(Or: How soon will translation apps make us all redundant?)
An applied linguist collecting data
In a book published on the first day of the new millennium, futurologist Ray Kurzweil (2000) predicted that spoken language translation would be common by the year 2019 and that computers would reach human levels of translation by 2029. It would seem that we are well on track. Maybe even a few years ahead.
Google Translate, for example, was launched in 2006, and now supports over 100 languages, although, since it draws on an enormous corpus of already translated texts, it is more reliable with ‘big’ languages, such as English, Spanish, and French.
A fair amount of scorn has been heaped on Google Translate but, in the languages I mostly deal with, I have always found it fairly accurate. Here for example is the first paragraph of this blog translated into Spanish and then back again:
En un libro publicado el primer día del nuevo milenio, el futurólogo Ray Kurzweil (2000) predijo que la traducción hablada sería común para el año 2019 y que las computadoras llegarían a niveles humanos de traducción para 2029. Parecería que estamos bien en el camino. Tal vez incluso unos años por delante.
In a book published on the first day of the new millennium, futurist Ray Kurzweil (2000) predicted that the spoken translation would be common for 2019 and that computers would reach human translation levels by 2029. It would seem we are well on the road. Maybe even a few years ahead.
Initially text-to-text based, Google Translate has more recently been experimenting with a ‘conversation mode’, i.e. speech-to-speech translation, the ultimate goal of machine translation – and memorably foreshadowed by the ‘Babel fish’ of Douglas Adams (1995): ‘If you stick a Babel fish in your ear you can instantly understand anything said to you in any form of language.’
The boffins at Microsoft and Skype have been beavering away towards the same goal: to produce a reliable speech-to-speech translator in a wide range of languages. For a road test of Skype’s English-Mandarin product, see here: https://qz.com/526019/how-good-is-skypes-instant-translation-we-put-it-to-the-chinese-stress-test/
The verdict (two years ago) was less than impressive, but the reviewers concede that Skype Translator will ‘only get better’ – a view echoed by The Economist last month:
Translation software will go on getting better. Not only will engineers keep tweaking their statistical models and neural networks, but users themselves will make improvements to their own systems.
Mention of statistical models and neural networks reminds us that machine translation has evolved through at least three key stages since its inception in the 1960s. First was the ‘slot-filling stage’, whereby individual words were translated and plugged into syntactic structures selected from a built-in grammar. This less-than-successful model was eventually supplanted by statistical models, dependent on enormous data-bases of already translated text, which were rapidly scanned using algorithms that sought out the best possible phrase-length fit for a given word. Statistical Machine Translation (SMT) was the model on which Google Translate was initially based. It has been successful up to a point, but – since it handles only short sequences of words at a time – it tends to be less reliable dealing with longer stretches of text.
An early translation app
More recently still, so-called neural machine translation (NMT), modelled on neural networks, attempts to replicate mental processes of text interpretation and production. As Microsoft describes it, NMT works in two stages:
- A first stage models the word that needs to be translated based on the context of this word (and its possible translations) within the full sentence, whether the sentence is 5 words or 20 words long.
- A second stage then translates this word model (not the word itself but the model the neural network has built of it), within the context of the sentence, into the other language.
Because NMT systems learn on the job and have a predictive capability, they are able to make good guesses as to when to start translating and when to wait for more input, and thereby reduce the lag between input and output. Combined with developments in voice recognition software, NMT provides the nearest thing so far to simultaneous speech-to-speech translation, and has generated a flurry of new apps. See for example:
https://www.youtube.com/watch?v=ZAHfevDUMK4
https://www.cnet.com/news/this-in-ear-translator-will-hit-eardrums-next-month-lingmo/
One caveat to balance against the often rapturous claims made by their promoters is that many of these apps are trialled using fairly routine exchanges of the type Do you know a good sushi restaurant near here? They need to be able to prove their worth in a much wider variety of registers, both formal and informal. Nevertheless, Kurzweil’s prediction that speech-to-speech translation will be commonplace in two years’ time looks closer to being realized. What, I wonder, will it do to the language teaching industry?
As a footnote, is it not significant that developments in machine translation seem to have mirrored developments in language acquisition theory in general, and specifically the shift from a focus primarily on syntactic processing to one that favours exemplar-based learning? Viewed from this perspective, acquisition – and translation – is less the activation of a pre-specified grammar, and more the cumulative effect of exposure to masses of data and the probabilistic abstraction of the regularities therein. Perhaps the reason that a child – or a good translator – never produces sentences of the order of Is the man who tall is in the room? or John seems to the men to like each other (Chomsky 2007) is not because these sentences violate structure-dependent rules, but because the child/translator has never encountered instances of anything like them.
