Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!

 


Forgot your password?
Close
typodupeerror
×
Technology

Universal Translators? 102

Posted by CmdrTaco from the here's-your-babel-fish dept.
bughunter writes "Carnegie Mellon University is announcing a 'spontaneous' translation system that allows speakers of different tongues to converse in natural language in real time. " I never liked the idea of putting aquatic creatures in my ear anyway.
This discussion has been archived. No new comments can be posted.

Universal Translators? More

Universal Translators?

Comments Filter:
  • Surely the whole point of the babelfish was that everyone had to have one?
    At least you can share this around......
  • Hm. Interesting idea, but there are some instances where the device can't possibly work in "real-time". Simple example that's close to my heart: translating from German to English on-the-fly. Problem: German has the rather odd habit of smacking verbs at the end of the sentence, often a loooonnnng way away from the subject, whereas English prefers to have the subject and verb at the beginning of the sentence (especially in spoken English).

    The story goes that an English visitor to Germany in the 19th century went to see a political debate with Otto von Bismarck, with her translator in tow. The debate went on, with Bismarck himself saying nothing, until finally he rose to speak at length about some minutely detailed point of law. The visitor craned forward to hear her translator, who for quite some time said nary a word--until the visitor became impatient and asked what Bismarck was saying, to which the translator replied, "Please, madam, I am waiting for the verb!"

    QED.

    Ethelred [surf.to]

  • Yeah, assuming it works. I'd still rather see
    us all learning a common _second_ language such
    as esperanto.

    Note: I said second language - not all learning
    a common language.
    Look at http://www.esperanto.org if you're not familiar with esperanto.
  • My favorite German grammer story:
    A german man had just written a two volume book. His publisher was short on cash, and asked if they could publish volume one now, and volume two next year. The author's horrified reply was "Heavens NO! Volume two is all the verbs"
  • I don't want voice-to-voice translation. I want a unit that I can carry around, the size of a Palm Pilot, which, when someone speaks in a language other than English, prints out what they said.

    I want the benefit of hearing the voice of the person who is speaking, but reading what they actually said on my translator unit.

    This voice-to-voice stuff turns me off. I want realtime real-life subtitles!
  • As I'm writing my Master's thesis on Machine Translation (MT), I found out that idioms are pretty easy to translate. In most MT-systems developed today, idioms are just exceptions to the rule which make up such systems. Simply put:

    if (strcmp("John kicked the bucket", str)==0)
    printf("Jan ging de pijp uit\n");
    else
    .
    .
    .

    Other difficulties, which can be much harder than idioms, are the following:

    • Word order problems, as in "John likes Mary" - "Marie plait a Jean" (lit: Mary pleases John)
    • Headswitching cases, such as "John likes to swim" - "Jan schwimmt gerne" (lit. John swims gladly).
    • Expressions for which an exact translation simply doesn't exist. For instance, the difference between "John has been living here for five years" and "John has lived here for five years" cannot be expressed in some languages.
    • Translation irregularies, for instance the German "Schimmel" translates into the English "white horse", but "white horse" doesn't necessarily translate back into "Schimmel", since the literal "weisses Pferd" is also correct.
    • etc.
    • etc.

    And there are the problems which occur in any Natural language program, resolving ambiguities and context being the main one.

    In my experience, it isn't hard to create an MT-system that solves one of these problems. It's putting them together and still have a relatively good system that's the tough part.

  • Esperanto would be great but I doubt that it will ever work, at least not for the US. Look at the metric system... it's a well thought-out universal system of measurement. Universal, that is, except for the US where we insist on clinging to our archaic system of measurement. I actually know seemingly intelligent people that feel metric is "less accurate" because you generally round numbers when converting from the English system. They can't grasp that if everyone used metric there would be no conversion.

    I can't imagine the general public ever grasping the need for a universal language. Most of the US thinks that english is.

  • So far I have never seen an automatic translater
    that doesn't screw up. I bet that even this one
    can make ridiculous translations.

    Enter "Washington Post" at
    http://babelfish.altavista.com
    and get it to translate in French.

    The translation is "Le poteau de Washington"
    poteau like in electrical post or fence post.

    Some other translation I've heard before :

    "Eventail de course" for "Racing fans"
    Eventail is a fan allright but ...

