BRENNEN

Industries / Finance and business services / Translation services

Document translation

Translators turn written material such as manuals, reports, websites and the text inside software from one language into another, usually working in software that offers them earlier translations to reuse.

7

Milestones achieved, from 1978 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1978

    Store every past translation

    Peter Arthern, head of the English translation division at the Council of the European Communities, proposed at a London seminar that an organization keep every text it produces in one computer store together with its translations. The computer would compare each new text, sentence by sentence, with the store and print the nearest earlier translation, so translators would fill in only what was new. This is the idea behind translation memory, the database of earlier sentence pairs that most translation software now checks first.

    aclanthology.org

    Credit

    Peter J. Arthern, Council of the European Communities

  2. 1992

    Memory software reaches freelancers

    Trados, a Stuttgart firm founded in 1984 by Jochen Hummel and Iko Knyphausen, released the first edition of its Translator's Workbench in 1992 and a Windows version in 1994. The European Commission bought 200 licenses of its first product, the terminology database MultiTerm, and more later; in September 1997 Microsoft adopted Trados for its own localization work and bought a 20% stake. Clients began to ask for a "Trados discount" on text the memory already held.

    LessonTwo earlier memory tools, the Translation Support System of ALPS in Salt Lake City and INK's TextTools in the Netherlands, were dropped in the 1980s because the computers of the day were too limited for them.

    aclanthology.org

    Credit

    Jochen Hummel and Iko Knyphausen, Trados

  3. 2008

    A shared file for translation

    OASIS, a standards consortium, approved XLIFF 1.2 on 1 February 2008: an XML file format that carries the text to be translated, its translations and notes between the tools of different vendors, so a job can move from one company's software to another's. XLIFF 2.0 replaced it on 5 August 2014.

    docs.oasis-open.org

    Credit

    OASIS XLIFF Technical Committee; XLIFF 2.0 editors Tom Comerford, David Filip, Rodolfo M. Raya and Yves Savourel

  4. 2008

    Users translate Facebook

    Facebook let users suggest translations of its pages while they browsed and vote on each other's suggestions. By Facebook's count, nearly 1,500 Spanish speakers translated the site in less than four weeks, and visitors from Spanish-speaking countries began seeing it in Spanish on 11 February 2008. IEEE Spectrum reported in 2011 that Facebook had 252 volunteer Brazilian Portuguese translators who had submitted more than 150,000 translations, with outside consultants reviewing the voted results.

    about.fb.comspectrum.ieee.org

    Credit

    Facebook

  5. 2011

    The standards body folds

    The Localization Industry Standards Association (LISA), a trade body that published file formats shared by translation tools, was insolvent and closed on 28 February 2011, its managing director Michael Anobile wrote. In July 2011 ETSI, the European telecommunications standards institute, set up a group to carry on LISA's work, naming TMX, the format for moving translation memories between tools, and SRX, the format for rules that split text into sentences.

    LessonFormats that every tool depends on need a home that can outlast one trade body's funding.

    cetra.cometsi.org

    Credit

    LISA; ETSI

  6. 2024

    Translators report lost work

    In a survey run through January 2024, the Society of Authors, the UK writers' union, found that about a third of translators who answered had already lost work to generative AI. The survey drew 787 responses from its members.

    thebookseller.cominternationalauthors.org

    Credit

    Society of Authors

  7. 2025

    The market stops growing

    Nimdzi Insights, an analyst firm, estimates the language services market at USD 72.6 billion in 2025 and expects growth of less than 1% a year to 2030. By its count the 100 largest providers grew 1.1% combined, the providers ranked 51 to 100 shrank 4.3%, and 22 of the 100 lost revenue; the largest, TransPerfect, had USD 1.32 billion in 2025 revenue. The new revenue came mostly from AI and data services, such as preparing data to train AI models, and from interpreting, which made up 33.6% of the 100 largest providers' revenue. Nimdzi writes that price pressure and "the mind-numbing nature of post-editing AI output have driven many linguists to leave the profession". Slator, another analyst firm, counts a narrower market of USD 30.85 billion in 2025 and expects it to grow 2.65% a year to 2031.

    nimdzi.comslator.com

    Credit

    Nimdzi Insights (Marjolein Groot Nibbelink and Laszlo K. Varga); Slator

Watch for

  1. Canada's in-house AI translator

    Public Services and Procurement Canada (PSPC), the department that houses the federal Translation Bureau, announced GCtranslate on 25 September 2025: an AI tool meant to handle routine business documents so translators can focus on complex texts, then being tested in six departments and agencies. A PSPC spokesperson later called it one of the department's most-used tools. The bureau had reported a 17% drop in demand for its services in 2023–24, and its union says the bureau's five-year plan would cut about a quarter of its roughly 1,300 staff through attrition.

