Field Guide
Writing Manuals for Translation
The Short Answer
Translation Cost Is Set at the Source
A translator, or a translation tool, can only work with what the source gives them. Text that is consistent, unambiguous, separate from its layout, and divided into pieces that change independently is cheap to translate and easy to check. Text that repeats itself in slightly different words, hides inside pictures, or is assembled from fragments is expensive in every language, every time it changes.
None of that needs a translation project to start. It is authoring discipline, and it pays back in the first revision cycle. This guide gives no prices because they vary by language, subject, and provider; it explains what moves the cost up or down so you can ask better questions.
Cost Drivers
What Actually Moves the Cost
- Volume
- Every word that reaches a translator is multiplied by the number of languages. Duplicated chapters, copied warnings, and near-identical variant manuals all count as new work unless something recognizes them.
- Repetition
- Translation tools store past translations and offer them again when a sentence repeats exactly or nearly. The more your source repeats itself on purpose, the more of each order is already answered.
- Terminology
- A part called three different things in the source gets three translations, and each needs a decision, a review, and a correction when the customer asks why. One name per thing is a cost control as much as a quality rule.
- Text inside graphics
- Words drawn into an illustration cannot be extracted. Someone must find them, retype them, and redraw the art for every language and every revision.
- Layout and rework
- Translated text is rarely the same length as the source. Tables, callouts, and tight columns that only fit the original language mean hand layout work after translation, on every language.
- Revision scope
- If a one-line change sends the whole document back, you pay for the whole document. If it sends one topic back, you pay for one topic. How the source is divided decides which of these happens.
- Ambiguity
- A sentence that can be read two ways comes back as two questions to the author, or as a confident wrong answer that is found in the field. Clear source removes both.
The case for a controlled source language rests mostly on the repetition, terminology, and ambiguity rows above; our guide to writing to ASD-STE100 covers it in detail. This page covers the rest: the files, the formats, and the structure around the sentences.
How the Tools See Your Text
Translation Memory, TMX, and XLIFF
Professional translation runs through tools that split your text into sentence-sized segments, look each one up in a translation memory, and present the translator with stored matches. Two open formats carry the data between tools, and they do different jobs.
TMX (Translation Memory eXchange) moves the memory itself. It began in the former LISA OSCAR group; the ETSI specification GS LIS 002 V1.4.2 (February 2013) recreates it as version 1.4b, with the stated purpose of describing translation memory data exchanged among tools and translation vendors with little or no loss of critical data. TMX is how a memory leaves one tool or provider and reaches another. Treat it as the asset you own: the accumulated, approved translations of your content.
XLIFF (XML Localization Interchange File Format) moves a job. OASIS describes it as a vocabulary for storing localizable data and carrying it from one step of the localization process to the next, with interoperability between tools. In version 2.1, the OASIS Standard dated 13 February 2018, a unit holds the extracted translatable text and a segment pairs one source with zero or one target, its translation. Version 2.2 appears in the OASIS listing as a Committee Specification dated 13 March 2025.
The round trip is therefore: your files are parsed, translatable text is extracted into XLIFF, segments are matched against the memory and translated, and the result is merged back into your file format with the layout intact. Every problem on the cost list above shows up somewhere on this path. Text that cannot be extracted never reaches the memory. Markup tangled into sentences breaks segmentation. Files that cannot be rebuilt cleanly create layout rework at the end.
Structure Helps
What Structured Source Gives a Translation Workflow
Structured XML source separates content from formatting, which is what extraction wants. It also adds three things that word-processor files cannot give you reliably.
Language and translate flags on the content. In DITA 1.3 the xml:lang attribute specifies the language of the element content, and the translate attribute indicates whether the content of an element should be translated (yes or no). A product name, a code sample, or a part number can be marked once at the source instead of being guessed at by each tool. S1000D data modules carry a language and country code alongside their module code and issue number, so a language version is an identifiable module, not a loose copy.
Topic-sized units. When content is divided into topics or data modules, a revision changes a small number of units. Only those units need to go out for translation, and each language version can be tracked against the source issue it came from. Our guides to single-source publishing and maps, keys, and conref explain how reuse works; the translation point is that a reused topic is translated once, however many manuals include it.
