User guide · Part 1 bAIbel AV
English
Draft

Translate a document, start to finish — privately

bAIbel AV works like the CAT tool you already know. You import a file, translate it with a translation memory and a termbase, then export. The difference is privacy. Before any text reaches a language model, bAIbel AV can hide the parts that must stay confidential. This guide walks the whole flow and keeps that privacy thread in view at every step.

The pipeline at a glance

Every project moves through four stages. Privacy is one of those stages, not an afterthought.

Figure 1. The four stages of a project. You import a file. You apply privacy so confidential content is hidden. You translate with the translation memory, the termbase, and machine translation. You export the finished file, with the real content restored.

Create a project and import your file

Open the Projects panel and choose Create New Project. A project holds your files, your translation memories, your termbases, and your privacy settings together.

Figure 2. The Projects panel. Use Create New Project to start, then open the project to add files.

Add your documents to the project:

  1. Choose Import Files for one or more documents, or Import Directory for a whole folder.
  2. For each file, set the language pair — the source language and the target language.
  3. Confirm the import. bAIbel AV reads the document and splits it into segments for you to translate.

bAIbel AV imports the common professional formats, including Word (.docx), PowerPoint (.pptx), and the XLIFF family used to exchange work between translation tools.

Figure 3. The import dialog. Each file gets its own source and target language.

The confidentiality promise

This is what sets bAIbel AV apart from an ordinary CAT tool. Machine translation sends text to a language model. For confidential work, that text must not contain the names, figures, or wording that identify your client or their business. bAIbel AV solves this by hiding sensitive content before the model ever sees it.

The core promise

The original confidential content is not sent to the language model. For a hidden name, the model receives a suitable stand-in instead. That stand-in is built from the attributes you define and the surrounding context, so the translation still reads correctly. After translation, bAIbel AV puts the real content back.

One honest caveat

While you work, prompt content — segment text, glossary terms, and client material — is written to disk in readable form so the app can run. For confidential work, turn this off where the app offers it, and clear these files when you finish. bAIbel AV tells you this plainly in the app, and so do we here.

bAIbel AV gives you three ways to protect content. They work together.

FeatureWhat it hidesExample
Privacy Profile Named entities — people, organisations, places, and similar. A client’s company name becomes a believable stand-in.
Numerical Obfuscation Numeric values — currency, percentages, and counts. A revenue figure is scaled or replaced, then restored exactly.
Boilerplate Text Obfuscation Repeated template wording that should not leave your machine. A standard confidentiality clause is masked.

You manage all three from the Privacy tab, under Privacy and Confidentiality. Each has a Change… button so you can attach a profile to the current project.

Figure 4. The Privacy and Confidentiality panel. Anonymization, Numerical Obfuscation, and Boilerplate Text each have their own profile.

Set up a Privacy Profile

A Privacy Profile decides which named entities to hide and what to replace them with. You build one in three screens.

Screen 1 — Classify Privacy Entities

bAIbel AV scans your source and lists the entities it found. You sort the ones that matter into categories such as person, organisation, or location.

You decide what matters

You do not have to classify every entity. Sort only the terms that must be hidden, and leave the rest unclassified. You can come back and classify more at any time.

Figure 5. Classify Privacy Entities. Move the entities that need hiding from the unclassified list into a category.

Screen 2 — Set Entity Attributes

For each entity, you describe its attributes — for a person, that might be a title, a role, or a gender. These attributes are what let the model produce a believable stand-in without ever seeing the original.

Figure 6. Set Entity Attributes. Select one or more entities and define the attributes that describe them.

Screen 3 — Set Alternative Names

Finally, you choose the stand-in each entity will use. You can type one yourself, copy and replace, or let the model generate alternatives from the attributes you set.

The original never leaves your machine

Even when the model generates the alternatives, the original entity is not sent to it. The model works only from the attributes you defined and the context you provided.

Figure 7. Set Alternative Names. Use Generate for Selected to have the model suggest stand-ins from your attributes.

Hide the numbers, too

A name is not the only thing that can identify a client. A revenue figure or a market share can do the same. Numerical Obfuscation protects numbers while keeping the document usable for translation.

You pick a strategy that suits the work. Two are worth knowing first:

StrategyUse it when
Placeholders (max safety for LLM) You want the strongest protection. Each number becomes a unique token and is restored from a saved mapping.
Scaling (preserve ratios) The relationships between numbers must stay intact. Every value is scaled by one factor.

Before you commit, the Preview step shows the document obfuscated, side by side, or original, and lets you run Test Restore. Restore proves the real numbers come back even after a model rephrases the text around them.

Figure 8. The Numerical Obfuscation preview. Test Restore confirms the original values are recovered.
More detail when you need it

Numerical Obfuscation has a full seven-step wizard with several more strategies. A separate guide covers it in depth. For the core flow, a Privacy Profile plus one of the two strategies above is enough to get protected.

Translate with the memory and the termbase

With privacy in place, translation works just as you would expect from a CAT tool. Everything the tools see is already protected.

Attach your resources

In the project, add a translation memory with Add, New, or Import TMX. Add a termbase with Import MultiTerm XML. Read and write permissions are set per resource, so you control what gets updated.

The editor

The editor shows your document as a grid of segments, with the source on one side and your target translation on the other. Two views are available, and you switch between them freely:

A status bar tracks your progress, counting segments that are unconfirmed, queried, locked, and confirmed.

Figure 9. The editor. The view toggle switches between Standard View and Formatting View. The status bar tracks unconfirmed, queried, locked, and confirmed segments.

Translation memory matches

As you move through segments, the TM Results panel shows matches from your memory. Each match carries scores that tell you how close it is. Insert the match you want into the target with Ctrl+1 through 9.

Figure 10. The TM Results panel. The scores rank each match: Trad is a fuzzy-match percentage, and the others measure retrieval strength and edit distance.

Termbase hits

The Termbase Hits panel shows approved terms found in the current segment. Insert a term at the cursor with Alt+Shift+1 through 9. You can add new pairs to the project glossary as you go.

Figure 11. The Termbase Hits panel. Approved terminology is one keystroke away.
Privacy stays on through translation

Because content was protected at the privacy stage, the memory, the termbase, and machine translation all work on the protected text. Confidential names and numbers never travel to the model.

Export the finished file

When the translation is ready, right-click the file and choose Export File. bAIbel AV restores the real content as it writes the deliverable, so what you hand over reads correctly — the stand-ins are gone and the originals are back.

A second option, Export Anonymized Source File, deliberately keeps the content hidden. Use it when you need to share the source for review without revealing confidential material.

Figure 12. The file menu. Export File produces the restored deliverable; Export Anonymized Source File keeps content hidden on purpose.
Watch for export warnings

bAIbel AV may warn you before export — for example, if some segments still need attention, or if required formatting codes are missing in Formatting View. Resolve these so the exported file is complete.

Terminology used in this guide

Project
A workspace that holds your files, translation memories, termbases, and privacy settings together.
Language pair
The source language and the target language set for a file.
Privacy Profile
A saved set of rules for hiding named entities and choosing their stand-ins.
Numerical Obfuscation
Hiding numeric values, such as currency and percentages, in a way that can be restored.
Translation memory (TM)
A store of previous translations that bAIbel AV matches against your current segments.
Termbase
A store of approved terminology that bAIbel AV surfaces as you translate.
Standard View
The editor view that shows plain source and target text.
Formatting View
The editor view that shows inline codes and tags for precise formatting.