The first guide covered the core flow: import, translate, export. This guide goes deeper into machine translation. You will see how dividing a document into sections gives the model the context it needs, and how to choose the right strategy for the document in front of you. The choice comes down to one question: is this document close to work you have already done, or is it new?
Everything in this guide runs on top of the privacy stage from guide 1. Whatever you send to the model is the protected text, never the confidential original.
A language model translates better when it understands the shape of a document. An isolated sentence gives it little to work with. A section — a heading, an introduction, a body passage — gives it purpose and context. bAIbel AV translates section by section, carrying that context into each request, so the result is more consistent than translating one segment at a time.
When you import a document, bAIbel AV divides it into sections for you. You see them as headers in the editor grid. If two sections should be one, right-click and choose Delete Section to merge a section into the one before it.
Each section has its own context, which you edit in the Section Context panel. The panel has three tabs: Prompting, Functional Role, and Glossary.
Three kinds of context make the biggest difference to machine translation. You set them once and they flow into every section’s translation.
Choose Summarize and Tag Entire Document from the toolbar. bAIbel AV reads the whole document and writes a short summary. That summary travels with every section’s translation request, so the model keeps the same understanding from start to finish.
A functional role describes what a section is for — a title, an introduction, a legal clause, a marketing pitch. Telling the model the role shapes the register and voice of the translation. You can let bAIbel AV generate a role from the surrounding text, or write your own. The role is free text, so you are not limited to a fixed list.
| What you set | Where | How it helps the translation |
|---|---|---|
| Summary | Document level, from the toolbar | Keeps meaning and terminology consistent across the whole document. |
| Functional role | Per section, in Section Context | Sets the register and voice that fit the section’s purpose. |
| Additional information | Per section, in the Prompting tab | Adds any extra guidance the model should follow for that section. |
| Glossary | Project level, shown for reference | Holds the approved terminology the translation should use. |
bAIbel AV gives you two ways to translate a document with the model. Choosing well saves time and protects the quality of your approved work. The choice depends on the type of document.
| Document type | Strategy | What it does |
|---|---|---|
| Closely matches your memory | Fix TM Matches | Adapts your existing, approved translations to the small changes in the new source. |
| New, with little memory coverage | Auto Translate | Produces fresh, context-aware translations for empty segments. |
Use Fix TM Matches when the document closely tracks work you have already approved, with only small edits in the new source. Rather than translate again from scratch, bAIbel AV takes your existing translation and adjusts it to match the change. Your approved wording is respected, and drift is kept to a minimum.
Fixing is faster and keeps you close to translations you have already approved. It is the right choice for updated contracts, revised reports, and new versions of documents you have handled before.
Use Auto Translate when the document is new and your memory has little to offer. bAIbel AV translates the empty segments from scratch, using the summary, the functional roles, and the glossary you set up earlier. Large sections are split into manageable batches automatically, so you do not have to manage that yourself.
Real documents are rarely all old or all new. A revised report might reuse most of last year’s text and add a few fresh sections. Fix the parts that match your memory, and Auto Translate the parts that are new. Throughout, the translation memory and the termbase keep working, so approved wording and terminology are applied wherever they fit.
Both strategies send only the protected text to the model. The privacy stage from guide 1 governs everything here, whether you fix matches or translate from scratch.