    "He's shaking it" ---> "Il se branle"
    A french would hear :"He's masturbating"



  • It would be relatively simple to say both of the possible meanings. In that specific case, "body" can be used to mean corpse as well, so the still simpler solution of always using "body" presents itself.
  • Let's sneak some funny code into the Chinese-English version at the next summit. For example we could replace the code for "unclear" with the code for "nuclear", then watch the wackiness unfold!

  • Hopefully they won't make it as hard to get as that stupid fish was. Argh!!!!

  • Try 50 years or more! One of the reasons governments started putting money into computers in the early days was for automatic translation.
  • "The spirit is willing but the flesh is weak."
    ==> "The vodka is good but the meat is rotten."

    Just to see what it'd do, I took this phrase and passed it through each of the languages Babelfish supports, and then had it translate that back to English. The results:

    • German: The spirit is ready, but the flesh is weak. (Der Geist ist bereit, aber das Fleisch ist schwach.)
    • Spanish: The alcohol is ready, but the meat is weak. (El alcohol está dispuesto, pero la carne es débil.)
    • French: The spirit is laid out, but the flesh is weak. (L'esprit est disposé, mais la chair est faible.)
    • Portuguese: The spirit is made use, but the meat is weak. (O espírito é disposto, mas a carne é fraca.)

    Not as funny as the famed English-Russian-English translation, but still interesting. (Note that it came closest with the translation through German, probably because English is more closely related to German than to the other languages.)

  • As a holder of a BA in Linguistics, I am highly skeptical of such a device between two closely related languages, much less anything more than a tiny fraction of the 6,000+ languages in existance. Humans will be employed in this regard for a VERY long time, I can assure you.
  • As a holder of a BA in Linguistics,

    As such you ought to know that how to spell "liquor"...

    Sorry.

    Ethelred [surf.to]

  • From simple xenophobic tendancies to straight out evasion or the girls lockeroom. You cannot lose.

    The "cloaking device" had to be one of the most pointless "inventions" of Star Trek. Think about it. You have two warships, both cloaked, trying to fight each other.

    The battle ends when someone says, "What's this button do?", or both sides get bored to death.

    Ethelred [surf.to]

  • Comment removed based on user account deletion
  • The problem with Esperanto is that it's so *horribly* ugly. I guess it's ugly as only designed languages can be, and this particular one seems designed for ugliness (Tolkien made up some languages which are both euphonic and look good on the page. Esperanto is neither).

    If the solution is that everybody learns the same second language, what's wrong with English? I did that, and it's working for me.
  • This is a very difficult thing for a translator to do, even within the same language. You've not only got to deal with exaggerationsl, but inverse meanings.

    -Imperator

  • I'm not Russian, and I don't speak Russian, but I learned of this phrase from a fairy tale.

    The village idiot in some small town did something weird and ended up castrating some wizard guy. This pissed the wizard off, so he cursed the idiot with a flying penis that would follow him around. When the village idiot noticed what it was, he started laughing moronically. At that moment, the penis flew into his mouth and gagged him. (Maybe killed him, I don't remember.)

    Thus, "Don't go catching any flying penises in your mouth." (Phonetically "Vof-lee luv-it," I think.) becomes, "Don't act like the village idiot." Or "Don't act like you don't know what's going on."

    -Robby

  • for(;;); writes:
    This is another example of marketing folks (or, god forbid, the programmers) trying to hype the features of a product and, by neglecting to mention its limitations, ending up lying about what the program can do.
    Given the word "releases" in the URL I would guess this is a press release written by some PR hack in the university's public relations department. You have to get those reporters and TV cameras in so the alumni are proud and send in their checks. :-)
  • Now all it needs to do is make the peoples lips move at the same time like in star trek.
  • I couldn't even make through half the posts here. Much bitching about how the system can't possibly do true real-time, accurate translation from any given language to any other. Duh. The point here is these people are setting up a focused, usable system for use by perhaps two groups: high-level diplomats who probably speak more than one language apiece and thus can tweak their use of colloqialisms as needed, and tourists who will not be using the system to get the latest guffaw local stories but rather to find food, shelter, and locations of interest. Such applications tend to restrict the vocab needed and countries involved.
  • > Until you have good algorithms, all the Beowulf
    > clusters in the world aren't going to do you a
    > damn bit of good...

