The following is a transcript of an interview I did with an Artificial Intelligence expert at Macalester College, Susan Fox, to get some details on how computers handle translation.
 | How would the best machine translation software operate? What are its parameters? Would
it be an amalgam of Internet servers? Translation is an incredibly difficult task, if it is to be done well. It is very hard for people to do, much less a machine. So I believe that full-scale, high-quality translation systems are many many years in the future. The most promising translation systems I have seen take a particular language and convert it to an internal "language-independent" representation, and can then produce reasonable translations into any of a number of target languages. The difficulty of this is in designing an expressive internal representation. Good translation requires an incredible amount of knowledge about the world in general and the languages in particular. That's the real bottleneck for doing *general* translation. It might help to have many different machines working together, each with its own particular information. However, in some specific subject areas machine translation is a more tractable task, and is actually being used. |
| Are certain types of printed material, like newspapers or manuals, easier for computers
to translate? That is, would poetry and stories be more difficult because of the allusions
and rhyming? News articles (though not features or commentaries), manuals, legal documents, business documents: all those are classes of text that are easier to translate because they follow particular patterns, use a more restricted set of words, and tend to avoid metaphor. The question of translating poetry or literature is a much more difficult one. There is more to the problem than just allusions and rhyming: there is the use of words to set a tone, the rhythmic structure as well as rhyming structure, metaphor, idioms. A machine translation system can translate the words, and simple phrases, but couldn't capture the tone, the rhythm, might miss metaphors. But then, translating literature is a very highly skilled job for people, and we've all probably run across *bad* translations. |
| What are the biggest problems facing an accurate translation from a machine? World knowledge, analogies, and metaphors, I would say. When we write, especially poems and stories, we often leave things out: between one sentence and the next we expect our readers to fill in details that make sense out of what we actually have written. Machines can't fill in those details right now, and that may lead to a misunderstanding of what follows, and a mistranslation. (This is less of a problem for non-literary texts, which may be why some machine translation systems are in use in those contexts.) |
| What common features would be included in an AI professional's definition of
"accurate translation" when dealing with machine translation? First and foremost, getting the literal sense of the original text *at the sentence level*. It's not enough to literally translate each word... the sentences have to work together. Correctly translating common idioms and phrases would be good. Other features such as the tone of the writing would be nice, but are out of our reach at the moment. |
| Where is the future of machine translation? Will it always remain an assisting tool, and
not an independent service? For quite a while I believe it will be an assisting tool, even in domains like legal or business writing where it can currently do a reasonable job of translation. Reasonable is not really good enough, right? Restrictive domains will be the first place where independent machine translation will become available. The general translation problem is what we in AI would call "AI-complete"... that means that to get an independent system that would do machine translation it would have to be an *intelligent* system: you'd have to solve the whole artificial intelligence problem in order to do it. |
| How do you see machine translation affecting the study of literary translation? In other
words, could you see machines replacing humans someday? Do you think the increase in the
world economy will increase the demand for more research into machine translation? I think I've already said that I think machines replacing people for literary translation is many many years in the future, if it happens at all. However, I think that there will be (and in fact has been) more and more research into machine translation for more "mundane" translation tasks. |