Can We Translate English To English?(rjlipton.wordpress.com)
rjlipton.wordpress.com
Can We Translate English To English?
http://rjlipton.wordpress.com/2010/11/12/can-we-translate-english-to-english/
12 comments
Why is that idea disgusting? When people know what they say will get translated, they tend to think in terms of the end result. This would promote the use of proper English, because that's how it would be read and written. When I code in a programming language, I am constantly thinking about what the language constructs mean, not what they literally say. The disjointed net-speak would just become a form of shorthand (http://en.wikipedia.org/wiki/Shorthand).
Perhaps it depends on the person? My "disgust," or worry, in my original post, came from the vision of someone internally or unconsciously deciding, "The computer will make my communication effective for me, so I needn't bother learning to communicate effectively myself." Without an empirical understanding of how to communicate, a person would also lack the trappings of that learning process, and, I think, would be forever marred as a result.
When you code, mustn't you be aware of what the code constructs literally say, because that is how the compiler will interpret then?
When you code, mustn't you be aware of what the code constructs literally say, because that is how the compiler will interpret then?
When I say literally, I mean that a for loop is constructed of the English word "for." I don't think of it like that. I think, "Okay, I need to iterate over the elements of this sequence, so I'll use a for loop." The code is literally just text, but I must understand their semantics to know what will happen.
Similarly, if someone was using what you proposed, they would have to keep in mind that the other person would see the proper English, not the net-speak. So they would - I think - end up thinking in proper English first, then coding the proper English using net-speak. They must think in terms of the end result, in terms of what the person they're communicating with will read. For that reason, I disagree with your assessment that people would be able to think "I don't need to bother learning to communicate effectively."
Similarly, if someone was using what you proposed, they would have to keep in mind that the other person would see the proper English, not the net-speak. So they would - I think - end up thinking in proper English first, then coding the proper English using net-speak. They must think in terms of the end result, in terms of what the person they're communicating with will read. For that reason, I disagree with your assessment that people would be able to think "I don't need to bother learning to communicate effectively."
Ah, I see. You're saying that the shorthand exists only for brevity of input; when a tween types "lol ur rite," they first think of what they mean ("That's funny! You're right!"), and only then do they translate it into the corresponding shorthand so they can type it faster. That's a very good point. I'll have to ruminate on that for a while.
Edit: It's worth noting that something about it still feels horribly wrong to me, but I can't seem to find the right words to express it. So I will concede... for now! :-)
Edit: It's worth noting that something about it still feels horribly wrong to me, but I can't seem to find the right words to express it. So I will concede... for now! :-)
Yes, that's exactly what I'm saying. Perhaps your reservation comes from the fact that now there is no translations, so you associate net-speak with what it is now: stilted communication optimized for brevity of input time and length.
I think my reservation may come more from what I'm saying in this comment:
http://news.ycombinator.com/item?id=1898053
Even if a computer could translate netspeak to "Proper English," it would still be computer-translated "Proper English." The idea of allowing a computer to choose which words to use, or where punctuation should go, or anything like that, is unsettling to me. The translation model works for programming languages because compilers have no emotions, but any message that starts out in shorthand with the intention of being read by a human is intrinsically hamstrung.
In fact, this can be true for programming languages, too! C is like short-hand; the compiler is what turns our netspeak (C) into "Proper English" (assembly/machine code). It usually works, but sometimes you just have to dip into assembly to really express what you mean.
http://news.ycombinator.com/item?id=1898053
Even if a computer could translate netspeak to "Proper English," it would still be computer-translated "Proper English." The idea of allowing a computer to choose which words to use, or where punctuation should go, or anything like that, is unsettling to me. The translation model works for programming languages because compilers have no emotions, but any message that starts out in shorthand with the intention of being read by a human is intrinsically hamstrung.
In fact, this can be true for programming languages, too! C is like short-hand; the compiler is what turns our netspeak (C) into "Proper English" (assembly/machine code). It usually works, but sometimes you just have to dip into assembly to really express what you mean.
