The machine-translation pipe dream

An article in Sankei News about NEC putting machine translation onto mobile phones (Japanese) has created a bit of buzz on Honyaku (a mailing list for J<>E translators).

Every time some new development in the machine translation world comes out, translators start to worry about whether they're going to be put out of work. Let me spare you the suspense: not for the foreseeable future. And perhaps more importantly, by the time machine translation has replaced human translators, just about every knowledge-intensive job — from programming computers, to flying jetliners, to performing brain surgery — will also have been replaced by a computer.

The reason is simple: machine translation is an AI-complete problem, which means that in order to solve it you've got to be able to make the computer as smart as a human. And although I think that this day is coming — probably within my lifetime — when it does come, everybody is going to be out of work. Hopefully that will be a good thing, because the alternatives would be very, very bad.

Despite what the relentless barrage of "articles" from the PR machines of megacorps would have you believe, machine translation hasn't gotten much better since the 1960s. The reason that so many smart people go to work in this field is that the solution always seems so tantalizingly close. That's been true ever since the first "machine translation" system — really just a very small phrasebook — was demoed. Fresh, bright young minds enter the field, find out how intractable the problem really is, and hopefully move on to a more productive field, or at least lower their expectations :)

What is changing, and what the article about NEC shows, is that information and services are becoming ubiquitously available. A glorified voice-activated phrasebook on your mobile phone is just a small part of a transformation of our society and culture whose end result will be something that frankly, I can scarcely imagine.

4 comments to The machine-translation pipe dream

  • Have you ever read any Charlie Stross? I think you’d like his stuff.

  • Thanks for the link, Adam. Great site. I’m going to have to try one of his books.

  • Ben

    What you say above might be true for some kinds of machine translation, but the statistical method used by Google translation is actually pretty powerful. It’s good enough to that a reader can make sense of its translations from English to German and vice-versa. Sometimes the English is even correct. The Japanese to English translation is not so powerful, but I’ve actually used it to get a start on some kinds of translation work. It also has a gigantic vocabulary, so it can be used as a kind of super-dictionary.

    So I don’t think it’s correct to say that machine translation hasn’t moved on since the sixties.

  • Yes and no. Statistical methods have been around in machine translation for a long time. They kind of came into their own in the 90s, when commonly available computers finally got powerful enough to implement them cheaply.

    Then they starting blowing away the old hand-coded types of solutions in competitions aimed at getting MT systems up and running quickly. But in terms of ultimate quality, I’d say that statistical MT is at best an incremental improvement over hand coding. And we’ve been doing incremental improvements for decades. What we need is radical improvement.

Leave a Reply




You can use these HTML tags

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>