References
Adams, D. (1995) The hitchhiker’s guide to the galaxy. London: Heinemann.
Chomsky, N. (2007) On Language. New York: The New Press.
Kurzweil, R. (2000) The Age of Spiritual Machines: When Computers Exceed Human Intelligence. Penguin.
An A-Z of ELT 
Hey Scott,
Strangely enough I was reading a piece about a human who perhaps didn’t do that good a job at translation (https://koreaexpose.com/deborah-smith-translation-han-kang-novel-vegetarian/) just before this. I know it’s not the same as speech to text translation, but I think it highlights that some things about translation are really hard (especially with high-context languages) and possibly also things that are hard for computers to be good at.
The other thing to consider is the reasons people learn languages. It’s often because we really like the language and like the process of learning it too. Even if we did have babelfish, I think people would still be interested in learning languages. Hopefully ELT will be around for a long time.
Thanks, Tim… yes, I guess an approximate analogy is the ‘self-driving car’, mentioned by Jack below. There will be people who will carry on driving because, well, they just like it. And others (like me) who don’t like driving but will prefer it to entrusting one’s life to an automaton. But there will be a LOT of people who will happily embrace driverless cars (even literally!) consigning driver-driven cars (is that a retronym?) to the expensive, niche end of the market. Will language classrooms go the same way?
Very interesting, as always, Scott! Funnily enough, I was marvelling at the improvements made by Google Translate recently when asked to do an informal bit of Italian-English translation by a friend who is an artist. Although not perfect, I would guess that it got 90% right first time. I wonder what will happen to the likes of Trados in future.
On a similar note, have you seen the new translation gadget that is seemingly all over the Internet, called Ili (see here for a rather jarring promo video https://youtu.be/VpFhdjkfWq8). It certainly seems like it could develop into something big in future when it offers more languages.
PS Do translators have a place in the EFL classroom? I think no but curious to hear your thoughts!
Regards,
Steve
Thanks, Steve – and thanks for the link. I note that the Ili gadget is limited only to travel language so in a sense it is really an automated phrasebook with voice recognition.
As for your last question, I have blogged about the role of translation in the classroom here: https://scottthornbury.wordpress.com/2010/10/15/g-is-for-grammar-translation/
I wonder why those working on machine translation haven’t embraced Larsen-Freeman’s explanations of language acquisition as enthusiastically as you have. No doubt this explains the delay.
On the contrary, Geoff, the view that complex systems emerge from the ‘coordinated activity of many different simple mechanisms’ seems to underpin all the work now being done in artificial intelligence and, by extension, machine learning and machine translation. See for example this long piece about ‘Google Brain’ from The New York Times: https://www.nytimes.com/2016/12/14/magazine/the-great-ai-awakening.html?_r=0
And the delay you refer to – do you mean the delay in posting the blog? That’s because I’m now in the US and forgot to change my settings. Unlike that of machines, my own intelligence has its limits!
“Complex systems emerge from the ‘coordinated activity of many different simple mechanisms’”. An easy step from this briliant summary to finally cracking machine translation and understanding language learning, which was the delay I referred to.
‘you’re going to leave a lot of translators on the freeway’. That’s enough. Fascinating and like the self-driving car, coming soon. https://www.abctales.com/story/celticman/go
I wonder if coordinating simple mechanisms is simple. (Sorry to hijack, Scott).
That’s an odd quote from the Economist. Do they really say “go on getting better” or is that machine translationese? And I can’t imagine what they mean by users making improvements to their own systems. Nevertheless a provocatively interesting post.
Hi James – yes it is an odd wording, but that’s what they say.
And, with regard to the point about users making improvements, this is how they continue: “For example, a small but much-admired startup, Lilt, uses phrase-based MT as the basis for a translation, but an easy-to-use interface allows the translator to correct and improve the MT system’s output. Every time this is done, the corrections are fed back into the translation engine, which learns and improves in real time. Users can build several different memories—a medical one, a financial one and so on—which will help with future translations in that specialist field.”
The full text is here: http://www.economist.com/technology-quarterly/2017-05-01/language
hi Scott
if translation apps have not made translators redundant then what are the chances language teachers have to be worried? 🙂
according to projections for 2024, in the US, English language and literature teachers will see ~10.5 percent job growth and Foreign language & literature teachers ~11%. These compare favourably with (traditionally safe) jobs such as Accounting & Auditing which project about 11%. Interesting to see for Interpreters and Translators the percentage is about 29%! [http://www.projectionscentral.com/Projections/LongTerm].
not sure if you are pulling our legs with that final paragraph? – using poverty of stimulus argument to +support+ usage/emergentism theory : )
ta
mura
Thanks for the cheering predictions, Mura!