You do not need to worry. No computer algorithm can actually add information not present in the original writing. An algorithm can chew on information given and perhaps make it more presentable or accessible, but absolutely nothing can create information from whole cloth; it's fundamentally impossible. If you think you see it happening, look for the curtain to pull back.
uh, what about the information existing in the probability distribution of the english language...
... what about it? Your question is so fuzzy I don't have a clue what you're trying to get at.
maybe i misunderstood your original comment. can you give an example of the sort of task you argue an algorithm is incapable of completing?
Do you know information theory, by which I mean the actual mathematical study of information? The definition of information is a wild and woolly beast; if you understand the definition you already know what I mean, if you don't you'll argue endlessly without ever having understood what you're arguing against.
But what it basically boils down to is a counting argument; if you are a human and you have ten thoughts of various kinds that in your illiteracy all get "translated" (putting it politely) to "lolz that suckz", no computer algorithm, or indeed anything else, can translate that back out to the original ten thoughts. It could pick one of them and use that, and that chosen though may be extremely eloquently expressed, but that's all it can do. A fixed number of inputs can only go to a fixed number of outputs; there's no magical process whereby anything, not even a human, can take truly useless semantic garbage text and turn that into accurate informative well-written text without loss. There's radically more well-formed, coherent thoughts than there are fuzzy thoughts and there simply isn't any sort of mapping at all from fuzzy to coherent. If someone wishes to communicate precisely, they will always have to put the effort into it, there is no magical shortcut. There's no need to worry about the effects of the magical shortcut on people or on society, no such thing exists.
In fact this is actually just another form of the following argument in heavy disguise: http://matt.might.net/articles/why-infinite-or-guaranteed-fi... I'm making the same argument in the other direction; that argues there is no such thing as a "guaranteed compressor", I'm arguing in the other direction, that there can not exist a decompresser than can take some smallish set of gibberish and map it to the much richer set of interesting thoughts. Information theory is fun for being able to make arguments about this hypothetical translator and a hypothetical perfect decompresser the same argument in the end.
The obvious counterargument is that a human might be able to use more context and extract more details to get a richer sense of the details; a stream of poorly written text can still mean more than one bad tweet. The problem is that the source information is everything that a human or anything else uses to make its decisions. Anything a human could use, a program could in principle use as well, even if we couldn't write it today. Differences between humans and computers are only issues of how good the translator may be; the theoretical limit on both is unaffected by those differences.
But what it basically boils down to is a counting argument; if you are a human and you have ten thoughts of various kinds that in your illiteracy all get "translated" (putting it politely) to "lolz that suckz", no computer algorithm, or indeed anything else, can translate that back out to the original ten thoughts. It could pick one of them and use that, and that chosen though may be extremely eloquently expressed, but that's all it can do. A fixed number of inputs can only go to a fixed number of outputs; there's no magical process whereby anything, not even a human, can take truly useless semantic garbage text and turn that into accurate informative well-written text without loss. There's radically more well-formed, coherent thoughts than there are fuzzy thoughts and there simply isn't any sort of mapping at all from fuzzy to coherent. If someone wishes to communicate precisely, they will always have to put the effort into it, there is no magical shortcut. There's no need to worry about the effects of the magical shortcut on people or on society, no such thing exists.
In fact this is actually just another form of the following argument in heavy disguise: http://matt.might.net/articles/why-infinite-or-guaranteed-fi... I'm making the same argument in the other direction; that argues there is no such thing as a "guaranteed compressor", I'm arguing in the other direction, that there can not exist a decompresser than can take some smallish set of gibberish and map it to the much richer set of interesting thoughts. Information theory is fun for being able to make arguments about this hypothetical translator and a hypothetical perfect decompresser the same argument in the end.
