BRENNEN

Industries / Finance and business services / Translation services

Machine translation and editing

Computers translate text from one language into another, and people correct the output, a step called post-editing, until it is fit to use.

7

Milestones achieved, from 1954 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1954

    Russian into English, publicly

    On 7 January 1954, at IBM's headquarters in New York, an IBM 701 computer translated more than 60 Russian sentences, most of them about organic chemistry, into English, using a dictionary of 250 word stems and endings and six grammar rules. Newspapers across the United States reported it, and Léon Dostert of Georgetown University, who had set up the simultaneous interpreting at the Nuremberg trials, was quoted forecasting that in "five, perhaps three, years hence" machine translation in important areas "may well be an accomplished fact".

    LessonA system built for a hand-picked set of sentences says little about how it will handle ordinary text.

    hutchinsweb.me.uk

    Credit

    Léon Dostert and Paul Garvin of Georgetown University, and Cuthbert Hurd and Peter Sheridan of IBM, by the account of the historian John Hutchins

  2. 1966

    A review ends funding

    The Automatic Language Processing Advisory Committee (ALPAC), set up in April 1964 by a joint group of the US Department of Defense, the National Science Foundation and the CIA, reported in November 1966 that "there is no immediate or predictable prospect of useful machine translation". It cited a Georgetown project whose machine output, once corrected by people, took slightly longer and cost more than ordinary human translation, and it recommended spending on aids for human translators and on methods for evaluating translations instead. By Hutchins's account the report ended substantial US funding for machine translation research for about twenty years.

    LessonPromises made from a small demonstration cost the field its support for a generation.

    hutchinsweb.me.uk

    Credit

    ALPAC, chaired by John R. Pierce of Bell Telephone Laboratories

  3. 1977

    Weather forecasts translated daily

    From 25 May 1977 the METEO system, built by the TAUM group at the University of Montreal under a 1975 contract from Canada's Translation Bureau, translated Environment Canada's English weather bulletins into French around the clock; it covered the whole country from June 1978. It passed any sentence it could not analyze to a human translator. By about 1981, of the 18,000 words the Translation Bureau's weather section translated each day, about 11,000 went to METEO, which translated about 80% of them with no human change, by the account of Benoît Thouin.

    LessonMachine translation first paid its way on a narrow, repetitive kind of text with a small vocabulary.

    aclanthology.org

    Credit

    TAUM group, University of Montreal; Canada's Translation Bureau

  4. 1990

    Learning from parallel text

    Researchers at IBM described a system that learned to translate French into English from statistics instead of hand-written grammar rules. They extracted about three million pairs of matching sentences from the record of Canada's parliament, which by law is kept in both French and English, and in a first test estimated some 81 million parameters, numbers giving how likely each word is to translate as another, from 40,000 of those pairs.

    aclanthology.org

    Credit

    Peter F. Brown, John Cocke, Stephen and Vincent Della Pietra, Fredrick Jelinek, John Lafferty, Robert Mercer and Paul Roossin, IBM

  5. 2016

    Neural networks go live

    On 27 September 2016 Google said its Neural Machine Translation system, which handles a whole sentence as one unit instead of translating phrase by phrase, now produced all Chinese-to-English translations in Google Translate, about 18 million a day. It replaced the phrase-based statistical method the service had used since its launch in 2006. In Google's own side-by-side ratings by bilingual people, it cut translation errors by 55% to 85% on several major language pairs.

    research.google

    Credit

    Quoc Le, Mike Schuster and colleagues, Google Brain and Google Translate

  6. 2017

    A standard for post-editing

    ISO 18587:2017 set requirements for "full, human post-editing of machine translation output" and for the competences of post-editors, for use by translation companies, their clients and post-editors. ISO 17100, the general standard for translation services, leaves raw machine output and its post-editing out of its scope, so this standard covers that work.

    webstore.ansi.org

    Credit

    International Organization for Standardization (ISO)

  7. 2025

    Machines join the top group

    At the 2025 Conference on Machine Translation (WMT), whose yearly shared task tests translation systems on the same new texts, organizers evaluated 60 systems and had professional annotators rate whole documents, chosen to be hard, in 16 language pairs. Google's Gemini 2.5 Pro, a general AI model, was in the top group for 14 of the 16 pairs and on par with or better than the human translation in 10 of them; the human translations were in the winning group for only 6 of the 15 pairs that had them. The organizers found that top systems still struggled with non-standard input, specialized terms and gender agreement. In 2026 the human translations, some now made by editing AI output, were back in the top group in 8 of the 12 pairs that had them, and the best system, Wayfinder, built on Gemini, was in the top group for 17 of 23 pairs.

    LessonA system trained to please automatic scores, Shy-hunyuan-MT, ranked first on them in all but one pair yet fell well behind in the human ratings.

    aclanthology.orgaclanthology.org

    Credit

    the WMT25 and WMT26 organizers, led by Tom Kocmi, with researchers from Charles University, Google, Microsoft, the University of Edinburgh and others

Watch for

  1. Open models built for translation

    Google released TranslateGemma, open translation models with 4, 12 and 27 billion parameters built on its Gemma 3 models and trained with rewards from Google's own quality scores, MetricX and AutoMQM. Google says the 12-billion-parameter model beats the larger Gemma 3 27B on its WMT24++ test of 55 languages as scored by MetricX; its own human evaluation in 10 language pairs found gains in most pairs and a step back in Japanese to English. It also trained on nearly 500 more language pairs for which it says it has no confirmed quality figures. In a preliminary comparison for WMT 2026, the organizers' automatic judge (GPT-5.4) scored TranslateGemma 27B lowest of five open models, 67.1 against 82.9 for Qwen3.5; no human ranking has appeared.

