jakequist's picture
Add Kanjiland silver annotations (9388, KFTT-derived, CC-BY-SA 3.0)
45e68c6 verified
|
Raw
History Blame Contribute Delete
2.23 kB
metadata
license: cc-by-sa-3.0
language:
  - ja
  - en
task_categories:
  - translation
  - token-classification
tags:
  - japanese
  - reading-comprehension
  - furigana
  - grammar
  - annotation
  - distillation
pretty_name: Kanjiland Silver Annotations
size_categories:
  - 1K<n<10K

Kanjiland — Silver Annotations

9,388 Japanese sentences annotated in the full Kanjiland reading-comprehension format: morpheme segmentation, furigana (ruby on kanji runs), per-token contextual glosses, word groupings, an English sentence translation, and grammar-pattern labels from a closed 120-rule inventory.

Part of Kanjiland, a from-scratch Japanese reading-comprehension engine.

Format

Each line of silver_annotations.jsonl is:

{"ja": "<Japanese sentence>", "wire": "<tagged annotation string>"}

The wire string uses Unicode Private-Use-Area separators (⟨T⟩ token, ⟨W⟩ word, ⟨S⟩ sentence, ⟨G⟩ grammar). The format is specified in docs/FORMAT_SPEC.md; the grammar rule inventory is docs/GRAMMAR_RULES.md.

How it was made (hybrid supervision)

  • Source sentences: KFTT (Kyoto Free Translation Task), formal/Wikipedia domain.
  • Deterministic layer: MeCab + UniDic (offline) — segmentation, kanji-run ruby, POS, lemma.
  • Judgment layer: an OpenAI teacher model — contextual glosses, translation, grammar labels.
  • Gate: every annotation passes a strict format linter (93.9% of 10,000 sentences passed).

License & attribution

CC-BY-SA 3.0. Derived from KFTT (© Graham Neubig, CC-BY-SA 3.0, http://www.phontron.com/kftt/) — attribute KFTT and this project, and share derivatives alike. English glosses/translations are model-generated (OpenAI); deterministic labels use MeCab (BSD) + UniDic. See the repo NOTICE.md.

Limitations

  • Domain: formal/Wikipedia (KFTT) — not conversational or web Japanese.
  • Silver, not gold: teacher-generated; the linter catches structural errors but not every semantic mislabel. Spot-audited, not exhaustively verified.