Coffeehouse Thread

20 posts

Forum Read Only

This forum has been made read only by the site admins. No new threads or comments can be added.

Universal Translator

Back to Forum: Coffeehouse
  • User profile image
    Shining Arcanine

    I think that in the next decade or two, the next major breakthrough will be the development of the Universal Translator, most likely by Google.

    What is Microsoft doing in this area?

    Edit: By the way, when I say Universal Translator, I am referring to a computer that will sit in the telephone line and translate phone calls in near-real time, making it sound like the other person is speaking your language to you and like you are speaking the other person's language to the other person.

  • User profile image
    jason818_25​3.33

    Please define universal translator.
    Are you thinking of something like Alta vistas babble fish? Where a person is able to type in the language and then choose what language they wish for it to be translated into?

    I just read your edit and that would be a cool thing to have.

    Getting a computer between you and the other caller is probably not the most difficult thing to do. Seeing how many people already use VoIP (Voic over internet protocol) for making phone calls. Would this be a program that sits on your local machine that does the translating or something owned and operated by some one like Google or a phone company like AT&T?

  • User profile image
    Soviut

    Full-on Star Trek.

    I must admit, it would be very impressive.  If it could actually resynthesize your voice formants to make it sounds like you, yourself,  were speaking another language, that would be even better. That in itself would be a huge leap forward in voice analysis and synthesis.

  • User profile image
    Bas

    I wonder if it's even going to be necessary. I'm suspecting that English will become more and more widespread, and that in the next two decades more and more people will speak it. Especially considering that the internet is becoming more and more accessible to more people. More exposure to the English language results in more people learning to speak English.

    If, eventually, we can all communicate to each other using the same language, then there's not such a big need for a translating service anymore, meaning less research, meaning no universal translator.

  • User profile image
    JonesJ
  • User profile image
    AndyC

    Sven Groot wrote:
    Perfect machine translation. Currently automatic translation is laughably bad even on the most ideal inputs. Spoken text is often ungrammatical and contains colloquialisms and expressions.



    And let's not forget that's when it know's which language it's converting from. For a truly transparent UT system you'd need to be able to identify the source language from within a few words. For some languages that's probably doable, but for others it might be extremely difficult (especially given all of Sven's other points).

    To be honest I really can't see this happening in my lifetime, if at all. I suspect that we would see a convergence of natural languages long before we reached a point that technology can solve this for us.

  • User profile image
    Sven Groot

    I don't think this will happen that quickly.

    In general, such a device would require the following, in order of increasing difficulty:

    1. Perfect (or at least not-annoying) voice synthesis. This is possible with today's technology (even though MS's TTS tech in Windows would lead you to believe otherwise), so it's not a problem.
    2. Perfect speech recognition. It may not misrecognise even a single word or you might greatly insult whoever you're speaking to. It must function perfectly even in less-than-ideal circumstances (e.g. a person speaking on a cell-phone on the train). Current technology does not allow this, but I could see it benig developed in the next decade or two.
    3. Perfect machine translation. Currently automatic translation is laughably bad even on the most ideal inputs. Spoken text is often ungrammatical and contains colloquialisms and expressions. A machine translator must have an extreme amount of real-world knowledge to recognize these and map them into another language. We're not even close on this one. Personally I believe that  unless we can create a functioning replica of the human brain the amount of data and the type or reasoning needed to really do this without lots of mistakes will make this impossible. This will take 50-100 years at least, if ever. IMO, of course.
    4. Knowledge of the future. You want real-time translation. Unfortunately, not all languages use the same word order. In Japanese, verbs always come at the end of the sentence. I can make a huge compound sentence, and the verb of the main sentence would be all the way at the end. In English, this verb is one of the first things in a sentence. So at the very least, the translator must wait until the sentence ends before it can translate. And determining where a sentence ends is not always an easy thing to do in colloquial speech.

    I'm in general very optimistic about the human capacity for solving impossible problems, but this particular one I do not expect to see happening in my lifetime.

