Skip to content
Interface language: English
Audit / Signal

AI document translation: checking names, numbers and meaning

Review the meaning, numbers and names in AI translation. Prepare a glossary and check for omitted conditions or passages.

Article language: English

A gardener adjusts the water flow at a prototype made from roots beneath a red greenhouse canopy.
In this article

A comma in a value can matter more in translation than an elegant paragraph. A model name, unit and “draft” label must also arrive intact. When accepting an AI translation, compare its meaning with the source: completeness, names, numbers, conditions and the scope of obligations. A beautiful sentence with the wrong unit still sounds good. It is less useful.

Before starting a tool, define the translation's purpose and use. A working aid for understanding needs different acceptance from a document given to a client or used to take action. AI can prepare a draft, but the person approving it needs access to the source and skills appropriate to the intended use.

Define what must remain unchanged

List items requiring special attention: names of entities, products and parts, identifiers, amounts, units and dates. Record passages that should not be translated and the expected terminology. Do not assume a model recognises a proper name merely from a capital letter.

For a technical document, specify what terms mean in that field. “Port” can describe different things; a correct general dictionary does not settle the context. A concept map helps the model, while an explicit criterion helps the reviewer judge the word choice.

Also establish the permissible scope of editing. A translation should not independently shorten a list of conditions or resolve ambiguity by choosing one interpretation. An ambiguous source may need a question for its author. A translator need not solve a puzzle with an elegant sentence.

A glossary is a team agreement

Prepare a small glossary with an approved equivalent and an example of use. Include context and who resolves disputes. Two columns may not suffice for a term with several meanings. The model and reviewer should receive the same version. Check inflection and actual use in sentences: a correct dictionary equivalent can be wrong in a particular passage, so a term list does not end the review.

Microsoft describes a custom document translation glossary that can specify equivalents, names left untranslated and capitalisation. This is a feature of a particular product. For another tool, check how the glossary is supplied and whether it actually affects the result.

Place the approved glossary in the company knowledge architecture, with an owner and update rules. The guide to organising company knowledge (in Polish) explains how to identify sources, owners and review dates. Do not replace it with a private employee conversation the rest of the team cannot later reproduce.

Check material before translation

A document may contain digital text, a scan or both. If source extraction confuses a digit or omits a paragraph, fluent translation cannot repair the input. Separate checks of OCR (optical character recognition) from review of the translation itself. The guide to OCR quality checks before AI (in Polish) explains what to check in the extracted text.

Google's document translation documentation identifies limitations in preserving the layout of scanned and complex PDFs and differences by file type. Before choosing a workflow, check support for the specific format rather than assuming every accepted file was processed completely.

On a sample, compare titles, paragraphs, notes, table entries and text in graphics. Check whether the model received material in an order that preserves relationships. Two columns read alternately can produce sentences the author never wrote.

Check names and numbers in a separate pass

After translation compare values with the original, including minus signs, decimal separators, ranges, units and the association of a number with the correct item. The presence of every digit does not prove correctness. An amount attached to another service option remains wrong.

Adapt dates unambiguously for the reader while preserving day, month and year. If the source is unclear, ask for confirmation rather than guessing the author's local convention. Preserve agreed spelling and links to a case or product identifier for names.

For an AI-assisted sales proposal, check the separation of recurring and one-off items, variants and conditions. The article on checking AI-generated sales proposals (in Polish) discusses how to verify commercial commitments. Translation should not change the price even if the new version looks more persuasive.

Read conditions and negatives carefully

Mark sentences containing “if”, “unless”, “before”, “after”, “at least” and “no more than”. Compare which action they qualify in both versions. A small change in a condition's scope can reverse a paragraph without a grammatical error.

Use an artificial instruction in an exercise: an action may start only after confirmation by a specified person. Check that translation preserves the sequence, need for confirmation and person's role. It is not enough for all three to appear somewhere; their relationship must remain the same.

In your team’s AI verification training, include pairs of apparently similar sentences. The guide to training a team to check AI answers (in Polish) describes how to practise verification. Reviewers need to recognise the difference in meaning. When a document needs specialist knowledge or a particular translation procedure, provide the right person and process rather than extending the tool's role.

Assess fluency after fidelity

Readable syntax and natural language matter, but after content is preserved. Separate the stages. First check against the source, then improve rhythm, unnecessary repetition and terminology. After editing, revisit changed passages for meaning.

Back-translation, translating again into the original language, can reveal a problem but does not replace acceptance by a competent person. Two systems can make similar errors or smooth over the same imprecision. Treat it as an additional clue, not automatic proof of correctness.

Pay particular attention to customer-facing replies, such as complaint handling. A polite request can become a firm refusal without changing the message's main topic. Tone and function are part of meaning. In the AI complaint-handling article (in Polish), a customer's account is separated from confirmed facts.

