What Is Glossary AI? The Engine Behind Consistent Novel Translation

In Korean, 이현 is a perfectly ordinary character name. In MTL output, it can become Lee Hyun in Chapter 1, Yi Hyeon in Chapter 5, and Ihyeon in Chapter 10 — three different people, by the time the reader gets to the climax, who are actually the same person.

This is the kind of thing that turned reading machine-translated novels into an act of patience for years. Names drift. Genders flip. A "Headmaster" becomes a "Principal" three chapters later. By page 200, you're keeping a mental scoreboard instead of enjoying the story.

We built Glossary AI to solve it. This post is the long version of what it is, how it works, and why fiction needs it more than any other kind of text.

The consistency problem nobody had really solved

Every general-purpose translation tool — Google Translate, DeepL, even raw ChatGPT or Claude — translates text in chunks. A paragraph here, a chapter there. They are excellent at producing fluent prose for any given chunk. They are terrible at remembering that the chunk they translated yesterday referred to the same character with a specific romanization.

Long-form fiction breaks this. A novel is hundreds of pages of recurring people, places, organizations, sects, factions, made-up terminology, cultivation realms, magic systems, in-jokes, family relations, and titles that all need to land the same way every single time. A single character's name can appear hundreds of times across a book. If a tool picks a slightly different transliteration on attempt 437, the reader notices.

The result, for years, has been the same set of community complaints we kept hearing across every web-novel and light-novel community we talked to: inconsistent naming, misgendered characters, and forgetting key context. We noticed it as readers ourselves. So we went about solving it.

What Glossary AI is

Glossary AI is the system inside AI Novel Translation that automatically builds and maintains a per-project dictionary of every named entity in your novel — characters, places, organizations, unique concepts, cultivation terms, magic systems, in-world jargon — and uses it to translate those terms the same way every single time, across every chapter, in every batch.

We pioneered the approach. It's now the thing that makes a Glossary AI–assisted translation feel like a real book instead of a machine dump.

How Glossary AI works (the four-step loop)

Glossary AI runs every time you translate a chapter. The loop is:

Scan the source text for key terms.
Check those terms against your existing glossary. Anything already known is locked in. Anything new is flagged.
Propose new entries with suggested translations and gender tags.
Show you the proposed entries before translation begins so you can accept, edit, or override.

After that, the translation runs — and every term in the glossary is translated the exact same way it was last time.

The glossary grows with each chapter. By the time you've translated five chapters, it has captured most of the recurring vocabulary of the book. By chapter twenty, the model is rarely surprising you with anything new — and when it is, that's exactly the moment you want to be paying attention.

You can manage the glossary directly, too. Add terms manually in the input row at the top of the table. Edit any cell by clicking on it. Delete a row with one button. Search and filter to find a specific term across thousands. Import a glossary as a CSV — useful if you've already been keeping one, or starting from a draft generated elsewhere. Export it to back up, share with a collaborator, or use as a head-start for a sequel.

One practical detail that matters:

One project, one language pair. Glossaries are tied to a project, and projects translate a single source language into a single target language. If you're going from Korean into both English and French, that's two projects with two glossaries. You can export and import to share terms between them, but they stay scoped — which is the only way the consistency guarantee actually holds.

Why long-form work needs this more than any other kind of text

Most translation use cases are short. A business email. A product description. A page of legal copy. The model sees the whole thing in one pass, makes its choices, and you move on. Consistency is essentially trivial.

Book-length work is the opposite of that. The defining feature of a book — a novel, but equally a textbook, a business report, or any other long-form document — is long-running internal consistency. A character introduced in Chapter 2 has to be the same person in Chapter 47. A faction's name has to be recognizable across an entire trilogy. A defined technical term in chapter one shouldn't quietly change wording by chapter twelve. Fiction is the hardest version of this problem — invented names, worldbuilding, honorifics — but any long document with recurring terminology runs into the same thing.