    EDIT: To demonstrate the difficulty of point four, consider the following sentence in Japanese:
    "kinou konbini de katta."
    It means "I bought it yesterday at the convience store". What "it" is isn't specified; the sentence doesn't have an object, which is allowed since Japanese is highly contextual. "Katta" is the verb here, the past tense of "kau", to buy (tranlated word-for-word the sentence reads "yesterday convenience store at bought").
    To make a compound sentence in Japanese, all I have to do is put a noun at the end of an existing sentence:
    "kinou konbini de katta hon."
    I just added "hon", which means book. Now the sentence is "The book I bought yesterday at the convenience store". So even if the translator heard "katta" it cannot safely assume I'm done with the sentence. And any noun I say after "katta" must come before the verb in the English translated text.

    Let's continue that:
    "kinou konbini de katta hon wo yonda" = "I read the book I bought yesterday at the convience store"
    "kinou konbini de katta hon wo yonda hito" = "The person who read the book he bought at the convience store" (I didn't actually specify who bought the book, so we can only assume the person who read it also bought it, but maybe not, maybe I bought it and gave it to him).
    etc.

  • User profile image
    wkempf

    Sven Groot wrote:
    

    Perfect machine translation. Currently automatic translation is laughably bad even on the most ideal inputs. Spoken text is often ungrammatical and contains colloquialisms and expressions. A machine translator must have an extreme amount of real-world knowledge to recognize these and map them into another language. We're not even close on this one. Personally I believe that  unless we can create a functioning replica of the human brain the amount of data and the type or reasoning needed to really do this without lots of mistakes will make this impossible. This will take 50-100 years at least, if ever. IMO, of course.



    I'm not dismissing all of your points, but this one stand out for me for some very personal reasons.  See, when I was a Freshman in College we had one of the pioneers of Computer Science who happened to be an alumn of our school come and give a very fascinating class room lecture.  This man was smart to the point of being scary, and had vast knowledge on a lot of subjects.  One subject he was very knowledgeable in because of a handicapped sister who could benefit from the technology was speach recognition.  He was aware of all sorts of research being done on the topic, and he was convinced that we'd have functioning software but it would take 25-50 years.  In fact, one of the commercial uses for it he was convinced would be replacing telephone operators.  We all walked out of the lecture and thought he was nuts for predicting this would be within 25-50 years.

    Four years later, while I was a Senior, AT&T was replacing operators in Fla. with a new computer system that employed speach recognition.

    I now hear predictions like this, and always think to myself "divide the lowest estimate by 5", and you know what?  That often turns out to be an accurate estimate!

  • User profile image
    Sven Groot

    wkempf wrote:
    
    Sven Groot wrote:
    

    Perfect machine translation. Currently automatic translation is laughably bad even on the most ideal inputs. Spoken text is often ungrammatical and contains colloquialisms and expressions. A machine translator must have an extreme amount of real-world knowledge to recognize these and map them into another language. We're not even close on this one. Personally I believe that  unless we can create a functioning replica of the human brain the amount of data and the type or reasoning needed to really do this without lots of mistakes will make this impossible. This will take 50-100 years at least, if ever. IMO, of course.



    I'm not dismissing all of your points, but this one stand out for me for some very personal reasons.  See, when I was a Freshman in College we had one of the pioneers of Computer Science who happened to be an alumn of our school come and give a very fascinating class room lecture.  This man was smart to the point of being scary, and had vast knowledge on a lot of subjects.  One subject he was very knowledgeable in because of a handicapped sister who could benefit from the technology was speach recognition.  He was aware of all sorts of research being done on the topic, and he was convinced that we'd have functioning software but it would take 25-50 years.  In fact, one of the commercial uses for it he was convinced would be replacing telephone operators.  We all walked out of the lecture and thought he was nuts for predicting this would be within 25-50 years.

    Four years later, while I was a Senior, AT&T was replacing operators in Fla. with a new computer system that employed speach recognition.

    But that's speech recognition. Those phone answering computers don't need to understand more than a few built-in phrases, and it works well enough. This isn't even as complicated as dictation, and even dictation is orders of magnitude easier than translation. Recognizing speech is pretty hard. I've never seen a dictation system yet that can infer where to put commas and full stops, you always have to say them explicitly. Still, this I estimate to be possible. We're getting there for speech recognition.

