0) of official-style sentences plus flip-generated errors, split by document. The dataset separates a hard split (words appearing in both usages) as the headline diagnostic.

Evaluation

Span-level accuracy (predicted kind vs. gold kind):

Evaluate test for generalization; hard cross-cuts the training split and is optimistic. For the project-level comparison against uniform replacement (prh) and a jōyō-kanji checker, see the benchmarks.

Intended use & limitations

• Intended: assisting drafters of Japanese official/public documents by flagging context-dependent kanji/kana choices with a cited clause of the norm; output is phrased as "the norm would write it this way," never as a verdict on the writer.
• Proof-of-concept scale: trained on a small synthetic corpus. It demonstrates that the distinction is learnable from context, but is not production-grade on natural text. For real deployment, train on a licensed real corpus (see scribe/collect.py) and expand the monitored vocabulary (data/norms/lexicon.json).
• Coverage: limited to the monitored formal nouns and auxiliary verbs; other orthographic issues (numbers, punctuation, okurigana) are handled by Scribe's deterministic rule layer, not this model.
• The base model is multilingual; a dedicated Japanese encoder may improve quality if paired with a custom offset aligner.

License

Apache-2.0. The normative documents' rights belong to their issuers; this model references clause numbers only.