    Well, that might not be true. AI research constantly butts its collective head against NP-complete problems. If we had non-deterministic machines (which people are working on, and probably at a faster pace than any proofs of P=NP), it would bring new life to the subject. My college AI textbook was filled to the brim with heuristics for avoiding exponentially large computation in what were essentially NP-complete problems.

    But in the context of that guy's post, I agree with you. Moore's law is a false hope here.

  • Re:reeducate the speaker. (Score:1)

    by Anonymous Coward
    > we all need to talk either scientific journal
    > like precission, or only allowed limited
    > extravagance scripts. let's just say we all talk
    > like greek drama actors.

    Unfortunately, that wouldn't solve it either. Strange and complex sentences can be built up even from proper English. Take the sentence "Now is the winter of our discontent made glorious summer by this sun of York." The word "sun" is a pun here. The "son of York", the son of a guy whose last name is "York", has just won the War of the Roses. But "sun of York" describes the change in weather from winter to summer. This weather change, in turn, mirrors the change in the social climate of England from war to peace. Translating that sentence to (for example) French, and keeping all the nuances, is something that human translators can do better than computer translators. Doing this by machine would entail writing programs with deep understanding of the nuances of the languages and of the subject matter being described.

    In another example, the first two lines of Camus' The Stranger, originally written in French, translate to: "Mother died today. Or was it yesterday?". However, the original nuances of the French words Camus used in that sentence, and the context of the novel as a whole, would be better represented by the translation "Mommy died today. Or was it yesterday?". This is a small difference, but an important one for a good translator.

    The problem isn't necessarily the vocabulary used. It's the meaning of words, taken in context, that is hard to figure out.
  • This is pretty impressive, but nothing too surprising I guess. I wonder how customisable the 'voice' of the translator is.. wouldn't want a big bunch of tourists walking around sounding the same :)

    Actually, it would be quite cool if it adapted to the wearers voice and made it sound like them with a foreign accent..

  • Cool, but how does it SOUND? All the computer speach products I've heard sound very monotonous.

  • It would come in real handy at those Star Trek conventions for people who don't want to go to the trouble of learning Klingon.

    I wonder if it does a better job than AltaVista's babelfish.
  • If this is the software I think it is, I have some experience with it from 18 months or so ago. The speech synthesis being used there was monotone in the sense of a person speaking with little intonation, but of fairly good quality (the accuracy is, for what I worked with, more a function of the raw synthesis data than anything else). It's also modular, so that the synthesizer can be replaced with another one with minimal effort.
  • Now, I want a phaser and warp accelerators.
  • rather have the fish in my ear. Next thing you know they will try to come up with a replacement for towels...
  • the article certainly seems to imply it does a better job. It says that earlier generations required 'perfect syntax and speaking style' and this one handles colloquial speech, which seems to put the translation far beyond the efforts of babelfish.

    on the other hand: 'The ums, urs, interruptions, hesitations and stutterings of spontaneous speech are automatically recognized, filtered and properly prepared for translation.' I would hope that it handles a lot more than my ums and urs....what about all my 'likes'?
  • People have been working toward this kind o fthing for a long time , with slow and painful successes . They are still no where near having a reliable solution . A friend ( and member of our LUG ) is a Mathematical Linguist and works on Automated Machine Translation software . His comments over the last year give me good reason to be skeptical .

    I am betting that this turns out to be the kind of thing that voice to text did . Wonderful idea , almost worked first time out . Almost worked two years later . Almost works five years later . Still wouldn't waste my time with it now .

    Besides , who wants to wisper sweet nothings through a throat box ...
    Your squire
    Squireson

  • How, exactly, does it handle colloquials? (Score:3)

    by Angwe ( 18648 ) on Tuesday July 13, 1999 @08:08PM (#1803673)

    Okay, so it can handle colloquialisms, but what does it do with them?

    Does it do a literal translation of the colloquialism? (Russian phrase "Don't go catching any flying penises in your mouth.") Or does it try to find the closest idiom in the other language? (Said Russian phrase would best be translated, "Don't act stupid, like you don't know what's going on." It's from a Russian fairy tale.)

