CoolFace
Datasetpublic

NagaYu/scribe-koyobun-usage

Scribe usage-judgment dataset Span-level data for judging context-dependent kanji/kana usage in Japanese official writing. Generated by scripts/build_dataset.py in the GitHub repo. What the claim rests on The center of this dataset is the hard split: occurrences of words that appear in both usages (kana and kanji) across the corpus. Because uniform dictionary replacement collapses a word to a single spelling, it is structurally forced to mislabel one side of this… See the full description on the dataset page: https://huggingface.co/datasets/NagaYu/scribe-koyobun-usage.

sourceHugging Facecc0-1.0updated 10d agoView on Hugging Face
0likes37downloads
Dataset Card

Scribe usage-judgment dataset

Span-level data for judging context-dependent kanji/kana usage in Japanese official writing. Generated by scripts/build_dataset.py in the GitHub repo.

What the claim rests on

The center of this dataset is the `hard` split: occurrences of words that appear in both usages (kana and kanji) across the corpus. Because uniform dictionary replacement collapses a word to a single spelling, it is structurally forced to mislabel one side of this split. Accuracy here is the headline metric.

Fields

fielddescription
textthe sentence
target_start / target_endcharacter offsets of the target word (half-open interval)
surfacesurface form of the target (the wrong spelling in flipped examples)
lemmacanonical form (こと, いただく, …)
categoryformal_noun / aux_verb
gold_kindnorm-correct kind, kana / kanji (determined by context, independent of surface)
label0 = kana, 1 = kanji
clause_idsupporting clause (C-KEISHIKI-1, …)
flippedwhether this is a flip-generated error
is_hardwhether it belongs to the hard split
doc_idsplit unit (a sentence and its flip stay in the same split)
source / licenseprovenance and terms

How it is built

  1. 1.Positives — official-style sentences authored for this project (SEED_SENTENCES in scribe/collect.py), annotated with each target's correct kind. Spans are derived from a morphological analyzer and cross-checked against the annotation, so offset mistakes are impossible.
  2. 2.Flipped errors (scribe/flip.py) — flip only the target to the opposite spelling, preserving position and gold label. Surface changes but context does not, so a model must learn from context rather than memorize surface.
  3. 3.Hard tagging (scribe/collect.py:mark_hard) — mark occurrences of words that appear in both usages.
  4. 4.Split — by doc_id into train/validation/test (no flip leakage). hard is a separate diagnostic split that cross-cuts the others (it overlaps train), so evaluate generalization on test and use hard as a diagnostic.

License and redistribution

  • —All bundled data is synthetic seed text (CC0-1.0); no body text of government documents is included.
  • —For real-corpus training, see DocumentSource / KNOWN_SOURCES in scribe/collect.py and re-confirm each source's terms at fetch time. Non-redistributable sources are used for local training only, never stored. Provenance and terms are recorded in sources.json.

Sources of the norm

  • —"Considerations for Creating Official Documents" (Council for Cultural Affairs recommendation, 7 Jan 2022)
  • —Cabinet Directive No. 1 of 2010 on kanji use in official documents — the operative lists of words written in kana / kanji
  • —"Rules for Okurigana" (Cabinet Notification) and the "List of Jōyō Kanji" (Cabinet Notification No. 2)

This dataset carries mappings to clause numbers but does not reproduce the normative text verbatim.

Limitations

  • —The seed sentences are a small, controlled set and do not represent the distribution of natural text.
  • —The monitored vocabulary is limited to data/norms/lexicon.json (representative formal nouns and auxiliary verbs).