Repository navigation
Improve the context for names and genders #7
Description
Hi! First of all, your skill is awesome! I translated a required book and read it full.
But there a few problems what I found, both related to the independent chunks:
- from chunk to chunk the names of the people could be different, like Tai, Taig, Taighi, Taiger etc. If in original the name is Taighi then for e.g. in russian translation it will have not stable translation
- gender in some chunks can be male/female randomly, because of from the context of one chunk is not clear to Claude how to translate correctly.
May I ask you please to somehow fix those problems in new versions of the skill? Maybe it could be solved by creating shared short context of the original book which could be used for all chunks. Or maybe it could be solved on validation step.
Would be nice for e.g. to check the final result to be more organic. Thanks!
Activity
neoOpus commentedon Apr 20, 2026
That's the first thing that came to my mind, as it is obvious that this is going to be a problem the way this project does things... The sub-agents must share some kind of information, and the whole project must have a shared knowledge base; otherwise, the translation quality will suffer. It shouldn't be doing this in a sort of batch processing that every simple script can do, but by doing things intelligently and having an oversight AI that orchestrates the operations between agents by providing them with the proper instructions on how to proceed, while it gathers the most relevant data from the chunks to put in a .md or DB, that distills some of the important information, like characters, events, locations, etc. but since it is not just about literature, then it must be versatile enough to extract other relevant information too and have it shared between all the agents for them to use and do their job properly. And it keeps enriching the database over time and altering, correcting, and improving its contents, and when something that was mis-translated earlier, it must be acknowledged and assigned to another agent so they could fix it.
deusyu commentedon Apr 21, 2026
Thanks, this is a valid issue.
You are right that the current chunk-by-chunk translation can cause:
I am currently working on the first part with a glossary-based consistency pass. The second part will likely need a small shared context layer for each chunk.
If you are OK sharing it, could you tell me which book this happened on, and the source language -> target language pair? That would help me reproduce and test the fix on a real example.
If you prefer not to share the book publicly, a small excerpt with the problematic cases would also be useful.
docs: Add roadmap for issue #7 cross-chunk consistency
docs: Inline roadmap into README and drop standalone file
deusyu commentedon Apr 23, 2026
Quick update — Phase 1 (cross-chunk term consistency, addressing the "Tai / Taig / Taighi / Taiger" drift) is now on
main. Sub-agents emit per-chunk meta files;scripts/merge_meta.pymerges them per batch with conservative alias/confidence rules, and the next batch translates against the updated glossary.I added a regression baseline (
tests/baselines/diligent-dick, commitccf3330) whereDick / Richard / Stuartco-occur across ~9 chunks — same shape as your case. If you can re-translate your original book onmainatccf3330+ and tell me whether the names converge by batch 2, that would be very useful.The pronoun/gender half of your report is not yet fixed — it's Phase 2 (inject neighbor-chunk excerpts as read-only context). If you have one chunk-pair where it bounced, sharing it would directly inform the design.
neoOpus commentedon Apr 25, 2026
maybe opting to translate them to Phonetics then from Phonetics to any language, this will ensure better quality and higher consistency than relying one whatever the model decides to translate names to... Just a thought!
deusyu commentedon Apr 27, 2026
Thanks for sticking with this.
The instinct behind the IPA idea — wanting a deterministic mapping so the model can't invent new spellings on its own — is the right instinct. But I don't think IPA itself delivers on that:
Readis /riːd/ or /rɛd/,Worcesteris /ˈwʊstər/. Disambiguating pronunciation needs context, so it just moves the failure to a new place.Holmesis /hoʊmz/, but Chinese has long settled on "福尔摩斯" (a 19th-century Fujianese transliteration). A phonetics-first pipeline would lose all such established renderings.Sherlock HolmesstaysSherlock Holmesin French/German/Spanish translations — no conversion step exists to standardize.And the drift you originally reported isn't really caused by transliteration quality — it's caused by sub-agents making independent naming decisions in isolation. Phase 1's shared glossary addresses that directly.
That said, your first comment is still very much on point. The "oversight AI gathering shared knowledge into a DB" framing is exactly where Phase 2 is heading. Two variants I'm weighing:
If you have a preference based on your use case — or that pronoun-bouncing chunk-pair from your original book — it would directly shape which path is worth the engineering cost.
neoOpus commentedon Apr 28, 2026
You're absolutely right about everything you just said. And it applies to every language that I know about. I couldn't have said it better.
deusyu commentedon May 23, 2026
Update: the main parts of this issue are now implemented on
main.What shipped:
scripts/chunk_context.pyinjects read-only previous/next chunk excerpts into each translator prompt for pronoun, gender, and entity-resolution context.scripts/run_state.pytracks which glossary/entity state each chunk used, so glossary changes can selectively re-translate only affected chunks instead of redoing the whole book.I also tested this with the checked-in Sleepy Hollow baseline and the full unit suite (
205 tests OK, skipped=3).I am keeping this issue open for now because the remaining part is real-book validation: whether the final translated result feels more organic on books that previously showed name or gender drift. If you can re-run your original book on current
main(52fb0fdor later) and share whether the drift is gone, that would be the best signal for whether this can be closed.docs: Add roadmap for issue deusyu#7 cross-chunk consistency
docs: Inline roadmap into README and drop standalone file