The obvious counterargument is that a human might be able to use more context and extract more details to get a richer sense of the details; a stream of poorly written text can still mean more than one bad tweet. The problem is that the source information is everything that a human or anything else uses to make its decisions. Anything a human could use, a program could in principle use as well, even if we couldn't write it today. Differences between humans and computers are only issues of how good the translator may be; the theoretical limit on both is unaffected by those differences.
the reason there isn't a guaranteed compressor is that a legitimate compressor has to be injective, and there isn't an injective function from, say, {0,1}^[n] to {0,1}^[m] where m<n. in contrast, there are many guaranteed "decompressors," if by a decompressor you mean an invertible function that expands the length of a string.
you seem to contradict yourself a bit, first you agree that an algorithm could take "lolz that suckz" and produce "one of" the original thoughts that got transcribed as it, then you say there "simply isn't any sort of mapping" from "fuzzy thoughts" (sentences produced by transcribing thoughts?) to coherent ones. but obviously there are many such mappings, it's just that they aren't guaranteed to produce the same thought the author of the "fuzzy thought" had when he wrote down what he was thinking.
now, what is the actual cause for fear? is it the idea that we might someday be able to communicate without taking too much effort (what you say is impossible), or that people with nothing going on between their ears will be able to produce computer-mediated output that is indistinguishable from coherent thought? i think the latter is more cause for concern/more interesting but i dont think anything youve said really addresses it. tbh i also dont think your brief discussion of information theory does much in the way of arguing against the former either. (your argument could be adapted to show that morse code is impossible, because messages are shorter in morse code than they would be in a straightforward ternary ascii encoding. but actually, it's just that the messages we typically wish to send are generally shorter.)
you seem to contradict yourself a bit, first you agree that an algorithm could take "lolz that suckz" and produce "one of" the original thoughts that got transcribed as it, then you say there "simply isn't any sort of mapping" from "fuzzy thoughts" (sentences produced by transcribing thoughts?) to coherent ones. but obviously there are many such mappings, it's just that they aren't guaranteed to produce the same thought the author of the "fuzzy thought" had when he wrote down what he was thinking.
now, what is the actual cause for fear? is it the idea that we might someday be able to communicate without taking too much effort (what you say is impossible), or that people with nothing going on between their ears will be able to produce computer-mediated output that is indistinguishable from coherent thought? i think the latter is more cause for concern/more interesting but i dont think anything youve said really addresses it. tbh i also dont think your brief discussion of information theory does much in the way of arguing against the former either. (your argument could be adapted to show that morse code is impossible, because messages are shorter in morse code than they would be in a straightforward ternary ascii encoding. but actually, it's just that the messages we typically wish to send are generally shorter.)
I think it would be similar to the phenomenon of Chinese people forgetting how to write characters that they can read and type. Perhaps not disgusting, but something is lost in the process.
If there was a program to go from legalese to lay that was as accurate as a lawyer then it would change the lawyer's profession forever.
Honestly, the more I think about this, the more I think that we'll have an English-to-English translating algorithm as soon as we have a "classical-masterpiece"-writing algorithm or a "painting-so-beautiful-it-makes-you-cry"-creating algorithm; that is, probably never. I'm no expert writer, but I know a few, and it seems to me that excellent writing style is more than reading Strunk & White cover-to-cover while reciting mnemonics for choosing between "who" and "whom." It's also knowing when to break or bend the rules, how to break or bend the rules, and why to break or bend the rules. That's the kind of spontaneous creativity that I don't think computers are even close to emulating, much less learning.
Have you read http://www.unc.edu/~mumukshu/gandhi/gandhi/hofstadter.htm ?
As Wittgenstein said, "language is a form of life" -- it is not just a system of encoding.
But something like Photoshop filters for text style might be possible. You can imagine the cheesy use of, say, gothic-romanticising, like overdone HDR tone mapping. But there are useful image manipulations too, when well deployed, and there must be similar potential for text, maybe more so given its more structured form . . .
But something like Photoshop filters for text style might be possible. You can imagine the cheesy use of, say, gothic-romanticising, like overdone HDR tone mapping. But there are useful image manipulations too, when well deployed, and there must be similar potential for text, maybe more so given its more structured form . . .
