NewBilingual Book Maker
EPUB translation tool that can ask Jev which markup groups contain text worth translating.
Added to Jevfast
NewHow it uses Jev
Bilingual Book Maker can use Jev to classify uncertain EPUB text signatures in plan mode. It batches up to 12 signatures into one TypeSafe request; each signature gets a Choice among `translate`, `skip`, and `unsure`, using sample text and structural context. Jev classifies content only; a separate translation backend translates selected text.
What Jev decides
Decide which EPUB text signatures should be translatedExample answers · not a recorded Jev response · Source ↗
Question 1 · {signature key}
YOUR APP
INSTRUCTION
Classify this signature's samples as reader-facing content to translate or material to keep unchanged. The internal `unsure` abstain is not offered as a choice to Jev.
STATE
The shared state is the caller's context containing a page of at most 12 candidate signatures. Each has a key, occurrence and character counts, percentage and mean length, optional parent signatures and conditional CSS, and sample lines. Jev receives one Choice question keyed by each signature ID; the Jev adapter omits the abstain option from the offered choices and handles uncertainty through the confidence fallback.
JEV · CHOICE
- translate
- skip
App workflow
- Parse and collect EPUB text
The loader groups repeated text into signatures and collects occurrence statistics, structural context, and sample lines.
- Classify the plan
With Jev selected as the classifier, the app sends undecided signatures in pages of up to 12. Decisions are stored in the plan; `unsure` leaves a candidate undecided, and low-confidence non-fallback choices can fall back to `translate`.
- Translate selected content
The translation stage uses the saved plan to decide which text to translate. Jev supplies the classification, not the translated prose.
Why it is interesting
This is a targeted publishing workflow where a small decision controls what enters a potentially large translation job. The request uses book structure and samples to distinguish reader-facing prose from page furniture, and the plan records decisions for audit and resume. The confidence gate is deliberately asymmetric: low-confidence non-fallback choices revert to translation so useful content is not silently dropped.