Managed terms and variables. A term defined once and referenced everywhere can be reviewed once and translated once; see managing defined terms. If you are choosing a content system, our CCMS requirements checklist includes the translation hooks to ask for, and the migration guide treats the translation memory and language versions as part of what must move.
What Breaks It
Habits That Make Translation Expensive
- Sentences assembled from fragments
- Reuse that builds a sentence from a reused opening and a variable ending works in English and fails in languages that change word order or inflect the ending. Reuse whole sentences or whole topics, and keep variables to values such as names and numbers.
- Layout done with spaces and manual numbering
- Spaces, tabs, hand-typed step numbers, and manual line breaks end up inside the text sent out. They confuse segmentation and are lost or corrupted on the way back. Let the stylesheet generate numbering and spacing.
- Long, stacked sentences
- Several actions, conditions, and exceptions in one sentence translate poorly and match poorly. One action per sentence gives translators a clear unit and gives the memory something it can reuse.
- Informal and idiomatic wording
- Phrasal verbs, metaphors, and shorthand are the usual sources of translator queries. Controlled language removes most of them before the text leaves the building.
- Fixing the translation instead of the source
- A correction made only in one language leaves the source and the memory wrong, and the error returns at the next revision. Fix the source, then re-send the changed text.
- Untracked language versions
- A translated set that no longer knows which source revision it came from cannot be checked, updated selectively, or defended in an audit.
Graphics deserve their own mention. Numbered callouts with the words in a key, rather than words drawn into the art, mean one drawing serves every language; see technical illustration. Warnings belong in text for the same reason: they can be found, managed, and translated with care.
Before the First Handoff
A Preparation Checklist
- Settle the terminology first: a single approved name for each part, tool, and function, with the terms you never translate (product names, part numbers) marked as such.
- Write the source in a controlled form with a short style guide, so every author produces the same kind of sentence.
- Move text out of graphics. Use index numbers or callout keys in the art and put the words in a table or list that is ordinary text.
- Strip manual formatting from the files you will send, and check that numbering, spacing, and tables come from the stylesheet.
- Mark non-translatable content and set the language on the content, so tools do not guess.
- Do a small pilot in one target language: send a representative chapter, round-trip it through the whole toolchain, and review what came back before committing the full set.
- Record, for every language version, the source revision it was translated from and the date.
Which languages you need, and in what form the instructions must be supplied, depends on where the product is sold and what your contract or the applicable rule says. Our guide to instruction language requirements covers the EU texts; read the text that applies to you.
Where Shakewell Fits
Source Content That Translates Cleanly
Shakewell prepares translation-ready source content: structured, controlled-language manuals that translation tools and translators can work with cleanly. The translation itself stays with the language provider you choose.
FAQ
Questions We Hear
What is a translation memory?
A database of past source sentences paired with their approved translations. When a new sentence matches a stored one exactly or closely, the tool offers the stored translation for the translator to reuse or adjust. It only helps if your source is consistent enough to match.
What is the difference between TMX and XLIFF?
TMX carries the memory, the stored pairs of source and translated sentences, from one tool to another. XLIFF carries a job: the specific text extracted from your files, with its translations as they are produced, from step to step in the process. They solve different problems and are often used together.
Does my manual need to be XML to be translated?
No. Translation tools handle many file types. Structured XML source makes the handoff cleaner, because translatable text is separated from layout and whole topics can be sent and tracked individually, but a well-written, consistently formatted document in another format can still be prepared well.
Is XLIFF 2.2 a standard yet?
OASIS lists XLIFF 2.1 as an OASIS Standard dated 13 February 2018. Version 2.2 appears in the same listing as a Committee Specification dated 13 March 2025. Check the OASIS listing for the current status, and ask your language provider which version their tools expect.
Does using Simplified Technical English guarantee cheaper translation?
It makes the source more consistent and less ambiguous, which helps matching and cuts queries, but it does not remove the other cost drivers. Text in graphics, fragile reuse, and poor revision scoping can still cost you. See our guide to writing to ASD-STE100.
Which languages do I have to translate into?
That depends on the markets you sell into and the rules that apply there, not on this page. See our guide to instruction language requirements, and read the text of the rule or contract that applies to you.
Get In Touch
Fix the Source Before You Send It
Send us a sample chapter and a note on the languages you need. We will tell you what in the source is likely to cause trouble downstream.