    Also, one encounters the possibility of there being no equivalent phrase. And what about weighted concepts? Many Asian languages, due to cultural influence, have an inherent extra emotional meaning attached to declarations of honor, but there is no easy way to translate this into English.

    This sounds like a HUGE step forward, but I'm still gonna be skeptical and double check machines with people.

  • You'll have to speak up, I'm wearing a towel.

    I'm sure once you got used to a fish stuck in your ear it would be a lot more pleasant than lugging around a translator. Well, i'm sure they'd get smaller eventually, but you know what I mean! At least you can't lose the fish and get stuck in Paris asking where "le crapper" is.

  • Un which article are you referring too. Cold Fusion does exist but it doesn't generate more energy than it uses.
  • not to mention a transporter.
  • Wow! It'll be just like The Last Star-Fighter! Only . . . Without the star-fighting and the aliens and all.

    One question: How will this effect those people who like to get away with things by pretending they don't understand english?


    ---
    seumas.com

  • I really don't know, I've had a tough time studying all those Chomsky-ish phrase structure grammars and non Chomsky-ish categorial grammars...

    Anyway, deriving semantics from phonological information (speech -> phonemes -> morphemes -> lexemes -> semantics) and then converting back into another language never seemed to me as a trivial task.
    The plethora of phenomena in language are just intimidating.

    I still doubt it's an AI-complete problem ;)
  • This product has great practical value for terrorists who want to talk to their american captives without hiring interpreters. (who wants to work for terrorists? We have all seen enough bond movies to know what happens to ex-terrorst employees.)

    Quick, put this product on the export restrictions list!!!!!!

  • "Don't go catching any flying penises in your mouth."
    ...
    It's from a Russian fairy tale.

    OK. I really need to read this fairy tale. Any translations (to English) online? :-)

    "And remember children....Don't go catching any..."

    dylan_-


    --

  • Got any URLs to pages with useable (ie complete schematics, documentation.. that kinds of thing)?
    I'd love to fiddle with cold fusion a bit. Not to mention the looks you'd get..

    Some shmuck: "Hey man, how you been?"
    You: "I'm working on cold fusion at home. So, what have you been up to lately?"
    *rofl*
  • There are experimental phasers.. there was an article here... ionise [sp?] a path through the air, and send a current through it..

    It stuns the poor victim by shorting out their muscles... theoretically, you could pump more juice through the air at them. I dont know about vaporization though..
  • but will it allow me to understand 'women talk' as well? Nnow that would be an acheivement.
  • If that was sarcasm ignore this but...

    I'm pretty sure that slang won't be given up just for this. English's ability to constantly evolve is why it is now the international language.

    Also, you do realize "got to" & "ought'a" are slang right?

  • What are you talking about?

    If you read up on it for 5 minutes you'd see it has English & German as input languages(I'll give you the two very similar language point there) and Japanese, Eng. and Ger. as output. It also is close in Spanish and Korean. English=Germanic. Spanish=romantic. Korean/Japanese=neither.

    More than 20% of the world's population speaks one of these languages.

    More than 10% of the world has used a telephone! Think about it. China is 50% of the world's population. At least half that country(all but the western part probably) has used a telephone. As has almost everyone in the US. And a majority of India has as well. As has Europe. Think before you type.


  • That's what I want to know.

    And simply put, there is NO WAY that it will be able to catch them all. Colloquial phrases are always changing. Remember "Don't have a cow, man"?

    And no-equivalent-phrase is always a pain to deal with. I studied advanced French and came across the occasional "not-really-translate-able" phrase. Those are not fun.

    Online friends of mine are blind and use text-readers. They have enough problems with mispronunciation and tripping over *emphasized* words on a system that doesn't allow for HTML formatting.

    And Babelfish online has made some interesting mistakes. Can't remember the URL, but lokisdottir over on Geocities posted some translations of her writing to another language and back again to show how badly Babelfish can mess up.

    Instant translation? Not anytime soon. *shrug*
  • Take this case
    A guy has met a girl he likes a lot and
    they are in the bed about to do some interesting
    exercise. He want to complement her but doesn't
    speak a word of english.
    using his little translater he enters :
    "T'as un beau corps"

    The translater says :
    "You have a beautiful corpse"

    How long do you think he will keep his hard on?

