It would be really cool if Doulingo had its very own Translator such as Google Translate, as it would be extremely convenient for Users...
Also, as great as Google Translate may be, its obviously not perfect, and I feel that a community of people who can speak different languages would be able to make a Translator superior to that of Google's...
It would take Time and Effort, as being Open Sourced, Duolingo Translate would probably recieve alot of Spam and Innacurate Translations, but it would be worthwhile...
In the end, of course, its only a suggestion for the Convenince of Users...
DL had a feature, Immersion, that was a lot better than Google Translate (in fact the feature provided a link to Google Translate)... Immersion was not deemed 'profitable' and it was turned off a little more than 1 year ago. It was the most debated issue that I have seen in the history of DL and the most downvoted announcement... Happy New Year and keep owling!
I feel that a community of people who can speak different languages would be able to make a Translator superior to that of Google's...
Do you understand how machine translation works? There is no chance that a bunch of amateurs could surpass one of the largest and best-funded companies in the world at its own game. Translation algorithms are improved by usage data, and Google has more data on language use than anyone else. Within a decade Google will probably be indistinguishable from a human translator, except perhaps for poetry (but there'll be an algorithm to translate even that eventually).
Within a decade Google will probably be indistinguishable from a human translator, except perhaps for poetry (but there'll be an algorithm to translate even that eventually).
Indistinguishable does not necessarily mean better than a human translator. You mention poetry rightly as being more difficult, but I think it holds true for both prose and poetry that translating a large body of text is not only linguistics, but an art form in itself.
Now I don't want to be the romantically-minded guy that thinks humans can do everything better than computers; certainly not. But when it comes to translating a novel, I am of the opinion that the translator has to have a knowledge of multiple factors apart from the language, such as the cultural background or the author's style. A text can be translated in so many different ways, and one translation is not necessarily better than the other. A translator puts something of his own style into his translation; the computer would also be creating his own style, which doesn't have to be superior to the human one.
I can disagree with nothing you say. I didn't say 'better', I said 'indistinguishable'. 'Better' implies some sort of aesthetic or moral judgement, whereas an algorithm can only make a logical and probabilistic judgement. What matters is whether we can tell them apart. (In reality, of course, our own aesthetic and moral judgements are governed by these same computational principles.)
In terms of an author's style, a computer has the great advantage of being able to scan through the whole corpus of said author's works in the blink of an eye. All subtleties of style are eventually reducible to mathematics. In terms of cultural understanding, computers aren't anywhere near capable of that, yet; but within a decade, it is eminently possible, given how fast the field is moving (which many people greatly underestimate). It is quite possible that we shall have a general AI within a decade, although I'd hedge my bets and say 15 years. Every judgement can be reduced to a formula that a machine can replicate—the judgement as to what is a better translation in the workings of a machine is really no different to that in the brain of a human, given sufficient complexity; the computer would indeed be creating its own style.
True, I am not doubting the power of the AI, and the mere data it has to its disposition is a major advantage over any human being. But I'm wondering to what an extent the computer will be able to translate freely, or should I say in a literary manner.
A human translator might look at a sentence, translate it word for word, then think it sounds clunky or off in the target language, and choose to paraphrase the meaning. Such way of translating is not as exact, true, but in order to provide a readable text, it's sometimes necessary. Mainly in literature, but possibly in other texts as well.
I guess the point I'm trying to make, is that maybe in the future, a computer-translated text might not be indistinguishable for most people, but can come across as clunky, so that a trained eye might indeed suspect a text is not translated by a human.
But who knows, maybe AI will eventually learn how to write literature itself - I have heard that they already made a computer generate a chapter for a Harry Potter book. That was gibberish for the most part, but maybe we will see more serious attempts in the future.
A human translator might look at a sentence, translate it word for word, then think it sounds clunky or off in the target language, and choose to paraphrase the meaning.
Noticing that something sounds clunky is just another neural algorithim in our own brains, and one which we learn naturally through exposure to comparatively limited data, which we filter and process in undoubtedly very sophisticated ways. There's no reason why an A.I. couldn't master the same skill in principle.
It won't, to begin with, at least, have our tremendously-refined filtering systems, however, so it will likely consume millions of examples before it comes up with a comparable rule-based shortcut to identifying (at least some of the qualities of) 'clunkiness'. Doing this would have taken years a few years ago, but would probably take an hour or so today at Google HQ, and the speed at which these sorts of calculations can be performed continues to increase exponentially.
This is why I really do think that machine translation will become indistinguishable in every possible way within a relatively short time period; the possible complexity of any given language is mathematically limited, and computers will be able to greatly exceed our own brains' computational capacity many times over before too long (not to mention that they can process information a thousand million times faster). Unless people stop trying to do it, it is inevitable.
As I understand it, the new, current Google Translate (which is clearly much better than the old) relies on neural networks. These rely on masses of statistical relationships between words. As you say, this hold great promise for eliminating "clunkiness," as that is only identifiable as a pattern that does not fit with the great number of examples a reader has encountered previously (and which a computer can sift through an even greater number of every so quickly). However, what this method does not yield is a "structural model" of the language being processed. The computer doesn't understand the language it's chugging on; it's just chugging on it.
So when one language demands a disambiguation based on certain kinds of context, it would seem that neural networks are likely to fail (i.e. not "guess" right sometimes). Between Russian and Ukrainian and other languages, imagine translating "he waved at them with his right [рука]." Statistically, I bet a neural network gives you "hand" since it sees "wave," and in the mass of prior data the stronger association will be with "hand" over "arm." But what if this particular "he" lost his right hand several chapters back? I don't think neural networks have the ability to take that into account. A literary translator certainly should.
Multiply such an example by all the non-indicated subjects in a Japanese or "missing" evidentiality information translating from a language that doesn't mark it grammatically to one that does (a class which exhibits a certain correlation to having a likely lack of a large digitized corpus to train a system on and a lack of economic incentives on the order of an English, Spanish, or Japanese to attract Google's attention) and I think humanity will remain a nose ahead for a good while yet.
Google Translate for Russian, despite the big steps forward, clearly does not understand the instrumental case. This puts in pretty clear relief the drawbacks of "neural network onlyism"; all that massive corpus it's working on and it can't reliably render (I don't know that I've seen it succeed a single time) the meaning that gives the instrumental case its name. I think anybody unfamiliar with the matter would have assumed they'd just throw in some grammatical checks besides the statistics, but it seems there must be very significant reasons for not doing so.
I have a good understanding of how translations (including machine) work. Regarding human vs machine translations there are good arguments on both sides. Besides technical factors, there are several other facts that need to be kept in mind:
- political/economic factors that have nothing to do with technical innovations per se.
- Google is one of the investors behind Duolingo.
Personally, I have doubts that Google will come up with a tool that will be better than the Immersion feature that was available through DL. Yes, they have the money and the knowledge to do it (so, I would be more than happy if they do!-). Hopefully, some other player will see the value and bring it on...
would be able to make a Translator superior to that of Google's...
As much flaws as Google Translate might have a crowdsourced Duolingo translator would have next to zero chance on bettering it.
That said perhaps an integrated word translator with the current database would be nice. Doubt it's on their plans though.