the transla*on studies guide to disrup*on Dorothy Kenny Dublin City University Ireland

Transcription

1 the transla*on studies guide to disrup*on Dorothy Kenny Dublin City University Ireland

2

3 disrup've technologies /dɪsˈrʌptɪv tɛkˈnɒlədʒiz/ adj-noun technologies that result in worse product performance, as least in the near-term Generally, disrup*ve technologies underperform established products in mainstream markets. But they have other features that a few fringe (and generally new) customers value. Products based on disrup*ve technologies are typically cheaper, simpler, smaller, and, frequently, more convenient to use. (Christensen 1997:xv)

4 digital disrup'on /ˈdɪdʒɪt(ə)l dɪsˈrʌpʃn/ noun-noun a transforma*on that is caused by emerging digital technologies and business models These innova*ve new technologies and models can impact the value of exis*ng products and services offered in the industry. This is why the term disrup*on is used, as the emergence of these new digital products/services/ businesses disrupts the current market and causes the need for reevalua*on. (h\ps://blog.oxfordcollegeofmarke*ng.com/2016/02/22/what-is-digital-disrup*on/)

5 Digital disruptors: build be4er product experiences that create stronger customer rela*onships bringing it all to market faster h\p:// Digital-Disrup*on-McQuivey.pdf 2013

6 low-end disrup*on first low-margin market then address over-served customers with a lower-cost business model new-market disrup*on compete against non-consump*on e.g. free on-line MT

7 Totally Fic'onal Comparison Between Transla'on Types 10 9 Arbitrary Quality Score HT/CAT RBMT SMT NMT

8 What performance metric should appear on the y-axis? 10 Totally Fic'onal Comparison Between Transla'on Types 9 Arbitrary Quality Score fringe market performance demanded in premium market HT/CAT RBMT SMT NMT

9 AEMs BLEU has always been the primary metric used to rank par<cipa<ng systems (IWSLT, Ben*vogli et al. 2016a) SMT beaer than PBMT by +5.3 BLEU points on English- German (PBSMT vs NMT, Ben*vogli et al. 2016b) Error Analysis morphology, lexical, and word order errors (PBSMT vs NMT, Ben*vogli et al. 2016b)

10 Sie werden nicht disappointed, wenn Sie unser Sor*ment. ver*cal integra*on

11 the acceptance of machine-translated texts, especially post-edited ones, has a lot to do with user expecta<ons and varied needs for quality (or an increased pain threshold for poor quality) Austermühl 2013

12 the translator s dilemma upward vs downward mobility

13 complements vs subs*tutes Hal Varian (Google) seek to be an indispensable complement to something that s gepng cheap and plen<ful (in Brynjolfsson and McAfee 2014:200) Post-edi*ng MT is among fastest growing segments of language industry (CSA 2016)

14 Post-Edi*ng Enjoyment? Moorkens and O Brien 2017 an edit-intensive, mechanical task that requires correc*on of basic linguis*c errors over and over again it s mechanics, and if it s mechanic, there must be a way it could be done by a machine Informant F in Moorkens and O Brien 2017

15 the solu*on? augmented transla*on with translator at the centre using adap*ve (neural) machine transla*on (Lommel/CSA 2017)

16 transla*on pedagogy Frank Austermühl Future (and not-so-future) trends in the teaching of transla*on technology. Tradumà<ca, 11: Stephen Doherty and Dorothy Kenny The Design and Evalua*on of a Sta*s*cal Machine Transla*on Syllabus for Transla*on Students. The Interpreter And Translator Trainer, 8(2): Dorothy Kenny and Stephen Doherty Sta*s*cal Machine Transla*on in the Transla*on Curriculum: overcoming obstacles and empowering translators. The Interpreter And Translator Trainer, 8(2): Anthony Pym Transla*on Skill-Sets in a Machine-Transla*on Age, Meta 583: Syllabus for NMT? Integra*on of adap*ve MT? Diversifica*on? Today J Takeda; INSTB (Buysschaert et al.); Moorkens; Kageura

17 machine transla*on, meaning and transla*on theory

18 Rule-based MT (e.g. Eurotra) lexical transfer rules {gb_lu=need} => {ir_lu=teastaigh,ir_pformarg2=ó}. (complex) structural transfer rules S: {cat=sentence} [ V: {gb_lu=need}, ARG1: {role=arg1}, ARG2: {role=arg2}, REST: *{} ] => S: [ V: {ir_lu=teastaigh,ir_argformarg2=ó}, ARG2: {role=arg1}, ARG1:{role=arg2}, REST ]. euroversal treatment of e.g. tense, aspect, diathesis

19 reac*on from transla*on studies (to RBMT; 1980s) MT a chimera Linguis*c theories of transla*on myopic naïve universalist discredited by their associa*on with MT (see Kenny 2001)

20 1990s: Transla*on Memory Efforts now coalesced around mapping and materializing the translator s cogni<ve decisionmaking That past transla<on decisions informed present ones, and should thus be consulted, updated or repeated necessitated the detailed capture of transla<on prac<ce s physical traces translated texts (and the assump<on of fidelity that makes this equal to that ). (Mitchell 2010:243)

21 Translation Memory Tools

22 1990s & 2000s: Sta*s*cal Machine Transla*on

23 phrase tables h\p://

24 SMT doesn t need to deal with meaning as that is a problem that has already been solved transla*on solu*ons already available in parallel corpora stored on, e.g., on hard disks based on what you have, not what you know based on n-grams no need for linguis*callymo*vated units all instan*a*on and no rule

25 effect on theory? Maybe professional translators work like Google Translate. David Bellos 2011

