How it works

How Bunko makes its editions

Bunko (文庫) publishes learner editions of Japanese literature that is in the public domain: the original text with furigana, an English translation next to every sentence, notes on vocabulary and grammar, summaries, a JLPT difficulty rating, and simplified retellings of the easiest stories.

Editions like this used to take a translator and a teacher weeks per story, which is why so few exist. Bunko makes them with AI, and then puts the AI's work through independent checks before anything is published. This page explains exactly what the AI does, what it does not do, and how its work is checked.

What the AI writes

• The English translation of every sentence, written for this site following a brief that fixes the names, register and tone for the whole story.
• The vocabulary and grammar notes on each page, the part and story summaries, the introductions, reading tips and author notes.
• Corrections to furigana where a reading depends on context (一日, 何 or 上手 change their reading with meaning).
• The simplified retellings for N5, N4 and N3 learners.

What is not left to AI

• The Japanese text. It comes unchanged from Aozora Bunko's transcriptions of printed editions.
• Copyright decisions, which follow fixed rules applied to Aozora's catalog data.
• Base readings. They come from the printed source edition's own ruby where it has them, and otherwise from the Sudachi morphological analyser.
• JLPT levels of words, which come from the published word lists, and story difficulty, which is measured from those lists.

How every page is checked

A second AI model reviews every page

The translation and notes are written by one model (Anthropic's Claude Opus). A different model (Claude Sonnet), which did not write them, then compares every English sentence with the Japanese, checks every reading, gloss and grammar explanation, and flags wording that resembles a published translation. Every issue it raises is decided on its merits before the page can go further; a sentence flagged as resembling a published translation is always written again.

A dictionary checks every gloss

Each vocabulary note is looked up in JMdict, the open Japanese–English dictionary, and the reviewer sees the dictionary senses next to the note. Words are tagged with their JLPT level from the word lists, never by the AI.

Fixed rules block anything incomplete

Software checks run on every page before publication and block it if any sentence lacks a translation, any kanji lacks a reading, a note refers to a word that is not on the page, a grammar example is not an exact quote, or the text contains any trace of the production process. Retellings are checked word by word against the JLPT list for their level.

Published pages are checked again

Every week a sample of published pages is reviewed again from scratch. Confirmed errors are corrected, and a page that no longer passes is taken down and rewritten. The live site is also crawled every hour for broken or incomplete pages.

What can still go wrong

AI translation can still miss a nuance, choose a flatter word than the author did, or smooth over a deliberately odd phrase. The translations are meant to help you understand the Japanese, not to replace it: when in doubt, trust the original.

Glosses give the meaning that fits the passage, not a full dictionary entry, and difficulty ratings are estimates made from word lists that do not cover literary vocabulary.

Texts and copyright

The Japanese texts come from Aozora Bunko (青空文庫), the volunteer-run digital library of Japanese literature, and every work page credits the source edition and the volunteers who typed and proofread it. A work qualifies only if Aozora marks neither the work nor the author as protected, and if every credited person died in 1967 or earlier: under Japanese law their works entered the public domain before the term was extended from 50 to 70 years on 30 December 2018, and the extension did not revive them. Translations are excluded, because the translator holds a separate copyright. The decision and its basis are recorded for each work.

JLPT levels

The Japanese Language Proficiency Test has not published official vocabulary lists since 2010, so levels are measured against the widely used community lists. For each story we measure the share of vocabulary at each level, the share of words beyond the lists, the kanji levels and the sentence length, and combine them into a rating: the level at which a learner can read the original comfortably with furigana and notes. The measurements are shown on each story page. The reviewing model also gives its own opinion of the level; if it differs strongly, the rating moves one step towards it.

Sources and credits

• Aozora Bunko (青空文庫) — Japanese source texts, transcribed and proofread by volunteers.
• JMdict — Japanese–English dictionary by the Electronic Dictionary Research and Development Group, used under the Creative Commons Attribution-ShareAlike 4.0 licence, via the jmdict-simplified project.
• JLPT vocabulary lists — Community lists by Jonathan Waller (tanos.co.uk), in the JMdict-linked edition of the yomitan-jlpt-vocab project, CC BY-SA 4.0.
• Kanji levels — Kanji data by David Gouveia (kanji-data, MIT licence), based on KANJIDIC and Waller's lists.
• Sudachi — Japanese morphological analyser by Works Applications (Apache 2.0).

The English translations, notes, summaries and retellings on this site are original works of Bunko, written with AI and checked as described above.