As for your second point, the fact that the ‘stimulus’ includes few or no examples of certain strings doesn’t mean that it is impoverished. Is the 4 billion-plus NOW Corpus, for example, impoverished for not containing the string ‘is the man who tall is’? Or even the string ‘is the [noun] who [adjective] is’? Or does it simply reflect current usage?
Hi Scott,
Thanks for another interesting post.
Can I query your date on the origins of the Babelfish? As far as I know it was in the original book of the Hitchhiker’s Guide to the Galaxy, which first appeared in 1979, rather than 1995 as you have here. I’m guessing it was also in the original radio series, which appeared a year earlier. I’m not sure where 1995 came from, as I believe the Babelfish website started in 1997 (at least, according to Wikipedia, which is the best date I can currently find).
Sandy
Thanks, Sandy. Yes, I originally had 1979, based on information in Wikipedia, but then I checked Amazon for an actual edition of the book, and was directed to that 1995 one – which, as you imply, misrepresents Adams’ farsightedness.
Great post Scott. I think ultimately we’re talking about ‘singularity`. Which is said to be a time when computers become way way smarter than any human can even imagine through recursive self-improvement. It has been suggested that the only way to keep up would be to merge with the intelligent technology we’re creating. So in the future, we won’t have a translation app on our mobile device, but a translation chip in our body. I guess when this happens, learning languages would be redundant, a long with many other things. On the other hand, we will be able to travel to absolutely any place on the globe and communicate like a native speaker in any language.
Excellent post Scott. Douglas Adams was indeed prescient (actually in 1978 – radio series started March 8, 1978, book 1979). You can now actually buy your own BabelFish, which comes with 2 earpieces, one for your ear and one for your interlocutors’ ear. It is being launched in the US and will be available shortly (US first) from Waverly Labs. More details here: http://www.waverlylabs.com/pilot-translation-kit/ for $249.
As to the classroom impact of SET (speech-enabled translation) I think it will be a benefit for ELT teachers, not a job-killer. I am certain that being able to communicate immediately (at least on a basic level) will boost the confidence of reluctant language learners and spur them on to want to learn more and master it for themselves, once the novelty of ordering drinks on holiday using your Apple watch translation app starts to pall…..
Thanks, Michael – in fact, the Waverley earbuds you mention are seen being trialled in the YouTube link I included in my post. Pretty impressive.
As for your sunny prediction regarding the use of translation software in the classroom, I’d like to agree, but this will require some initiatives at the level of teacher education: how can such tools be integrated into the classroom ecology? This, in turn, may require an attitude shift, given many teachers (and their institutions) are still resistant to the idea of learners using electronic dictionaries in class.
that’s me Scott – always relentlessly sunny!
You’re right that there may well be resistance, but the rapid adoption of the translation apps by students (especially if they become wearable and embedded) means the teachers may find it hard to ignore them in class, & may find it more helpful to design lesson activities that embrace them instead. Hopefully ETP etc will provide lesson activity ideas…
Chuck in a few colloquial idioms or cultural specific items into an automated translation machine and it produces something totally different. You will still need to have human translators at the very least to proofread something that has been automatically translated so that the nuances of particular messages are conveyed.
For example, there’s a clear difference between how a British person would utter something to a colleague while he is leaving the office to a Korean equivalent. In English we would say, “Don’t work too hard!”. While in Korean, it’s the opposite, “Work hard!”. Again if you look at the cultural references to a message conveyed: if you automatically translate from one language to another, you will still get cultural ambiguity.
So, do I believe that automated translation will replace human translation? I don’t. It may change or evolve the method of translation but humans are still required to look at equivalence and appropriacy in a translated text.
Martin, the example you give is an example of pragmatics, I think. Another example would be the different uses of the word ‘well’ in English. I agree, machine translation still has a very long way to go before it can successfully translate pragmatic meaning.
I moved to Spain at the end of last year and still struggle with Spanish/Catalan. Translator apps have been very useful, google has helped me with a lot from talking to my neighbours to asking for directions. However, it hasn’t removed my need to learn Spanish. Translator apps are still unable to effectively facilitate language learning and so cannot replace language classrooms/teachers.
Embedded chips scare the heck out of me. If this becomes a thing, though, one area of language teaching that would have to change dramatically would be testing, which would probably not be such a bad thing! However, I would much prefer that changes to assessment occur due to better/sound educational motives than as a consequence of more invasive AI.
I often have to tell my students not to first do their writing in their native language and then translate it into English as this is a waste of time. I realize that translation for the sake of comparing grammars is helpful. But translating everything you say or write is not. Anyway, the point is, many of my students “translate” their writing by using a translation app, which makes it impossible to correct. I can correct bad English. I can even correct Konglish. But, I cannot correct “Applglish”. ;o) Plus, there is no real language learning taking place in that process.