Future of Machine Translation

[DanWang]

Anyway, what do you all think on the future of translation and interpretation? It seems, from what I've read and heard, that technology is rapidly replacing labor in this field and it might not be worthwhile to study so hard and end up competing with softwares and eventually lose out to it. Of course, current technology might not be so powerful as to replace simultaneous interpretation, but there have been recent developments that have resulted in softwares that can reach a 90% accuracy in interpreting spoken content from one language to another. It's just that it is not yet popularized. What if technology like these do invade the market a decade from now on? How will interpreters and translators survive then? It appears to me that becoming an interpreter might just be a bad career choice in its own right. 

[icebear]

How will interpreters and translators survive then? It appears to me that becoming an interpreter might just be a bad career choice in its own right. 

 

That's true of almost every field today; maybe the more complex ones are 20 years out. 

[Lu]

 

but there have been recent developments that have resulted in softwares that can reach a 90% accuracy in interpreting spoken content from one language to another

That means that at least 10% is wrong. 90% sounds like a lot, but it's much too low to be useful.

 

Google translate is useful if you want to have an idea of what a text is saying, but it's unable to accurately translate stuff, and that's written. Translation software is useful, but nowhere close to replacing actual translators. For reading contracts and other complex and important texts, translating anything that needs to be 100% correct and also in correct English/other target language (let alone good, readable English/target language), interpreting (let alone simultaneous interpreting), especially if the speaker doesn't speak the source language perfectly standard, translators will remain indispensable. No translation software comes close to replacing people.

[DanWang]

@Lu

 

True. Current technology is lacking but my speculation is that a decade from now on, things can be very different in this field. Thanks to Moor's law, that 90% accuracy will eventually come to 98-99%. It might take 8-9 years for current technology to mature and another 3-4 years to market. But in 15-18 years, I think it will be popularized. By then, interpreters will be screwed. 

[yueni]

Edited to remove personal info and add content...

 

@Lu & @DanWang

 

I'm going to have to agree with Lu here about technology replacing interpreters and translators. Sure, there will be improvements in technology, but given how flexible language is, I doubt that technology will be able to fully replace a human translator or interpreter in areas requiring a high level of lingual flexibility, subject matter knowledge, and accuracy.

 

I believe that technology will replace subpar translators for sure, because there are a lot of those in the market.  I think technology will raise the bar for market entrance for both translators and interpreters, and I see that as a good thing. However, the clients for whom high level accuracy and subject matter knowledge is paramount will not be replacing their translators and definitely not their interpreters with technology. And these are the clients I'd strongly prefer to work for because they generally will pay a lot better for your services because they understand the value of a quality translation or interpretation.

[Lu]

DanWang: This is rather off-topic but I'd love to continue this discussion. Yes, Google Translate is very helpful if you want to know what a given text is about, or if you need a few words translated and nothing is really hanging in the balance if you get them wrong. But it's has been around for years now, and it's not nearly as good as it thinks it is. Last year it asked users to help improve the machine by choosing which of two given translations was better. Of the options given, I don't recall seeing one completely correct translation, and all too often both options were just incorrect. For Chinese, on word level it's much less good than either Pleco (free dictionaries), Nciku or Iciba; for English/Dutch words it's okay, but not better than my paper dictionary. On sentence level, for Chinese it's usually bad to nonsensical. For English to Dutch, it still hasn't figured out a basic grammar point, which makes most sentences readable but incorrect. I tried it for reading a Swedish article in English about a year ago and while I could read the article (only a few sentences were really incomprehensible), there was a certain Swedish word that returned again and again and that google didn't translate, while I understood what it meant after a few instances. Perhaps there are languages that Google really translates well, but so far I haven't seen it.

 

And mind, that's for written, correctly-spelled Chinese sentences, with usually not even any weird grammar or slang words. Just normal Chinese text with a clear, unambiguous meaning, and it can't do it to a useful level. If your job is summarising the newspaper, you might need to worry a bit (but only a bit, because so far nobody's built a summarisation machine). But for any translation that's more than on the fly, let alone interpreting, let alone simultaneous interpreting, there's no need to fear machine translation.

[yueni]

@Lu

 

Speaking of machine translation etc. have you ever tried Google Translator Toolkit? It is much, much better than Google translate at the sentence level, but it is hampered because everything you put in it is entered into a massive language database. It's super cool to use, but I almost never use it because clients obviously do not want their proprietary data going into a language database that they don't control.

 

I did use the toolkit for classroom exercises when I was in school, and was quite impressed at its accuracy and how natural the language was in comparison. I don't know how it does going into Dutch, but I'd wager it's much better than Google translate.