    In French corps translate to body and corpse.

  • (Russian phrase "Don't go catching any flying penises in your mouth.")

    I don't know, the literal translation isn't so far off ("Try not to suck anybody's cock on your way to the parking lot!" -- Clerks)

  • "Je suis en beau joual vert" (I'm pissed off)
    to "I'm in nice green horse"

    "A sweepstake" to "Un steak balayé"
    "balayé" like with a broom.

    "Vaincu" (defeated) to "Twenty asses"

    "Harry" to "Poilu" (meaning hairy)

    "Racing fans" to "Eventails de course"
    Eventail is that fan with blades to cool you off.

    "Washington Post" to "Le poteau de Washington"
    poteau like in telephone post.

    "Micro Soft" to "Chiquito Blando"
  • I'm Russian myself (and 20 years of age) and never have heard such an expression. Really.
  • Calm down. The HHGTTG is a book as well, so instead of picking up all kinds of arbitrary items to set up in an elaborate system, you need only read to chapter 5. And presumably keep reading, because the book/series is so addictive.
    --
  • Hey! What's wrong with Babelfish? They don't cause any problems and they're highly cross platform (all you need is an ear of some sort). Sure they may not be avaliable in stereo yet, but who needs stereo when talking to a Vogan anyways.
  • I ran the same check and saw that all of the matches has correlation values in the low 50% range -- meaning they had the word cold or fusion, but not both. Most of them had cold. It seem seemed like a pretty loose journal, but not as bad as you've painted it.
  • Is there a delay between the uttered "foreign" speech and the translated version or what? Wouldn't this either result in either extremely lenghty conversations or confusing overlap?

    Hmm. I wonder whether the original speech could be masked so that it wasn't audible. A little chip could record the wave patterns of the speaker and chirp out their inversion. Original + inversion = silence. That would be so sci-fi!
  • Posted by 2B||!2B:

    In a course I took that emphasized having machines understand the big picture of what's being said (e.g. figuring out the context and leaning on a large set of rules about the real world beyond just basic vocabulary), it was apparent that you could have a voice synth put an appropriate emphasis on the right words. The problem with doing that is that it would be much more difficult to do in real time unless you gave the machine a while to hear the entire phrase before it attempts to say anything, after running everything through yet another complex rule engine which figures out points of emphasis. I've done enough English ->Spanish and Spanish->English translations on the fly (myself, not a machine) to know how difficult it can be for even a person, let alone software.
  • Depends on how real time it is. I'd think it would break the sentances down by clauses, in which case sentance long phrases wouldn't work very well. Hopefully it would just give a literal translation -- for most colloquial phrases (including your example), the meaning is pretty obvious, even if it's not something we're used to hearing.

    I took colloquialisms in this context to mean more informal, but hey.

    In regards to weighted speech (i.e. the masculine vs feminine structures in japanese), it's going to first be a tool for business, so it'll probably just translate to formal convorsational. People are going to be using this to talk about oil prices, not translate the Koran.
  • Yakman> This is pretty impressive, but nothing too
    Yakman> surprising I guess.

    I'll take issue with that. It's pretty impressive, and damned surprising, but probably untrue. If they actually had what they claimed to have, they'd be in line for the Nobel prize. The CMU press release makes no mention of any flaws in the product; for all the reader knows, this product can translate text from any language into text in another language as well as a human. Again, this would be damned surprising! If this were true, we'd be seeing this press release on the cover of Time, and almost every other publication. A program able to process natural language that well would be able to pass the Turing test. And think about it -- if a program could really translate anything (as well as a human) from English to French, why not modify it to translate from English to a recursively-enumerable subset of English (as well as a human)? Who would need programmers after that? If they really had what they implicitly claim to have, it would be a big deal.

    Writing a program that can translate natural languages as well as a human is a holy grail, and not something that should be claimed (or implicitly claimed) lightly.

    I'd bet anything that this is babelfish-level translation software with voice recognition software (probably of comparable quality) slapped on. This is another example of marketing folks (or, god forbid, the programmers) trying to hype the features of a product and, by neglecting to mention its limitations, ending up lying about what the program can do.