Maintain test cases and an acceptance record

Keep artificial documents showing common traps: a proper name, ambiguous term, numbers and a condition. Include expected resolutions. Add them to AI quality regression tests, particularly after changing the model, glossary or file-reading method.

Aeneia conceptual photography

For a method of comparing results after a change, see the article on AI regression tests (in Polish).

In AI supplier evaluation, check processing terms appropriate to the data. When minimising material, retain translation context while limiting unnecessary information. The guide to data minimisation before using AI (in Polish) discusses preparing the material. Cutting half a sentence can remove a confidential detail and destroy meaning at the same time. For questions about access and retention, see the AI supplier assessment guide (in Polish).

Add source, glossary and translation versions and the acceptance scope to AI audit materials. Include review and correction time when measuring benefits. Receiving a document quickly is the start of acceptance, not its result. For the distinction between active work and waiting time, see the AI work measurement article (in Polish).

The guide to preparing AI audit materials (in Polish) explains how to organise versions and the scope of what is shared.

Review passages together with their relationships

Splitting a long document helps work, but can separate a definition from later use. Decide what context the model receives for each part. Names, references and numbering must remain consistent. Separately correct sentences may not form a correct document if each describes the same item differently.

In an artificial example define a term in the first paragraph and use it in several later parts. Change the order in which parts are supplied. Check whether the glossary and context preserve meaning. This can reveal dependence on an earlier conversation that would be missed if the same file were always translated from beginning to end.

Check references such as “the above option”, “this term” and “except as stated below”. A changed layout can make them point to the wrong passage even when the sentence itself is faithful. Compare headings, numbering and connections after export. Acceptance covers relationships between elements, not just a list of correctly translated paragraphs.

When several people collaborate, agree terminology decisions. If one part uses another equivalent, check whether it really means the same thing. Consistency should not erase a meaningful distinction in the source. Different names can be deliberate and should sometimes be preserved rather than automatically replaced by one favourite term.

Collect questions for the author in a separate list. Do not hide ambiguities in notes that may disappear during later formatting. The source owner decides the meaning; the translator carries it to the right places. The next version need not guess the same issue again.

Check the finished file through the reader's eyes

Open the document in the form in which it will be used. Check that text fits fields, does not overlap values or disappear at page ends. The same information can need different space in the target language. If material mixes writing directions, separately inspect names, codes and addresses in bidirectional text (in Polish). A correct translation hidden outside a box cannot help its reader.

Compare tables row by row. A heading must still describe the right column and a note remain attached to the right item. Inspect labels, legends and text in images. Do not judge completeness from page count: layout may change without content loss, or remain similar despite an omission.

Check links and visible addresses where used. Translating a label should not change a link's destination without agreement. File names and version labels should identify source and translation. Avoid several different documents with the same name when language versions are prepared in parallel.

If someone changes the source after translation, identify the effect on the target version. Not every correction requires translating everything again, but every material change needs to reach the right passage and be reviewed again. Otherwise two fluent documents may describe different conditions. Separately review the AI-assisted comparison to identify deletions and conditions moved between versions. The article on checking AI-assisted document comparison (in Polish) discusses how to assess the result.

Before handover, ask an authorised person to perform a specific action using an artificial document: identify the deadline, starting condition and next step. If their understanding differs from the source, revisit text and layout. This practical usability test supplements linguistic review.

Match the review to the translation's use

For preliminary understanding, a draft with clearly described limitations may be acceptable. When a document will drive action or go to an external reader, set acceptance appropriate to the consequences of error. Do not judge by length alone. One sentence containing a condition may need more attention than several pages of general description.

The reviewer should know both languages and the necessary field context or have appropriate support. Conversational fluency alone does not establish ability to assess specialist terms. If the company lacks the competence to accept a kind of material, limit use of the draft or obtain an appropriate service. Fluency assessment cannot replace missing knowledge.

Decide which errors block use and which can be corrected in ordinary editing. A wrong product name, number or negative should trigger checks of similar places, not just repair of the one instance noticed. Investigate whether the cause lies in the glossary, source extraction or the way the document was split.

The acceptance record should identify the version, purpose and scope checked. If only selected parts were assessed, preserve that information with the document. Do not present partial checking as approval of the whole. A clear boundary helps the reader decide whether it suits the intended work and what further checking is needed.

Use shared language for feedback. “Sounds odd” calls for different work from “the starting condition changed”. Identify the kind of issue and source passage. This separates style editing from repair of meaning and helps select the right person to decide.

In a short trial ask the reviewer to explain one important sentence in their own words, then compare the explanation with the original. Fluent translation can conceal misunderstanding; a question about meaning can expose it quickly. This supplements, not replaces, full review.

Finally, identify which version goes to the recipient and who may change it. Changes after approval need appropriate review again. Otherwise one document can be carefully checked while another, just polished, is sent. A sentence can be smooth as a mirror yet reflect a different meaning from its source.

To examine such a process in your company, see Aeneia Audit.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear