Technical / IT Translation
Translate manuals, specifications, software documentation, and API references while keeping identifiers visible for review.
Why the subject matters
What changes in this kind of work
Technical prose often places ordinary sentences beside commands, variables, model numbers, and measurements. QA checks flag those strings when they change, while glossaries keep component names and interface labels consistent across the document.
In technical and IT documentation the difficult part is rarely the prose. It is the mixture. A sentence that explains how to restart a service sits next to the command that does it, a configuration key, an error message and a version number. The reader copies the command and expects it to work, so the explanation has to change language while the code stays exactly as it was.
DeepReference's model drafts the whole segment, code included, and nothing locks the code: commands, placeholders and file paths are not protected automatically. The quality checks flag tags, numbers and placeholders that changed, and the reviewer compares every code block with the source.
Most of the value in this field comes from consistency across releases. Manuals and help content change a little with every version. A glossary keeps component names and interface labels the same, translation memory (Translator Pro) reuses approved segments whose source has not changed, and revision comparison shows exactly which sentences of a new DOCX version are different. The reviewer's time then goes to the new text.
How it works
A project from upload to delivery
- 1
Prepare the source file
Upload the manual or specification as DOCX, PDF, PPTX or XLSX, up to 50 MB (75 MB on Business). Documentation kept in Markdown or in a help-centre tool has to be exported to one of these formats first. A PDF needs selectable text; an image-only scan is refused.
How to do it, step by step: 6. Document translation
- 2
Set up the terminology
Run terminology extraction on the source, or on an older source and its translation, and approve or dismiss the terms it finds. Add component names and interface labels exactly as they appear in the localised software. An existing termbase can be imported as CSV, XLSX, TMX or TBX.
How to do it, step by step: 3. Terminology extraction
- 3
Translate with the right settings
Choose the target language and a tone, for example Technical. Approved glossary terms that occur in a sentence are sent to the model with that sentence. On Translator Pro, an exact match from translation memory is reused instead of being translated again.
How to do it, step by step: 5. Text translation, tones and glossary
- 4
Review segment by segment
In the segment editor the reviewer sees source and target side by side, with glossary hits and quality-check flags. They compare commands, code and placeholders with the source, correct what changed, and approve each segment, or approve all segments at once after reading them.
How to do it, step by step: 7. Translation Editor
- 5
Check the delivered file
The file comes back in its original format, with tables, headings, numbering, headers and footers kept. Open it and look at code blocks, tables and captions. In a PDF, complex layouts need a manual check and some fonts are replaced by a similar one.
How to do it, step by step: 8. Document Editor
- 6
Update for the next release
When a new DOCX version arrives, revision comparison shows what changed in the source as a redline, a DOCX with tracked changes and a change list. In the mode with translation, used with translation memory on Translator Pro, only the affected segments are retranslated. Nothing is charged until you confirm.
How to do it, step by step: 9. Revision comparison
Document types
What we translate
- User guides and help contentInterface labels in the text must match the software the reader has open, or the reader cannot find the button.
- API documentationEndpoints, parameters and request and response examples stay as written; only the descriptions around them are translated.
- Installation and configuration guidesReaders copy commands, file paths and configuration keys, so one changed character breaks the step.
- Release notesShort and dense; deprecated, removed and breaking change have different consequences and must stay distinct.
- Technical specifications and datasheetsValues, tolerances, units and part numbers are the content; the number check flags changed figures and the reviewer confirms each one.
- Hardware manualsSafety warnings and the order of steps matter more than style; the reviewer checks that no warning or step was merged or dropped.
- Architecture documentsComponent and service names recur many times; a glossary keeps them identical from the first diagram caption to the last appendix.
- Security policies and proceduresWords such as must, should and may set levels of obligation, and a policy changes meaning if they are blurred.
- Engineering drawings and annotationsAnnotations in a PDF are translated only if they are selectable text; labels that are part of an image stay untranslated.
- Network specificationsLong documents with numbered requirements and cross-references. Numbering and headings must survive, and words such as ‘shall’ and ‘should’ must keep their different strength.
- Installation and commissioning guidesCommands, parameter names and values sit inside the instructions, and each one must match the source exactly.
- Alarm and parameter listsOften kept in XLSX. The descriptions are translated and formulas stay in place; the reviewer checks that codes and identifiers are unchanged.
How to do it, step by step: 6. Document translation
Quality controls
Checks available during review
Domain terminology
A domain glossary or termbase can be applied to keep preferred vocabulary consistent during review.
How to do it, step by step: 3. Terminology extraction
Format and structure checks
Document structure, numbering, tables, and section conventions remain visible for reviewer checks.
How to do it, step by step: 8. Document Editor
Translation memory
Approved translations can be reused when the same or similar text appears again.
How to do it, step by step: 4. Translation memory
Review records
Reviewer actions, translation settings, and approval timestamps are kept where the workflow supports them.
How to do it, step by step: 7.3 Project tools
Glossary filtering
Only glossary terms found in the current text are sent with the translation request.
How to do it, step by step: 2.15 Use a glossary correctly
Recommended settings
How to set up a job in this field
Domain for terminology extraction
Choose Technical when you extract terms from these documents. The domain is used for extraction only.