    Achieved whenan outside evaluation, such as a WMT shared task with human ratings, ranks TranslateGemma against other systems on the same test.

    arxiv.orgblog.googleaclanthology.org

    Claimed by

    Google, January 2026

  2. A specialist claims the lead

    DeepL says that in blind tests by professional linguists in June 2026 its translations won 89% (66 of 74) of head-to-head test groups across 18 language pairs against seven competitors, among them OpenAI's GPT-5.5, Google's Gemini 3.1 Pro, Anthropic's Claude Opus 4.8 and Google Translate. The page does not say how many raters took part, how they were recruited, or whether every competitor was tested in the same round.

    Achieved whenan independent evaluator publishes a comparison of the same systems with its method and ratings, or DeepL's system is ranked by human raters in a WMT shared task.

    deepl.com

    Claimed by

    DeepL, June 2026

Open

  1. 01

    Rerun the editing-time study

    A repeat of the 2022 DivEMT study with today's AI translation systems. Professional translators post-edit machine translations of the same English documents into Arabic, Dutch, Italian, Turkish, Ukrainian and Vietnamese, and translate some documents from scratch as a baseline, while an editor program logs every keystroke, pause and second. Publish the logs and the seconds per word for each system and language beside the 2022 figures.

    The brief
    Credit · unclaimed
  2. 02

    Rate the untested languages

    Human ratings of Google's open TranslateGemma models in ten languages they were trained on but never evaluated in, such as Wolof, Tok Pisin, Tswana, Tumbuka and Shona, using test sentences that already have professional translations. Two native speakers per language mark each error and its severity.

    The brief
    Credit · unclaimed

Open 01 · the brief

Rerun the editing-time study

Credit · unclaimed

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

What to build

A repeat of the 2022 DivEMT study with today's AI translation systems. Professional translators post-edit machine translations of the same English documents into Arabic, Dutch, Italian, Turkish, Ukrainian and Vietnamese, and translate some documents from scratch as a baseline, while an editor program logs every keystroke, pause and second. Publish the logs and the seconds per word for each system and language beside the 2022 figures.

Why it is still open

DivEMT had 18 professional translators edit output from Google Translate and the mBART-50 model in 2022, before today's AI models, and found that the time saved varied widely, with the least gain for languages unlike English. The public study that followed, QE4PE (2025), logged 42 post-editors but in only two translation directions. Translated, a translation company, publishes a falling "time to edit" trend from more than 2 billion edits on its own platform, but not the data or yearly values.

arxiv.orghuggingface.cotranslated.com

Built with what exists

DivEMT's released documents, logs and analysis code (github.com/gsarti/divemt); GroTE, the open logging editor built for QE4PE; output from current systems such as Gemini, DeepL, Google's open TranslateGemma and Mistral Large 4, whose published weights let anyone rerun the same version later (see the Note "A trillion open parameters").

  1. First step

    Download DivEMT, pick 20 of its documents, translate them with three current systems, and have one Italian translator post-edit them in GroTE to size the differences before recruiting the rest.

  2. Done looks like

    At least 18 professional translators across the six languages and at least three current systems; seconds per word and share of words edited for each system and language, with 95% confidence intervals; the change from DivEMT's 2022 figures; all logs published under an open license.

Open 02 · the brief

Rate the untested languages

Credit · unclaimed

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

What to build

Human ratings of Google's open TranslateGemma models in ten languages they were trained on but never evaluated in, such as Wolof, Tok Pisin, Tswana, Tumbuka and Shona, using test sentences that already have professional translations. Two native speakers per language mark each error and its severity.

Why it is still open

Google says it trained TranslateGemma on nearly 500 language pairs beyond the 55 it measured and has no confirmed evaluation figures for them; its own human evaluation covered 10 pairs. The only outside human ratings found, Alconost's MQM dataset of October 2026, cover 16 pairs out of English with 362 segments in all, mostly major languages plus Hmong and Belarusian.

huggingface.coarxiv.org

Built with what exists

TranslateGemma weights from Hugging Face or Kaggle; FLORES+, the Open Language Data Initiative's professionally translated test sentences in more than 200 languages, gated on Hugging Face under CC BY-SA 4.0; Google's open Anthea rating tool, which the TranslateGemma report used; chrF++ scores from the sacrebleu package for comparison; native speakers recruited through language communities and universities.

  1. First step

    Pick three languages from TranslateGemma's training list that FLORES+ covers and WMT24++ does not, translate 100 FLORES+ sentences into each with the 4-billion-parameter model, and have one native speaker per language rate them.

  2. Done looks like

    10 languages, 200 sentences each, two raters per language using the Error Span Annotation or MQM method; a score per language with a 95% confidence interval and the raters' agreement; the correlation between chrF++ and the human scores for each language; all ratings published.