    But translation is just such a completely different problem. I propose that it is not possible to truly accurately translate a text without understanding it. Rules and heuristics just won't cut it. And I've never seen a computer that understands, in the human sense, anything at all. And even understanding the text isn't enough! The complex and subtle differences between the languages must also be understood. I consider myself fluent in English at the level of a native speaker. But I still find it difficult translate a text between Dutch and English.

    I'm not a professional translator, but I do translations on occasion. One of my Japanese teachers said, "the important thing [when translating Japanese to Dutch] is first and foremost that it's correct Dutch; only then comes resemblance to the Japanese original." Because of this, because of differences in nuances, and many other tiny things, I find that making a good translation requires extensive real-world knowledge, about the languages involved, and also about the cultures that use the languages. And it requires creativity, because especially with languages that are as wildly different as Japanese and Dutch, some things have no direct translation, and you have to work around it.

    If you think we can teach a computer to do all that in 10 years, I admire your optimism. I do, however, not share it. Smiley

    wkempf wrote:
    I now hear predictions like this, and always think to myself "divide the lowest estimate by 5", and you know what?  That often turns out to be an accurate estimate!


    Ok, so when's never divided by five? Tongue Out

  • User profile image
    wkempf

    Again, experts in the field working on the subject matter, who had the best chance of predicting the timeline, got the speach recognition one wrong.  By a lot.  Technology increases geometrically, such that things we think are "impossible" today, are mearly difficult in a year, and a reality in 5.  There are so many things that are a reality today, that when I started in this field we were skeptical could even be done, much less in our life times.

    Everything you've said is true.  This is a very non-trivial problem.  Today, we have no idea how to even go about solving the problem.  But a break through in an unrelated area could change all of that in a very short period of time.  It's happened numerous times in the past, and it will happen numerous times in the future.

    I'm not predicting that the "universal translator" will actually happen in the originally predicted time frame.  Heck, like you I've got some amount of skepticism about it being done in our life time.  But experience has taught me to not make wild predictions like that.  If you're prediction has to stretch out more than 15 or so years, you're really just guessing wildly.  You have no real idea how long it will take.  You just know that we have no idea how to do it today.

  • User profile image
    jason818_25​3.33

    Bas wrote:
    I wonder if it's even going to be necessary. I'm suspecting that English will become more and more widespread, and that in the next two decades more and more people will speak it. Especially considering that the internet is becoming more and more accessible to more people. More exposure to the English language results in more people learning to speak English.


    great point. All the more reason to have a translator. a mono lingual world would be so boring.

    Sven Groot wrote:
    

    but this particular one I do not expect to see happening in my lifetime.



    why, how old are you? Big Smile
    You make good points to building a translator. The parts of a translator if you will.
    The problem you speak of with translation may be solved with ether intelligent design algarithms so the program recognizes the different grammar structure. or it needs copious amounts of data storage to handl the many different ways of saying things. here is a small hand held that has canned fraises stored with in it http://www.franklin.com/estore/dictionary/TGA-490/
     
    Also, needed is a nice sounding speech program so when listening on the other side of the phone it doesn’t sound like your talking to a recording.

  • User profile image
    Massif

    Hmm... I think we'll probably have a true artificial intelligence before we have a flawless translator (witness: human translators are very rarely flawless, and a universal translator would have to be flawless if any business was ever going to use it.)

    By which time it would be a moot point.

    But then predicting the future isn't exactly my forte.

  • User profile image
    wkempf

    No reason for it to be flawless.  Just look at the universal translator in Star Trek, for instance.  It's highly flawed.  Very bad translations have been the basis of several story lines, both on television and in the books.  Despite that, it's a heavily used technology.

    When those speach recognition programs were rolled out and replaced operators, they were hardly perfect.  Only specific responses would be recognized, and then not 100% of the time.  In fact, failure rates were frequent enough that some customers would complain loudly about it.  Despite that, AT&T still rolled them out and never abandoned them.  Today, such "operators" are common place in many businesses, and we don't think anything of them, or get more than midly annoyed when we have to repeat ourselves numerous times in order to be recognized.

    No, perfection isn't required.  Just look at how many manuals are translated the hard way, and botch something badly enough to cause laughs and other such reactions.  We understand the source of the problem, and generally ignore gaffs in translations.