Step by step in the user guideGlossary
Add the approved terms for this client and language pair, one translation per term, and select the glossary when you translate.
Step by step in the user guideTerminology examples
How key terms are handled
API endpoint
Paths such as /v2/orders and methods such as GET stay as in the source. Nothing stops the model from changing them, so the reviewer compares each one.
Placeholder / variable
Strings such as {user_name} or %s must come back unchanged. The quality check flags a placeholder that changed or went missing, and the reviewer restores it.
Authentication / authorization
Checking identity versus granting permission. Some languages use one everyday word for both, so a glossary entry for each keeps them apart.
Latency / throughput
Delay versus volume per second. They stay separate terms, and units such as ms or Mbit/s are not converted.
Deprecated
Still works but will be removed later. It must not become “removed” or “obsolete”; a glossary entry sets the target term.
Breaking change
A change that stops existing integrations from working. The target term is agreed per language, sometimes kept in English, and added to the glossary.
Default value
The literal value (true, 30, en-GB) stays as written. The number check flags changed figures; words such as true are checked by the reviewer.
Interface label (UI label)
Button and menu names should match the localised software. If the product is not localised, the label usually stays in English, and the glossary records that decision.
Error message
Readers search for the exact message the software shows. Many teams keep the original message and add a translation in brackets; the reviewer applies the team's rule.
How to do it, step by step: 2.4 Add terms · 2.15 Use a glossary correctly
Who uses this
Industries served
Software companies, IT service providers, engineering firms, data centers, cybersecurity teams, and standards bodies.
Standards and references
What reviewers can check
Reviewers can check commands, file paths, identifiers, code fragments, units, part numbers, and repeated technical terms.
Before delivery
What your reviewer checks
- Commands, code blocks, file paths and configuration keys are identical to the source.
- Placeholders and variables such as {name} or %s are unchanged and none is missing.
- Version numbers, part numbers, values and units match the source; nothing was converted or rounded.
- Interface labels match the localised software, or stay in English where that was agreed.
- Every segment flagged by the quality checks has been looked at and resolved.
- Glossary terms are used consistently, including in headings, tables and captions.
- Warnings, cautions and the order of steps in procedures are complete.
- Comments inside code samples are translated or left as they are, as the client asked.
- Text inside screenshots and diagrams is not translated; it is listed for the client.
- The downloaded file opens correctly; in a PDF, complex layouts and replaced fonts are checked.
How to do it, step by step: 7. Translation Editor
Transparency
What DeepReference does not do
- Code, commands and placeholders are not protected automatically. The model can translate a comment, a variable name or a word inside a command. The quality checks flag changed tags, numbers and placeholders, but not every change, so a person checks each code block.
- Software resource files such as JSON, YAML, PO, .strings or .resx cannot be translated as files. XLIFF can: the empty targets are filled and the same file comes back (Starter and up). DeepReference translates DOCX, PDF, PPTX, XLSX and XLIFF files and pasted text; it is not a software localisation tool.
- Units, number formats and dates are not converted by any rule. The output is what the model wrote, so the reviewer decides, for example, whether 1.5 should read 1,5 in German.
- Text inside images, such as screenshots and drawings saved as pictures, is not translated. An image-only scanned PDF is refused; upload the original file or a PDF with selectable text.
- There is no connection to Git repositories or help-centre platforms. Files go in and out by upload, or through the REST API on Business.
- There is no separate model for technical or IT content, and DeepReference does not provide translators or technical reviewers. The domain setting guides terminology extraction only.
Questions
Frequently asked questions
Will my code blocks come back unchanged?
Not automatically. The model translates the whole segment, code included, and nothing locks the code. The quality checks flag changed tags, numbers and placeholders, but a person still has to compare every code block and command with the source.
Can we upload JSON, PO or XLIFF files from our software?
XLIFF yes (1.2 and 2.0, Starter and up): the empty targets are filled and the same file comes back. JSON and PO no. Translatable files are DOCX, PDF, PPTX, XLSX and XLIFF, up to 50 MB (75 MB on Business), plus pasted text up to 100,000 characters. Documentation kept in Markdown or a help-centre tool has to be exported to one of these formats first.
Is unreleased product documentation sent to an outside AI company?
No. The model runs on DeepReference's own servers in the EU, and no outside AI company sees the text. Customer content is never used to train the model. Uploads are encrypted in the browser before sending and stored files are encrypted at rest; this is encryption between your browser and DeepReference's servers.
Is there a separate model for technical and IT content?
No. One model translates all fields. The Technical tone changes the register of the output, the glossary supplies your approved terms, and the domain setting guides terminology extraction only. It does not change how the model translates.
How do we handle a new version of a manual?
If both versions are DOCX, revision comparison (Freelance and up) shows the source changes as a redline, a DOCX with tracked changes and a change list. In the mode with translation, with translation memory on Translator Pro, only the segments the revision affected are retranslated. Nothing is charged until you confirm.
Can we connect DeepReference to our build or documentation pipeline?
On Business there is a REST API with API keys, so your own scripts can request translations. There are no ready-made connectors for Git or help-centre platforms, and the API output still needs the same human review.
Review framework
Open the platform from the dashboard when you are ready to work on files.
View quality documentationBegin projects: Technical / IT Translation
Open the platform from the dashboard when you are ready to work on files.