[simplet]

The idea that machines are going to replace simultaneous interpreters in the foreseeable future is frankly ridiculous. Like Lu said, Google translate hasn't made a lot of progress and this is only for written translation. Now imagine that we still can't even get speech regognition software to work correctly and see what that gives you. If you have the slightest error in the speech recognition, that error is going to not only appear in the translation, but it's going to propagate through the whole text. That's because the way machine translation works at the moment is flawed to the core. Basically it works by looking through the x preceding words in the text for each word and going through a corpus to find the most frequent translation for that word with the same combination of words preceding it. This means the slightest error will resonate through the whole process.

 

The problem here is that even if your speech regognition software is perfect, the speaker himself will make some mistakes, misspronounce some words, not finish his sentences; he might not be a native speaker himself. The worst for the machine will be all the syntactic breaks that are inevitable in oral speech. Remember that every small mistake will accumulate over time and make the translation more difficult. Once you've managed to master all this by some miracle, you'll need your computer to render the correct intonation, rythm, emotion, humor, innuendos in a way that is pleasant or moving for the hearer.

 

In the end I guess it's possible that you'll get a machine to do simultaneous interpreting, after we reach the singularity or whatever, but by then there probably won't be any more finance analysts, meteorologists, manual workers, economists, or teachers... All those things are a lot easier to do for a computer than interpret spoken language in real time. For all we know by the time computers get the hang of natural languages we might be uploading our thoughts to each other in real time.

[imron]

Thanks to Moor's law, that 90% accuracy will eventually come to 98-99%

Pedantic note:  This has nothing to do with Moore's law.  Moore's law states that "the number of transistors on integrated circuits doubles approximately every two years".  This is great for other areas of computing (e.g. computer graphics, simulations etc) because we've known how to do various things for decades we just haven't had the computing power to do them quickly.  As the number of transistors increases we get that computer power and now all sorts of things that were once only possible on a super computer, are now possible on the phone in your pocket.

 

The problem with machine translation is that we don't know how to do it well yet, so it will require a breakthrough in AI and natural language processing rather than just an increase in computing performance, although an increase in computing performance will certainly make it easier for that sort of breakthrough to occur, it's by no means guaranteed.

 

Personally I think automatic translation tools will continue to improve, but there will still be a need for human translators and interpreters for quite some time yet.

[Nathan Mao]

Besides all that, I study Chinese because it is cool to speak Chinese with Chinese people.  It is fun to walk on the streets of Beijing and be able to talk to everyone, to listen in to what they say about you, to have philosophical conversations with your masseuse, to hear the taxi driver curse out foreign drivers as idiots.  While it is not fun to argue, it is satisfying and gratifying that my wife can vent her feelings and frustrations in Chinese and I can understand it.  It is interesting to watch television shows and movies with different cultural assumptions, and fascinating to compare with US cultural assumptions and learn inobvious truths from the comparison.

 

That I can use the ability to get a job is just icing on the cake. The cake would still be worth baking/eating w/o the icing.

 

I fully agree with Imron's last post.  I was already composing my response in my head to the previous posts when I saw his perfect encapsulation.

[Lu]

Nathan Mao: Of course. But if machine translation will be good enough for simultaneous interpretation, it won't be that much of a stretch that it will also be good enough to understand taxi drivers and angry wives.

Which just makes it all sound even more unlikely imo.

[tysond]

I pay for translating companies to do lots of translation for my job.

If this Moores law argument were to hold we should be seeing decreased prices, faster turn around or higher quality.

We see none of those things at any significant pace. i don't think we will see anything but a moderate improvement in a 10 year span. Not even a factor of two. The translation of subtle/deep meaning and the production of natural sentences is a very difficult problem. I am not sure the current pattern matching approaches will scale well enough to solve it.

The NSA is not short of computing resources but is still hiring translators apparently.

[Nathan Mao]

@Lu

Maybe.  I guess you could have a microphone and an earpiece, and as they are talking it could translate it directly for you.

 

But that's not the same as being able to do it yourself, in your own brain.  For one thing, having a machine translate it for you would just add to the noise...could you understand it clearly as your spouse was yelling?  Wouldn't there be some delay?  Would a machine be able to include the emotional stresses in your speech as it translates back for you?

 

Anyway, my point is just in response to "it might not be worthwhile to study so hard and end up competing with softwares and eventually lose out to it." from the original post.

 

Right now, software AI has improved to the point that computers can compose pleasant music.  You can play a tone on any instrument into a microphone, and the computer can automatically harmonize it for you, and has been able to do this for years.  GarageBand lets you easily compose backing tracks so you don't need a band to rehearse with.  Rocksmith apparently has extremely smart AI accompaniment that will play differently depending on how you play, stopping when you stop, playing more aggressive when you play more aggressive, etc.

 

Yet people still practice to learn to play instruments better.  People still form bands.  People still pay money to get into a club where an unknown band is playing live music.

 

Job opportunities may well reduce substantially as machine translation improves.  But there is nothing like doing it for yourself, even if you never make a dime doing it.

 

Of course, there is also only so much time in the day.  You have to pick and choose your passions, and your hobbies. The Chinese language is my passion. I spend hours a week on it when I don't have to, just because I like learning and using it.