    This product is mildly interesting from a software development point-of-view, but not anything I'd go firing my human translator staff over.



    -----------------
    "The spirit is willing but the flesh is weak."
    ==> "The vodka is good but the meat is rotten."
  • Well, a cloaking device is simply the ultimate form of camoflage.(sp?) I personally liked the Predator's technology. That could be handy.
  • Yeah, that Nasa/Odessa/Nato triumverate could use these devices to turn ordinary tourists who want to speak in the language native to the country they're in into Ninja killer drones. Very upsetting indeed.
  • Err actually most languages evolve, and continue to evolve. This is not why English is so widespread. It's got more to do with empires, both political and economic. One of its advantages that I have found, as opposed to some other languages (I do speak others (Latin based, although I am trying to learn one Malay based language too))is that you can completely bastardise it and still be understood reasonably well.
  • So its not as groundbreaking as it sounds. (read the press blurb). The questions people have been asking here on how it handles colloquialisms and 'the general problem' are mostly irrelevant. I saw a demo of a system (possibly the same one, as it was working on the same domain) translating on the fly between English and German, and it worked just fine, over a year ago.

    The way these things work is, since they are in such a restricted domain, (in this case ordering plane tickets and the like), it isnt impossible to give the machine a fairly complete picture of all possible tacks a conversation at a ticket desk might go. It can then use this internal, and fairly complete, model of its domain to get a model of the meaning of the conversation. This intermediate step into an 'interlingual' model means that internally it doesnt need to care what language you want it to speak in.

    Plug in a module to translate from an interlingual response to eg French, et voila, the computer will respond in French. Translating _out_ is much easier than translating _in_, since you already have a precise representation of the response internally; but translating _in_ is made much easier as you dont need the machine to understand the entirety of the language.

    This is all a bit glib, and I'm not knocking the achievement; I'm just putting some perspective on what's been done. IANAL(ingistics researcher) ;o) . Anyway, what I'm trying to say is: the only reason they are making any progress is by restricting the domain to one where the conversations are simple and well-understood. That way, they can do a *far* better job than the Babelfish, but its not Star Trek technology yet.

  • > A long time? What's this amount to, 15 or 20 > years? It's not like this is a project that has > eluded humanity for centuries.

    Actually, it kind of is... software is just the latest and best tool in this quest.

    > Breakthroughs are always just around the corner.
    > Maybe your friend has given you good reason to
    > play skeptic. The again, maybe he's not the one
    > who makes the breakthrough.

    Oh, bullshit. Some reasonable skepticism is always healthy. It's not like his friend has given up trying.

    > By the power of Greyskull! I invoke Moore's Law!

    *sigh*

    When will you people learn that these things aren't just a matter of throwing more CPU at the problem? You need better algorithms first.

    Until you have good algorithms, all the Beowulf clusters in the world aren't going to do you a damn bit of good...
    ---
  • Dude, learn to count. Minnimum 5 BILLION people on this planet. There are, at most, 1.5 billion.

    I dunno where you went to school, and I didn't do well in math, but my calculator tells me that's not 1/2, or 50% either.

    While I think that you're probably right about more than 10%, I don't think it's much higher. When you get to 50%, that's something. By that time, ~6 billion, half that is 3bil.

    Shit, we need more telephone numbers

    bye
  • Translation, especially online, is a damn hard thing to do. I've studied with one of translation pioneers and have heard many stories of a multimillion dollar demo that concludes by translating 'Hello' into 'thirteen'. Part of it is the fact that translation is 'interlingua' based, where 'interlingua' is some hypothetical 'language of the mind', which presumably uses grammar to verbalize some of world knowledge. Now, the computer has no word knowledge. This is going to sound like babelfish - at best, a little bit usable... That's my 5 cents. I hope I'm wrong anyway, this would be great technology.

