One of the potential applications of computers that people have always found intriguing is the automatic translation of natural (human) languages. It is easy to understand the appeal of the idea; I certainly would love to be able to communicate easily with any person on Earth. Yet, despite some serious efforts, for many years computer translations were mostly a joke. There is a classic story, quite possibly apocryphal, of the program that translated the English phrase, “Out of sight, out of mind”, into the Russian equivalent of “Invisible insanity”; going the other direction, it translated the verse from the Bible, “The spirit is willing, but the flesh is weak”, to “The vodka is good, but the meat is rotten.”
Of course, to be fair to the machines, translation can be a tricky business. We have all probably puzzled over the assembly instructions for something, as translated from the original Chinese, or have encountered some decidedly odd variants of English in various places. I remember a sign in my room in a small German hotel, which requested, “Please not to smoke while being in bed.” This was the translation of the perfectly straightforward, “Bitte nicht rauchen im Bett”, which is more or less literally, “Please do not smoke in bed”. Humans are also far from perfect at the translator’s job.
Today’s Washington Post has an interesting article on the evolution of machine translation. Beginning in the early 1950s, the general approach to the translation problem was to build a rule-based system. That is, the system “knew” about the rules of grammar, how to conjugate verbs, and so on. (Of course any translation system must also have a comprehensive dictionary of some sort.) The idea was that, knowing the rules of the source language and the target language, one could be reliably transformed into the other. But it is fair to say that these systems never did much to threaten interpreters’ job security.
The problem is not that there are no rules, but that there are too many of them. According to the work done by Chomsky and others, there is a certain amount of deep structure common to all languages. But there are an enormous number of special rules, idiosyncratic to particular languages, that have to be taken into account. (Recall when you were learning to spell. Let’s see, it’s “I before E, except after C …” except when it isn’t. Mark Twain once remarked that he would rather decline three free drinks than one German adjective.)
As the article points out, machine translation has improved considerably, mostly because newer efforts have taken a different approach. Instead of trying to specify all the rules of a language, they approach translation as an exercise in statistical inference. By examining a large body of parallel text in two or more languages, the system could learn common constructions and words usages in each language. In a sense, the approach is like that used in trying to break an unknown cipher.
Warren Weaver, a mathematician at the Rockefeller Foundation, had first raised the idea of a statistical model for translation in a 1947 letter in which he wrote: “When I look at an article in Russian, I say: ‘This is really written in English, but it has been coded in some strange symbols.’ “
(I wrote last summer about the use of a similar technique in the attempt to decipher unknown languages, like those of some ancient civilizations.)
The new, statistically-based techniques are the basis of Google’s translation service, and are also a significant part of Yahoo’s BabelFish service. The quality of the results for European languages has also been helped as a side effect of the formation of the European Union. Because all official documents must be translated and made available in all 23 official and working languages of the EU, and because governmental organizations produce documents as routinely as cattle produce cow-pats, there is a very large and steadily growing body of text to use as a source. Having used some of the older systems, and the newer ones, I think it is fair to say that a significant improvement has been made.
I think, too, there’s an interesting parallel between the evolution of machine translation, and the evolution of “intelligent” systems — but that is a subject for a later post.