This is why pasting a book into a generic translator — or worse, into ChatGPT a chapter at a time — produces output that sounds fluent and is functionally unreadable. Each chunk is translated correctly. The document, as a coherent whole, is not.

Glossary AI is the missing component. It's the part that holds the book in its head while every other piece of the pipeline does its job.

How Glossary AI fits with the rest of the stack

Glossary AI is the consistency layer. It works alongside the rest of what AI Novel Translation does:

Structure AI detects chapters and section breaks in raw text and EPUBs, so the document comes out organized, not as one wall of paragraphs.
Editor AI runs a second pass over the translated text, catching awkward phrasing, missed nuance, and small errors.
Batch Translation runs all of it across 500+ pages at once, with format preservation for EPUB, DOCX, and TXT.
Custom Instructions let you steer voice, register, and stylistic choices on top of all the above.

You can use any of these on their own. But together they're the pipeline that has translated over 2 billion words for 10,000+ users and 10,000,000+ readers — and the reason a 90,000-word book can come out the other side as a single coherent novel rather than 90,000 disconnected words.

What this looks like in practice

A few of the user-visible outcomes once Glossary AI is doing its job:

Character names stay locked in across every chapter, no matter how many times they appear.
Honorifics, titles, and forms of address stay consistent — the same source term renders the same way every chapter, instead of one character being "Joe" in chapter 3 and "Joseph" in chapter 9.
Made-up worldbuilding terminology (cultivation realms, sect names, factions, magic systems) gets translated once and then stays translated that way.
Genders stay correct, which sounds basic until you've read the alternative.
When a new term shows up that the system isn't sure about, you see it before the translation happens — not after.

It's not glamorous and it's not flashy. It's the part that takes machine translation of long-form fiction from "technically possible" to "actually readable." Which, after a decade of MTL readers patiently inferring around drift errors, is the whole point.

Frequently asked questions about Glossary AI

What's the difference between Glossary AI and a regular glossary?

A regular glossary is a list of terms you maintain by hand. Glossary AI builds, maintains, and applies that list automatically. You can still edit anything you want — but you don't have to start from a blank page or remember to add the side character introduced on page 312.

Can I just use ChatGPT or Claude to translate my novel with a glossary?

You can try. The practical problem is that general-purpose chat tools don't reliably hold a long glossary across many separate translation calls, they lose formatting on every paste, and on the consumer interfaces they may train on your manuscript. AI Novel Translation doesn't train on your work, preserves EPUB and DOCX formatting, and applies the glossary on every chapter without you re-pasting it each time.

Do I have to build the glossary before I translate?

No. Glossary AI builds it during translation. The first chapter usually introduces most of the main cast and core terminology, so you'll have a working glossary to review within a chapter or two. Reviewing earlier saves work later — fixing a character's name in Chapter 1 is much cheaper than fixing it everywhere it appears in Chapter 47.

Can I import a glossary I already have?

Yes. Upload a CSV with three columns — original_language, translated_language, gender — up to 20MB. You can append to an existing glossary (duplicates are skipped) or replace it entirely. Useful if you've been maintaining a glossary by hand, picking up a series mid-translation, or seeding a sequel project from the first book's terms.

What about books in multiple target languages?

Each project is tied to one source language and one target language. Translating a Korean novel into both English and Spanish is two projects, each with its own glossary. You can export terms from one and import them into another to bootstrap — but they stay scoped to a single language pair, which is what keeps consistency consistent.

How much does it cost?

Pricing is based on character count. A 50,000-word book — roughly 300,000 characters — runs about $26 end-to-end. New accounts get free credits, so you can try Glossary AI on your own work before paying anything.

Try it on a chapter

The fastest way to see what Glossary AI actually changes is to translate one chapter with it on, and one chapter with a tool that doesn't have anything equivalent, and compare. We'll save you the suspense on which one reads better at chapter twenty — but you should see it for yourself.

Questions or feedback? Reach us at support@ainoveltranslation.com.