    Achieved whenan outside party, such as the Commissioner of Official Languages, publishes measured error rates for GCtranslate on a sample of real government documents.

    capitalcurrent.cathelogic.co

    Claimed by

    Public Services and Procurement Canada, September 2025

Open

  1. 01

    Score AI on app strings

    A public test set and scorer for AI translation of software text. Take the string files (gettext.po files) of open-source programs that already ship human translations, remove the translations, and have several AI models translate them. A script then checks every string for broken placeholders such as %s or {name}, missing plural forms (Polish, Russian and Arabic need more than English's two), damaged markup and lines over the length limit, and native-speaking translators grade the meaning of a sample without knowing which model wrote it.

    The brief
    Credit · unclaimed
  2. 02

    Measure layout damage in translation

    A public test of the "translate this document" features of common tools on 100 openly licensed Word and PowerPoint files full of tables, footnotes, text boxes, charts and headers. For each output, count the damage page by page: footnotes lost or moved, tables broken, text spilling out of boxes, headers or image text left untranslated, and numbers changed.

    The brief
    Credit · unclaimed

Open 01 · the brief

Score AI on app strings

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set and scorer for AI translation of software text. Take the string files (gettext.po files) of open-source programs that already ship human translations, remove the translations, and have several AI models translate them. A script then checks every string for broken placeholders such as %s or {name}, missing plural forms (Polish, Russian and Arabic need more than English's two), damaged markup and lines over the length limit, and native-speaking translators grade the meaning of a sample without knowing which model wrote it.

Why it is still open

The tests found are small or closed. The software company Localazy had language professionals score seven AI models on 10 strings into Polish and Croatian in March 2025 and found that "variables appeared to be problematic to some extent for nearly each LLM", and Claude struggled with plural forms. An AMTA 2026 paper by Mao Tian and Na Wu tested eight models as checkers of game text, built from strings of three commercial games from unnamed studios. This research found no open benchmark built on real.po catalogs.

localazy.comaclanthology.org

Built with what exists

.po files downloaded from GNOME's translation site (l10n.gnome.org) and other open-source projects; the Translate Toolkit's pofilter, which already checks printf placeholders, variables, XML tags, plural counts and length; AI models such as Claude Fable 5.2 (see the Note "Three working devices") and the open-weight Mistral Large 4, which can run on a team's own servers (see the Note "A trillion open parameters"); Python.

  1. First step

    Download the Polish.po files of three GNOME programs, strip the translations, translate them with two models, and count pofilter's placeholder and plural errors against the shipped human translations.

  2. Done looks like

    5,000 strings from 10 projects in 6 target languages, each translated by at least 4 models; for each model and language, the share of strings with a broken placeholder, a wrong number of plural forms or text over the length limit; meaning errors on 500 strings per language graded blind by two native translators, with their agreement reported; strings, outputs and scores published under the projects' own licenses.

Open 02 · the brief

Measure layout damage in translation

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of the "translate this document" features of common tools on 100 openly licensed Word and PowerPoint files full of tables, footnotes, text boxes, charts and headers. For each output, count the damage page by page: footnotes lost or moved, tables broken, text spilling out of boxes, headers or image text left untranslated, and numbers changed.

Why it is still open

Research so far covers PDFs and tagged sentences. ForMaT, a set of 3,956 PDFs in 15 language pairs presented at EAMT in June 2026, found that current systems often lose "the link between text and its visual context"; BabelDOC, an open-source PDF translator described in May 2026, was scored against other tools on a 200-page set. A 2023 MT Summit study by Raj Dabre and colleagues found that AI models placed formatting tags well inside sentences but risked "losing, hallucinating and corrupting tags". LayoutTranslateBench, a GitHub project begun in May 2026, scores 59 documents but so far only text translators given the layout. This research found no public count of layout damage in Word and PowerPoint files run through the document features people use.

aclanthology.orgarxiv.orgaclanthology.orggithub.com

Built with what exists

Document upload in DeepL, Google Translate and Microsoft Word's Translate feature, and an AI model that accepts files; python-docx and python-pptx to compare document structure before and after; LibreOffice in headless mode to render pages as images; a vision-language model to flag overflowing text for a person to confirm; public-domain US federal agency documents as test files.

  1. First step

    Pick 10 US federal agency Word files with tables and footnotes, run them through two tools into Spanish, and log every defect by hand to fix the list of defect types.

  2. Done looks like

    100 documents, 4 tools and 3 target languages, with defects per 10 pages reported by type and tool; the automatic counts checked by hand on 20 documents, agreeing at least 90% of the time; files, outputs and scores published.