  • User profile image
    Shining Arcanine

    Soviut wrote:
    Full-on Star Trek.

    I must admit, it would be very impressive.  If it could actually resynthesize your voice formants to make it sounds like you, yourself,  were speaking another language, that would be even better. That in itself would be a huge leap forward in voice analysis and synthesis.


    It has already been done by Scansoft in that they resynthesized Bill Clinton's voice to produce an audio track of him giving his last will and testament:

    http://www.theinquirer.net/?article=10109

    Using their technology to resynthesize someone's voice so they are speaking another language should not be that difficult, although I would expect getting the proper accent into the resynthesis to be an issue.

    Accurate mathematical modeling of human vocal cords would probably have to be done, although Scansoft might have already done it in order to resynthesize Bill Clinton's voice.

    Come to think of it, Microsoft should probably buy Scansoft before Google does.

    Sven Groot wrote:
    

    I don't think this will happen that quickly.

    In general, such a device would require the following, in order of increasing difficulty:

    1. Perfect (or at least not-annoying) voice synthesis. This is possible with today's technology (even though MS's TTS tech in Windows would lead you to believe otherwise), so it's not a problem.
    2. Perfect speech recognition. It may not misrecognise even a single word or you might greatly insult whoever you're speaking to. It must function perfectly even in less-than-ideal circumstances (e.g. a person speaking on a cell-phone on the train). Current technology does not allow this, but I could see it benig developed in the next decade or two.
    3. Perfect machine translation. Currently automatic translation is laughably bad even on the most ideal inputs. Spoken text is often ungrammatical and contains colloquialisms and expressions. A machine translator must have an extreme amount of real-world knowledge to recognize these and map them into another language. We're not even close on this one. Personally I believe that  unless we can create a functioning replica of the human brain the amount of data and the type or reasoning needed to really do this without lots of mistakes will make this impossible. This will take 50-100 years at least, if ever. IMO, of course.
    4. Knowledge of the future. You want real-time translation. Unfortunately, not all languages use the same word order. In Japanese, verbs always come at the end of the sentence. I can make a huge compound sentence, and the verb of the main sentence would be all the way at the end. In English, this verb is one of the first things in a sentence. So at the very least, the translator must wait until the sentence ends before it can translate. And determining where a sentence ends is not always an easy thing to do in colloquial speech.

    I'm in general very optimistic about the human capacity for solving impossible problems, but this particular one I do not expect to see happening in my lifetime.

    EDIT: To demonstrate the difficulty of point four, consider the following sentence in Japanese:
    "kinou konbini de katta."
    It means "I bought it yesterday at the convience store". What "it" is isn't specified; the sentence doesn't have an object, which is allowed since Japanese is highly contextual. "Katta" is the verb here, the past tense of "kau", to buy (tranlated word-for-word the sentence reads "yesterday convenience store at bought").
    To make a compound sentence in Japanese, all I have to do is put a noun at the end of an existing sentence:
    "kinou konbini de katta hon."
    I just added "hon", which means book. Now the sentence is "The book I bought yesterday at the convenience store". So even if the translator heard "katta" it cannot safely assume I'm done with the sentence. And any noun I say after "katta" must come before the verb in the English translated text.

    Let's continue that:
    "kinou konbini de katta hon wo yonda" = "I read the book I bought yesterday at the convience store"
    "kinou konbini de katta hon wo yonda hito" = "The person who read the book he bought at the convience store" (I didn't actually specify who bought the book, so we can only assume the person who read it also bought it, but maybe not, maybe I bought it and gave it to him).
    etc.



    Regarding point 4, I did say near-real time, as I realized that the word order would pose an issue. I would imagine that when this is done, it would be similar to having a human translator today, although somewhat faster, as the computer could listen and speak at the same time while a human translator would have to wait before the person stops speaking to translate. I would not expect to have more than a one to two sentence lag time at most.

    Bas wrote:
    I wonder if it's even going to be necessary. I'm suspecting that English will become more and more widespread, and that in the next two decades more and more people will speak it. Especially considering that the internet is becoming more and more accessible to more people. More exposure to the English language results in more people learning to speak English.

    If, eventually, we can all communicate to each other using the same language, then there's not such a big need for a translating service anymore, meaning less research, meaning no universal translator.