  • This would be an amazingly cool thing---if it worked. We'll have to wait for the demo to know for sure, but call me skeptical. There are too many jobs this system would have to perform, each of which is a potential point of failure. Here's a brief rundown:

    • Voice recognition. In each language. Note that the current state of the art in voice recognition involves a *lot* of hand-tuning, which would then have to be done in each of the six languages. Also, to get reasonable performance you really need to let the system get to know you---not practical in this case ("Um, hi, mr. travel agent, could you talk into this box for about an hour? Be back in a jiffy.")
    • Sentence parsing. The absolute best results these days have just under 90% accuracy; the best results (that I know of) with something approaching real-time speed are down around 80%. Some of that (I'd guess 5% or so) is legitimate ambiguity, that probably wouldn't affect a translation much, but much of the inaccuracy would cause problems.
    • Word-sense disambiguation. This is far more important to a real-time translator, but the state-of-the-art still isn't perfect, and more importantly, the SotA algorithms generally work on entire articles, not just sentences, so (particularly for short conversations) they'll be at a disadvantage. This is the source of most of the classic translation errors (like "the vodka is eager, but the meat is rotten"). More importantly, it's harder to forgive than most other errors, since e.g. pauses or slightly off grammar generally don't render the text unintelligible, but using the completely wrong word can totally obscure the sentence. The speech translation problem is slightly worse, too, because there are (typically, in most languages I know of) more homonyms than homographs.
    • Text-to-speech. This area's not so bad, in that computers can generally produce (relatively) understandable speech, but in a Talking Moose sort of way. Again, though, this has to be hand-tailored to each individual language, at least to some extent.

    My concern is that each of these things will introduce substantial error; and further that the SotA in speed is not the same as the SotA in quality, so that in order to do real-time translation, the quality degrades even further.

    Another, more cynical/paranoid concern, is the fact that (most of) the conversation will be between people who are, presumably, prepped for it. They can be told to speak slowly and in a clear voice (which is reasonable enough, I suppose), to restrict their vocabulary (which is not reasonable), and to limit their interaction to certain other languages (also not reasonable). The Heidelberg man-on-the-street (Straßenmann? ;) conversations should be more informative in this regard. But I wonder if their system can actually handle translations between any two languages, or only certain combinations? And if they can translate between any two, do they use a "hub-language" or translate directly? Two clicks at babelfish can show you just how much further the former would degrade the quality....

    Oh well. I'll certainly keep my eyes open for the results of the demo. Are any slashdotters going to be there? Make sure to post your impressions. :)

  • "Actually, it would be quite cool if it adapted to the wearers voice and made it sound like them with a foreign accent.."

    I doubt this one does that. Though I can envision how that could be done technologically, it would take a lot more processing power than would be available in such a situation.

    Oh, I wasn't implying that this one does that, chances are it doesn't.. just that it would be cool if it did :) Like someone else mentioned though, it is all probably 'too good to be true'. I mean, look how much trouble Babelfish has at times and that doesn't even have to do speech -> text first (not to mention text -> speech after translation). After reading enough of these things on the 'net you learn to treat them with skepticism until proven otherwise I guess :)

  • Hmmm, very revealing [cmu.edu]... According to the CMU website, the system (presumably Janus) is limited domain, which certainly makes things much easier. Still pretty cool, though.
  • I'd take this with huge lumps of salt, thank you.
    First of all: since when even *recording* speech
    to text has been solved? Without the text form
    translating between languages sounds like hogwash.
    Oh yes, we humans can do it, witness simultaneous
    translation in UN and so, but a machine doing it
    now? Maybe in few decades or so when they work out about gazillion big and small problems in speech processing and natural language processing, let alone translating. But now? Rubbish.
  • A long time? What's this amount to, 15 or 20 years? It's not like this is a project that has eluded humanity for centuries. Breakthroughs are always just around the corner. Maybe your friend has given you good reason to play skeptic. The again, maybe he's not the one who makes the breakthrough. By the power of Greyskull! I invoke Moore's Law!
  • You seem very sceptical.

    I saw a presentation of this at a conference some time ago. Two observations:

    • This is not some garage project of some small university, this is a HUUUGE project with many big companies and lots of funds. By far the biggest project in the translation field.
    • Their trick is to take a very small domain, in this case just the reservation of rooms at hotels. You clear out a lot of problems if you can at least expect one meaning for one word ('bank' is unlikely to mean riverbank in such conversations, to use a popular linguist example).

    I do not remember quite what it was, but I do think they did something with accents in the speech synthesis.

    --Sander.

Slashdot Top Deals

He has not acquired a fortune; the fortune has acquired him. -- Bion

Close