I’m thinking of a translation engine for internet message boards and such, where a Russian or Chinese or French etc user would type something in their language that would be translated (automatically) into contextually correct English and vice versa. Is this even possible in real time? Is this possible at all without human intervention?
I tried out www.babelfish.altavista.com last weekend and was quite impressed.
Currently it is definitely not possible, and it will probably stay that way for many years. For the task you gave those often questionable translations offered by e.g. Babelfish are the best thing available. You can get better results by restricting the domain, e.g. translate only business appointments, intructions for consumer electronics… but useful translation of unrestricted text is incredibly hard, even more so if the style is informal. From a certain point on you require something awfully close to general artificial intelligence and an enormous amount of knowledge about the world. Even if we don’t achieve a program that “truly understands” language (what does that mean anyway?), we might one day have systems that work reasonably well for many everyday applications.
I’m really not. Tools like babelfish are useful when you want to know the general meaning of a relatively long (because if it’s too short, a single poor translation can change completely the meaning) text written in a language you don’t understand.
I just used it for ** kellner **'s post above, for instance, and I got what he meant (I didn’t read previously the english version).
On the other hand, if it’s important that the precise content of your message will be correctly conveyed, you shouldn’t rely on such a tool. Especially not if the message is somehow technical.
I’m thinking, for instance, about people from another board sending e-mails to foreign hotels to make sure they’ll get the kind of room they want. Since they want to be nice/ are worried that the hotel staff maybe won’t understand english, they send their message translated by babelfish.
The result is generally impossible to understand. Sometimes, the most logical interpretation of the translation is something completely different from what they meant. Or even the contrary, depending on how the sentence is organized.
I think a computer program capable of that, would also be capable of passing the Turing test.
This is doubtless true. Good, general translation is AI-complete, because it presupposes that the program is capable of taking all of the context into account and divining the meaning of the text, not just the grammar and the definitions of all of the words and recognized idioms.
It would be nice if everyone learned a high-level language that could be autotranslated into any other human language, a symbolic language capable of expressing everything every human language can express in a common, uniform way. But that’s just the naive engineer in me speaking.
I agree; a computer program capable of perfect (or even just very accurate) translation would have to actually understand the meaning of the input in order to render a translation that preserves the idiomatic as well as literal meaning.
The task would be somewhat simplified if:
- The rules of grammar were absolute
- The rules of grammar were always observed
- A comprehensive international database of idiom existed
- People didn’t have an annoying habit of making things up as they go along.
- The context of the passage was always locally defined
As it stands though, language is rather organic and even competently bilingual humans can have problems, especially when the text to be translated is informal, emotional, laced with metaphor and/or dependent on assumed context.
I would go so far as to say its impossible seeing as how even human translators have occasional problems and misunderstandings. The problem is so ill-specified that there is no solution.
One interesting advance was made that I have read about in the past year, and that’s statistical-based language translation. You know the heuristics mechanism that runs the good spam filters lately? Imagine that on a huge, huge scale. It’s always been said that translating needs some type of intelligence, but this kind of demonstrates that it may not be necessary. Context, you say? Given a large enough sample, you can probably gleen the context statistically. I wish I remembered where I read about this…
Very large-scale heuristics may be intelligence.
Obviously, it’s possible to evolve a system that can do translations, but designing it from the ground up might be a bit tough. There’s still a lot we don’t know.
Trouble is, there’s a lot of context in communication that doesn’t ever show up in the verbiage. Where are you? Who are you talking to? Are you playing a game, or deadly serious? Is it day or night? Has the incessant downpour finally broken the the levee?
If it’s ever done, I bet it still couldn’t translate song lyrics to make them rhyme and fit the nuance of the musical notes.