    Many people have been talking about Spanish supplanting English as the dominant language in the United States. Although there would be benefits to everyone speaking fluent English, I doubt that it will happen.

    jason818_253.33 wrote:
    
    Bas wrote:
    I wonder if it's even going to be necessary. I'm suspecting that English will become more and more widespread, and that in the next two decades more and more people will speak it. Especially considering that the internet is becoming more and more accessible to more people. More exposure to the English language results in more people learning to speak English.


    great point. All the more reason to have a translator. a mono lingual world would be so boring.

    Sven Groot wrote:
    

    but this particular one I do not expect to see happening in my lifetime.



    why, how old are you?
    You make good points to building a translator. The parts of a translator if you will.
    The problem you speak of with translation may be solved with ether intelligent design algarithms so the program recognizes the different grammar structure. or it needs copious amounts of data storage to handl the many different ways of saying things. here is a small hand held that has canned fraises stored with in it http://www.franklin.com/estore/dictionary/TGA-490/
     
    Also, needed is a nice sounding speech program so when listening on the other side of the phone it doesn’t sound like your talking to a recording.


    Sounding like a machine could be a benefit if you receive telemarketing calls. I know a professor that had the misfortune of losing his vocal cords in some incident; he now uses a device to speak and sounds like Professor Stephen Hawking. He said that telemarketers hang up on him as they think he is a machine designed to take telemarketing calls. If a universal translator is developed, a robotic voice add-on might be one of the first add-ons developed.

    By the way, I notice no one from Microsoft has commented. Does this mean that Microsoft is already working on this?

  • User profile image
    brian.​shapiro

    Shining Arcanine wrote:
    I think that in the next decade or two, the next major breakthrough will be the development of the Universal Translator, most likely by Google.

    What is Microsoft doing in this area?

    Edit: By the way, when I say Universal Translator, I am referring to a computer that will sit in the telephone line and translate phone calls in near-real time, making it sound like the other person is speaking your language to you and like you are speaking the other person's language to the other person.


    Microsoft has been doing a lot of research on both natural language parsing and speech recognition, so I doubt Google will get ahead of Microsoft on this, unless they pull of hiring ex-Microsoft employees or there are new people in the field who have the answers that the veterans don't and support Google over Microsoft.

  • User profile image
    evildictait​or

    wkempf wrote:
    

    I now hear predictions like this, and always think to myself "divide the lowest estimate by 5", and you know what?  That often turns out to be an accurate estimate!



    Funny, I always multiply by three, but maybe that's because I listen to optimisists and you listen to pessimists?

  • User profile image
    evildictait​or

    brian.shapiro wrote:
    
    Shining Arcanine wrote:
    I think that in the next decade or two, the next major breakthrough will be the development of the Universal Translator, most likely by Google.

    What is Microsoft doing in this area?

    Edit: By the way, when I say Universal Translator, I am referring to a computer that will sit in the telephone line and translate phone calls in near-real time, making it sound like the other person is speaking your language to you and like you are speaking the other person's language to the other person.


    Microsoft has been doing a lot of research on both natural language parsing and speech recognition, so I doubt Google will get ahead of Microsoft on this, unless they pull of hiring ex-Microsoft employees or there are new people in the field who have the answers that the veterans don't and support Google over Microsoft.


    I suspect that it will be developed by universities in the States and will be joined with the existing algorithms at Microsoft or Google for a price. Speech recognition/translation is hot PhD stuff these days.

  • User profile image
    wkempf

    evildictaitor wrote:
    
    wkempf wrote:
    

    I now hear predictions like this, and always think to myself "divide the lowest estimate by 5", and you know what?  That often turns out to be an accurate estimate!



    Funny, I always multiply by three, but maybe that's because I listen to optimisists and you listen to pessimists?


    For me, it depends on the timeline.  Less than 15 years in the prediction, and multiply by 3, because in the near time people seem to have a habit of underestimating.  More than 15 years, divide by 5, because they're really "guessing wildly" and it's likely that the natural progression of technology will cause unforseen break throughs.

    Obviously, though, this isn't an exact science Wink.

Conversation locked

This conversation has been locked by the site admins. No new comments can be made.