26 The idea of genera<ng target sentences by transla<ng words and phrases from the source sentence in a random order using a model containing many nonsensical transla<ons may not seem plausible. In fact, the methods used are not intended (in our opinion, at least) to be either linguis<cally or cogni<vely plausible. (Hearne and Way 2011:206)

27 SMT to NMT Sept 2016 Early experiments in SMT at IBM IWSLT shared task evalua*ons start WMT shared task evaluations start Google Translate moves fully to SMT SMT is state of the art, outperforming all other approaches to MT Neural MT outperforms SMT in shared tasks Google Translate starts moving to Neural MT

28 Neural Machine Transla*on NMT is appealing since it is conceptually simple It reads through the given source words one by one un<l the end, and then, starts emipng one target word at a <me un<l a special end-ofsentence symbol is produced. (Luong and Manning 2015)

29 NMT analogy with neural networks suggests an approach to meaning that is rela*onal, associa*ve and distributed similarity of meaning (intra- and interlingual) supposedly represented by proximity in mul*dimensional space

30 Neural Machine Transla*on h\ps://sites.google.com/site/acl16nmt/

31 We kind of hope that in these hidden states, there s enough magic to capture the meaning of the input sentence to then produce every single word in the output sentence. (Koehn 2016)

32 Neural Machine Transla*on. NMT represents a further step in the evolu<on from rule-based approaches that explicitly manipulate knowledge, to the sta<s<cal/data-driven framework, s<ll comprehensible in its inner workings, to a sub-symbolic framework in which the transla<on process is totally opaque to the analysis. (Ben*vogli et al. 2016b)

33 Opacity opens the door to error and misuse You can t control what you don t understand (Domingos 2017:xvi) Opacity as (1) inten*onal corporate or state secrecy (2) technical illiteracy (3) a property that arises from the characteris*cs of machine learning algorithms and the scale required to apply them usefully (Burrell 2016)

34 NMT Material Requirements huge training corpora long training *mes (days, weeks, months) dedicated expensive hardware (such as GPUs, originally used as graphic processing units) significant energy consump*on e.g. Britz et al. (2017): took 250,000 GPU hours using Nvidia Tesla K40m and Tesla K80 GPUs. Joss Moorkens (pers. comm.) es*mates this would cost around 12,750 (17c per kwh) in energy alone, without factoring in cooling, memory costs, etc.

35 Customizable NMT

36 NMT and the materiality of language NMT may involve processing text as single characters or sub-word character sequences (byte pair encoding) less and less to do with meaning and more to do with graphic substance (Ca ord 1965) can lead to crea*ve transla*ons (Mitchell 2010)

37 Conclusions SMT was a disrup*ve technology NMT probably a sustaining technology, but a challenge nonetheless Transla*on pedagogy and theory trying to keep up

38 Thank you!

39 References Austermühl, Frank Future (and not-so-future) trends in the teaching of transla*on technology. Tradumà<ca, 11: Bellos, David Is That a Fish in Your Ear? London: Penguin. Ben*vogli, Luisa, Marcello Federico, Sebas*an Stüker, Mauro Ce\olo, Jan Niehues. 2016a. The IWSLT Evalua*on Campaign: Challenges, Achievements, Future Direc*ons. Proceedings of the LREC 2016 Workshop Transla*on Evalua*on From Fragmented Tools and Data Sets to an Integrated Ecosystem, Georg Rehm, Aljoscha Burchardt et al. (eds.), pp Ben*vogli, Luisa, Arianna Bisazza, Mauro Ce\olo and Marcello Federico Neural versus Phrase-Based Machine Transla*on Quality: a Case Study. In EMNLP arxiv: v1 [cs.cl] 16 Aug Britz, Denny, Anna Goldie, Minh-Thang Luong and Quoc Le Massive Explora*on of Neural Machine Transla*on Architectures. h\ps://arxiv.org/ pdf/ pdf [cs.cl] 21 March Brynjolfsson, Erik and Andrew McAfee The Second Machine Age. New York: W. W. Norton & Company Burrell, Jenna How the machine thinks : Understanding opacity in machine learning algorithms. Big Data & Society, January June 2016: Ca ord, J.C A Linguis<c Theory Of Transla<on. Oxford: Oxford University Press. Christensen, C. M The innovator s dilemma: when new technologies cause great firms to fail. Boston, MA: Harvard Business School. Common Sense Advisory Global Market Research Firm Common Sense Advisory Finds Post-edited Machine Transla*on (PEMT) Among Fastestgrowing Segments of the Language Industry. CSA Blog. h\p:// Conten\ype=Ar*cleDet&tabID=64&moduleId=392&Aid=36546&PR=PR 08 February Domingos, Pedro The Master Algorithm. London: Penguin. Hearne, Mary and Andy Way Sta*s*cal Machine Transla*on: A Guide for Linguists and Translators. Language and Linguis<cs Compass 5: doi/ /j x x/pdf. Kenny, Dorothy Lexis and Crea<vity in Transla<on. A Corpus-based Study. Manchester: St. Jerome. Koehn, Philipp The State of Neural Machine Transla<on (NMT). Omniscien Technologies Webinar. 24 November h\ps://omniscien.com/ more/resources/webinar/#-the-state-of-neural-machine-transla*on--nmt----technology-webinar-series Luong, Minh-Thang and Christopher D. Manning Stanford Neural Machine Transla*on Systems for Spoken Language Domains. In IWSLT15. Mitchell, Chris*ne Transla*on and Materiality. The Paradox of Visible Transla*on. Transla<ng Media 30(1): Moorkens, Joss and Sharon O Brien Assessing User Interface Needs of Post-Editors of Machine Transla*on. In Dorothy Kenny (ed.) Human Issues in Transla<on Technology. London and New York: Routledge, Pym, Anthony Transla*on Skill-Sets in a Machine-Transla*on Age, Meta 583: