Scribd Logo
Upload

Welcome to Scribd!

  • 37 views

    MT Impact - Horizon 2020 (And Beyond) : Rudy Tirry

    Uploaded by

    zemike
    Neural machine translation (NMT) is becoming more prevalent, found in tools from Google, Facebook, Amazon and others. NMT uses deep learning with recurrent neural networks rather than the separate models of statistical machine translation (SMT). Studies have found NMT makes 3-5 times fewer errors in areas like word ordering and morphology. However, SMT can still perform better on ambiguity and terminology. Post-editing of NMT output also presents challenges as higher fluency can mislead editors. By 2030, the speaker predicts MT will cover standardized, procedural and basic editorial content while creative content will still require humans. Translators may evolve to become augmented translators that add the human dimension to MT through various levels of post-editing

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    MT Impact - Horizon 2020 (And Beyond) : Rudy Tirry

    Uploaded by

    zemike
    37 views23 pages
    Neural machine translation (NMT) is becoming more prevalent, found in tools from Google, Facebook, Amazon and others. NMT uses deep learning with recurrent neural networks rather than the separate models of statistical machine translation (SMT). Studies have found NMT makes 3-5 times fewer errors in areas like word ordering and morphology. However, SMT can still perform better on ambiguity and terminology. Post-editing of NMT output also presents challenges as higher fluency can mislead editors. By 2030, the speaker predicts MT will cover standardized, procedural and basic editorial content while creative content will still require humans. Translators may evolve to become augmented translators that add the human dimension to MT through various levels of post-editing

    Original Title

    2_Rudy_Tirry

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Neural machine translation (NMT) is becoming more prevalent, found in tools from Google, Facebook, Amazon and others. NMT uses deep learning with recurrent neural networks rather than the separate models of statistical machine translation (SMT). Studies have found NMT makes 3-5 times fewer errors in areas like word ordering and morphology. However, SMT can still perform better on ambiguity and terminology. Post-editing of NMT output also presents challenges as higher fluency can mislead editors. By 2030, the speaker predicts MT will cover standardized, procedural and basic editorial content while creative content will still require humans. Translators may evolve to become augmented translators that add the human dimension to MT through various levels of post-editing

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    37 views23 pages

    MT Impact - Horizon 2020 (And Beyond) : Rudy Tirry

    Uploaded by

    zemike
    Neural machine translation (NMT) is becoming more prevalent, found in tools from Google, Facebook, Amazon and others. NMT uses deep learning with recurrent neural networks rather than the separate models of statistical machine translation (SMT). Studies have found NMT makes 3-5 times fewer errors in areas like word ordering and morphology. However, SMT can still perform better on ambiguity and terminology. Post-editing of NMT output also presents challenges as higher fluency can mislead editors. By 2030, the speaker predicts MT will cover standardized, procedural and basic editorial content while creative content will still require humans. Translators may evolve to become augmented translators that add the human dimension to MT through various levels of post-editing

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 23

    MT Impact – Horizon 2020 (and beyond)

    Rudy Tirry
    President EUATC
    Country Manager LIONBRIDGE BELGIUM
    Agenda

    • MT usage today
    • NMT vs SMT
    • Post-editing
    • MT usage tomorrow
    • Machine vs Human
    MT usage
    2017 Language Industry Survey - Trends

    35%

    30%
    30%

    25% 23%

    20%

    15% 13%
    11%
    10%

    4% 3%
    5% 3% 3% 3%
    2% 2%
    0%
    2018 – the year of MT

    No MT usage 2016 2017 2018 (prel.)


    LSP’s 80% 57% 36%
    Freelancers 69% 67% 48%
    Source: European Language Industry Survey
    No MT
    100%
    80%
    60%
    40%
    20%
    0%
    2016 2017 2018 (prel)
    Companies Individuals
    What about Neural ?
    Neural MT – why the hype ?
    Where do you find neural MT today?

