Home / Translation practice  / Neural networks and the future of translation
7 Comments
  • Tom Hoar

    Hi Pieter. Thank you for pulling together SMT/NMT ideas from many sources, and thanks for mentioning Slate Desktop. I want to point out that all of the evaluation statistics in the TAUS SlideShare presentation compare the changes in “big data” cloud systems when they migrate from SMT to NMT. They do not reflect a translator’s experience with personalized SMT.

    For that comparison, readers can first read Isabella Massardo’s recent blog about her first EN-IT experience with Slate Desktop. Readers can see evaluation statistics for her Slate engine similar to those in SlideShare. Here’s my re-post: https://slate.rocks/review-who-is-a-translators-new-best-friend/

    Then I compared Isabella’s Slate engine to Google’s NMT using Isabella’s own 2,353 human translated segments. She published my guest article on her blog and I re-blogged it here: https://slate.rocks/practical-mt-evaluation-for-translators/

    In short, the SlideShare evaluation statistics show NMT generically improves 10% to 20% over SMT in “big data” cloud systems. Our evaluation statistics using Isabella’s translations with her Slate engine show 200% to 700% improvements over Google’s new-improved NMT, depending on which scoring system you use.

    Finally, you missed a small but impressive “big data” cloud systems new-comer. Linguee recently launched the DeepL Translator. Translator reviews have declared a significant subjective perception of improved quality, but I can’t find any objective reports. In a few days, I hope to publish the same evaluation statistics results using the DeepL Translator on Isabella’s 2,353 segments. This will be an objective apples-to-apples-to apples comparison. Stay tuned!

    18 September, 2017 at 22.02 Reply
      • Andre Hagestedt

        Hi Pieter,
        DeepL is definitely worth a look. I was using Google NMT and was impressed but DeepL is definitely better, at least for EnglishGerman.

        27 September, 2017 at 04.05 Reply
  • Josephine Bacon

    Human translators are no more replaceable than human authors, in most circumstances. Granted, a robot can translate a weather forecast, but most translations require translator ingenuity and creativity. There is only one reason to use MT: to cheat translators out of money. Many translation agencies claim “we have to use CAT tools, our clients want it”. Their clients wouldn’t know a CAT from a dog! It is their own greed and they will come a cropper and serve them right.

    24 September, 2017 at 19.03 Reply
    • Ioan Prislopeanu

      Jo, there is a girl to my heart ! But – alas – there is a way around it – a narrow path that I have just discovered last week – after 35 years of being convinced that TRANSLATORS and MACHINES DON’T MIX ! Well, Josephine, they MIX and can positively interact ! But you know what? Machines – and the guys up there, the Pupeteers – need US – the Old School bastards to do the trick. Not for translating “avec panache” the recipe of a tropical cocktail or another – but for Will Shakespeare, for Goethe, for Baudelaire – even for Ernst Hemingway, the undisputed bastard of the XX-th. Century…

      26 September, 2017 at 02.31 Reply
  • Ioan Prislopeanu

    Yes, my friends ! There is a way for the Machine to steal our immortal souls from us – and this moment is not so distant as I thought; maybe a couple of decades or so. I was convinced that mechanic translations will never ever exceed the quality and complexity encountered when you ask for the Service Handbook of your new Japanese car….
    Thee IS a way which I saw against my will – and NOW I would very much like to see the guys from Google, Microsoft and the whole bunch – to negotiate with them. Because I am a translator and a soft developer . Tell me – if you know where they are ! ! prislopeanu@yahoo.com

    26 September, 2017 at 02.46 Reply
Post a Comment