Neural Machine Translation – NMT New Media Darling – NMD




  • Greater than 4 minutes, my friend!

    For months, barely a week has passed without a blogger, discussion group, or main-stream media asking if neural machine translation (NMT), with its artificial intelligence (AI) technology, will replace human translators. NMT is the new media darling. I don’t need to list the links, just search for yourself.

    It started with a Google research report in September 2016. Mind you, Google was not the first to experiment with or deploy an NMT system. Both Facebook and Microsoft had production NMT systems running before Google. Google’s NMT report stated,

    In some cases human and GNMT translations are nearly indistinguishable on the relatively simplistic and isolated sentences sampled from Wikipedia and news articles for this experiment(1).

    That sounds really balanced. Nonetheless, reports of Google’s report shakes the market. Why is there so much hype?

    In my opinion, the media needs the ratings. They engaged in what is now popularly called “fake news.” They ignored Google’s 18 pages preceding my quote above, They splashed the headline, “Nearly Indistinguishable From Human Translation”— Google Claims Breakthrough (2). Note the use of the words nearly indistinguishable, the quotes, and the twisted word order. Yet, search the original and the headline’s quoted text just isn’t there. The Slator.com quote is “fake news.”

    Kirti Vashee, a popular translation technology blogger, quickly identified this discrepancy and posted The Google Neural Machine Translation Marketing Deception (3). Regrettably, not even Kirti’s well-balanced and reasoned logic could forestall the flood of speculation and hype that ensued.

    When was the last time I witnessed a new translation technology capture our market’s imagination?

    In 2006. The University of Edinburgh released the Moses Toolkit, an open source software project for statistical machine translation research. The market sprang to life. New technology companies vyed to leverage the free software for profit. Within 3 years, a dozen or more companies popped up with cloud-based services offering “near human quality” translations for human translators to post-edit (4, 5). When SMT was the reigning media darling, an industry pundit predicted this Utopia:

    Looking into the future, I see a thousand MT systems blooming. I see fortune for the translation industry, and new solutions to overcome failed translations. I see a better world due to improved communications among the world’s seven billion citizens. And the reason why I am so optimistic is that the process of data effectiveness is joining hands with the trend towards profit of sharing. – Jaap van der Meer, 03 Nov 2009, Let a Thousand MT Systems Bloom (6)

    Where are those dozen companies now? Some have passed or morphed. Some are still around. Two have opened A-B testing to customers between tried-n-true SMT systems and new NMT systems (7, 8). KantanMT’s founder Tony O’Dowd doesn’t know if NMT is any better. Reading between the lines in his interview, these A-B tests present the appearance that KantanMT are up-to-speed with current trends, capture some attention with the newest media darling, and could justify (not not) a new NMT technology investment before realizing an ROI on SMT.

    The hyperbolic predictions aside, this is only the tip of the iceberg. The parallels run deep between NMT’s emergence in the SMT era, and SMT’s emergence in the now-bygone RbMT era (rules-based machine translation technology). I created this scorecard to help us track the technologies.

    Scorecard

    Key
    1 = possible but weak
    2 = stronger than 1
    3 = stronger than 2
    CAT = responsibility of the CAT or user application, not the MT technology
    N/A = not applicable with today’s technology
    impossible = early NMT technology can not not support this feature
    unknown = early NMT technology might work, but no proven use cases

    Feature

    RbMT

    Cloud
    SMT

    Slate
    Desktop

    NMT

    better for User Generated Content and broad domain material such as patents

    1

    1

    2

    unknown

    documentation and even software

    2

    2

    2

    unknown

    protects tags

    CAT

    impossible

    CAT

    impossible

    translates tags

    CAT

    impossible

    CAT

    impossible

    post-editing and durable changes

    CAT

    impossible

    CAT

    impossible

    on-the-fly translations of short-shelf-life content

    CAT

    2

    CAT

    impossible

    retains corrections to terminology

    CAT

    impossible

    CAT

    impossible

    use the most likely term, not the one you expect

    1

    2

    2

    impossible

    predictable but the sentences may awkward

    2

    1

    1

    impossible

    unpredictable but sentences are more fluid

    1

    1

    2

    3

    fast update, maintain (daily or more frequently)

    2

    N/A

    2

    unknown

    longer update cycles (once or twice a year)

    N/A

    2

    N/A

    unknown

    expensive to license

    2

    1

    1

    N/A

    can be free open source

    1

    N/A

    N/A

    1

    heavy on linguistic resources

    2

    N/A

    1

    impossible

    heavy on processing resources (see hardware below)

    1

    3

    2

    3

    makes more fluid sentences

    1

    2

    2

    2

    can handle bad grammar

    1

    1

    2

    unknown

    significantly better with controlled authoring

    2

    1

    3

    unknown

    choice for minority languages

    1

    1

    2

    unknown

    match for languages like French and Spanish

    1

    1

    2

    1

    performs better for Japanese, German, Russian, Korean

    1

    1

    2

    1

    language pairs out of the box

    30

    < 10,500

    980 +

    unknown

    ready to customize for your domain and preferred terminology

    2

    N/A

    2

    impossible

    hardware resources required

    legacy
    2000 PC
    cloud infrastructure modern notebook PC

    GPUs
    (special
    hardware)

    time to create new engine

    < 1 hour unknown
    (days?)
    6-8 hours

    hours
    or days

    Special thanks to Lori Thicke at Lexworks because I based the scorecard on her 2012 blog, 13 Differences between SMT and RBMT that You Need to Know (9). I updated Lori’s list based on our experience through 2017 and comments Alon Lavie (senior manager of machine translation at Amazon) and Chris Wendt (machine translation program manager at Microsoft) during a recent podcast interview on Globally Speaking Radio (10).


    End notes:

    (1) https://arxiv.org/pdf/1609.08144v1.pdf, bottom of page 18
    (2) https://slator.com/technology/nearly-indistinguishable-from-human-translation-google-claims-breakthrough/
    (3) http://kv-emptypages.blogspot.com/2016/09/the-google-neural-machine-translation.html
    (4) http://www.tauyou.com/Domain_adaptation.pdf
    (5) http://www.prweb.com/releases/2010/01/prweb3447604.htm
    (6) https://www.taus.net/think-tank/articles/translate-articles/let-a-thousand-mt-systems-bloom
    (7) https://translate.tilde.com/neural/en
    (8) https://kantanmtblog.com/2017/02/21/tony-odowd-talks-about-kantanneural/
    (9) http://lexworks.com/translation-blog/13-differences-between-smt-and-rbmt-that-you-need-to-know/
    (10) http://www.globallyspeakingradio.com/podcast/podcast-020-the-challenge-of-neural-mt-part-i

    Tom Hoar

    About Tom Hoar

    Language technology veteran serving translation professionals for 30 years with technologies that augment translators from IBM Selectric II typewriters to dictation and translation software.

    Leave a Reply

    The Open Mic

    Where translators share their stories and where clients find professional translators.

    Find Translators OR Register as a translator