    • Google Translate
    • Skype Translate
    • Microsoft Translator Live
    • Facebook
    • Amazon
    …
    • DeepL
    • Omniscien
    • SDL
    • Systran
    • KantanMT
    • Tilde
    …

    The new Wild Wild West in machine translation


    SMT vs NMT : main differences

    Statistical Machine Translation (SMT) Neural Machine Translation (NMT)


    • Phrase-based • Uses (recurrent) neural networks (« deep »
    • Separate language model, translation model and NMT uses several layers of neural networks)
    reordering model • Sentence-based
    • Fast training • One single sequencing model – simpler than
    SMT approach
    • ‘Predicts’ next word
    • Restricted vocabulary (max. 50,000)
    • More time needed for training
    • No easy solution for terminology
    • Less tolerant for low quality source
    • More pre- and post-processing required
    Is it really that good ?
    Findings Tilde (www.tilde.com/about/news/316)
    Findings DFKI/QT21 Project
    Phenomenon Occurrences Percentage correct
    NMT Moses
    Formal address 138 90% 86%
    Genitive 114 92% 68%
    Modal construction 290 94% 75%
    Negation 101 93% 86%
    Passive voice 109 83% 40%
    Predicate adjective 122 81% 75%
    Prepositional phrase 104 81% 75%
    Terminology 330 35% 68%
    Tagging 145 83% 100%
    Sum/average 1453 89% 73%
    Findings DFKI/QT21 Project

    Source:
    MultiLingual Jan ’18, John Tinsley
    So, is it better ?

    • NMT makes 3 to 5 times less errors in


    – Word ordering
    – Morphology
    – Syntax
    – Agreements
    Source: Tilde (EN>ET)
    This leads to more fluent translations, mainly on ‘difficult’ languages

    • BUT
    – Older techniques (RBMT, trained SMT) can perform better on ambiguity
    (source:PBML, Aljoscha Burchardt et al., June 2017), terminology and tagging
    – Still quite some ‘dangerous’ errors, such as negation (although better than SMT)
    – Traditional automated evaluation methods (BLUE score) do not always agree with
    human evaluation results
    – Uneven results, depending on language pair
    Let’s talk Post-editing
    Post-editing levels

    • Full MTPE
    – No distinction with full human translation
    • Light MTPE
    – Correct understanding
    – No effort on stylistic aspects
    – Varying practices regarding linguistic accuracy
    – Run automated QA rules (ex. check for missing negation)
    • Focused MTPE
    – Specific rules
    • Specific – highly visible – parts of the content
    • Check important elements like numbers, names, etc.
    • …
    Traps

    • Higher fluency misleads post-editor


    • Terminology
    • Tag order issues
    • « Target first » approach not ideal
    So where does that lead us ?
    MT penetration – a personal view
    100%
    Basic procedural
    2030 content

    Complex procedural Basic editorial


    content content
    2030

    2030 2030 Complex


    editorial
    2018 2020 content
    2018 (marketing
    brochures)
    2020 2020
    2030
    Standard-
    ized 2018 2020
    content Creative
    User-generated Controlled 2018 2018 content
    content language 2018 (ads)
    Future role(s) of translator

    « Machine translation will replace only those translators


    that translate like a machine »

    « The machine will take care of the keystrokes.


    The translator will add the human dimension - the cherry on the cake.»

    • Non-MT content (<20% - same as non-CAT content)


    – Transcreator / Copy-editor
    • MT content
    – Full post-editing
    – Light post-editing
    – Focused post-editing
    Translator = Post-editor = [Augmented Translator] ?

    • Do we need the same profile ?


    • Editorial translation
    • Full post editing
    • Light post editing (cf software testing)

    • Do we need the same training ?


    • Creative writing
    • Pattern recognition (search for typical MT errors)
    • Eye for detail and critical sense
    • General (world) knowledge: disambiguation, logical errors
    Workable MT is here. Are we ready to work with it ?

    Will translation buyer expectations become realistic ?


    Will translators embrace technology ?
    Will translation companies find a workable business model ?
    Will universities adapt training programmes to prepare future generations ?
    Will translation tool providers be able to integrate and standardise ?

    Come and see in 2, or rather 12 years !


    Q&A

    [email protected]

    You might also like