"This is one of the best written and slickest proposals I have ever read. Do NOT fund this work under any circumstances."- NSF grant reviewer
That's eminently quotable.
edit: "rant reviewer" => "grant reviewer"
edit2: proposal => proposals (even though it's proposal in the link, putting proposal[sic] just looks odd).
That's eminently quotable.
edit: "rant reviewer" => "grant reviewer"
edit2: proposal => proposals (even though it's proposal in the link, putting proposal[sic] just looks odd).
One thing some people do is take an article on the net, use Google translate to translate it into some other language and then back to English, and put it as an article on their website to attract money from Google ads. This counts as better, for a suitably restricted definition of "better" -- it makes approximate sense and is different enough from the original that Google doesn't spot it.
The idea was that perhaps they could help improve the quality of this or any other written piece.
I'd pay $10 without thinking twice for a chrome extension that simply ensured the correct versions of there, their and they're were replaced based on context when they were misused. Hmm, I wonder how hard that would be.
I'd pay $10 without thinking twice for a chrome extension that simply ensured the correct versions of there, their and they're were replaced based on context when they were misused. Hmm, I wonder how hard that would be.
A company called White Smoke has an application that will take non-native English and fix it:
http://www.whitesmoke.com/products.html
I've often thought the same, and set out to build song writing software with ML principles. I ended up with iRhyme, an iPhone app that analyzed a huge corpus of existing song lyrics to make a good song writing oriented rhyming dictionary. That worked out well but improving general prose seems difficult. I think domain specific applications could be good stepping stones.
Isn't this idea very similar to the grammar suggestions you can get from Word? The quality of those suggests it is a very hard problem.
Obligatory: http://www.translationparty.com/
Some classic works are very difficult to read in their native language. I've had Japanese friends say that Genji Motoganari - probably the most famous Japanese literature of all time - is basically unreadable in Japanese, and they didn't "get it" until they read it in English.
I feel the same about some older English works. For a long time, I've thought it'd be cool to have one talented author translate an older English book into modern Japanese, and a second author who hadn't read the original to translate it back to English. I bet it'd be more readable.
I feel the same about some older English works. For a long time, I've thought it'd be cool to have one talented author translate an older English book into modern Japanese, and a second author who hadn't read the original to translate it back to English. I bet it'd be more readable.
I read part of Genji Monogatari during my Japanese studies as an undergrad. It's nearly 1,000 years old and it's absolutely not the same language that's spoken in modern Japan. How much of anything written that long ago could a normal native English speaker understand?
It's just not a reasonable thing to expect from the general populace.
It's just not a reasonable thing to expect from the general populace.
I think it's possible, but you would need peices to draw off of: Google uses U.N. translations to train their algorithms, but how many documents do we have translating legalese into good english, or poor writing into great writing? Probably not more than a couple hundred.
We have documents that have gone through an editing process, with changes tracked from draft through to published article.
I'm thinking particularly here of the process inside, say, a news agency. Many of the changes made by sub-editors are fairly mechanical modifications, to conform to a standard style. They might well be sufficiently repetitive to train a 'translator'.
You could also take a corpus of 'good' and a corpus of 'bad' writing, and compare the features in each. sentence length, constructions used. Origin of words would be an interesting one ('never use a Latinate word where an Anglo-Saxon one will do').
Unfortunately, most of the difference between good and bad writing is in the ordering of ideas; you can't really deal with that on statistical grounds.
I'm thinking particularly here of the process inside, say, a news agency. Many of the changes made by sub-editors are fairly mechanical modifications, to conform to a standard style. They might well be sufficiently repetitive to train a 'translator'.
You could also take a corpus of 'good' and a corpus of 'bad' writing, and compare the features in each. sentence length, constructions used. Origin of words would be an interesting one ('never use a Latinate word where an Anglo-Saxon one will do').
Unfortunately, most of the difference between good and bad writing is in the ordering of ideas; you can't really deal with that on statistical grounds.
The developer in me thought, "Oh man. That would be effing sweet."