WATANI International
11 April 2010
Everybody has his own tale of terrible translation to tell — an incomprehensible restaurant menu in Croatia, a comically illiterate warning sign on a French beach. “Human-engineered” translation is just as inadequate in more important domains. In our courts and hospitals, in the military and security services, underpaid and overworked translators make muddles out of millions of vital interactions. Machine translation can certainly help in these cases. Its legendary bloopers are often no worse than the errors made by hard-pressed humans.
Machine translation has proved helpful in more urgent situations as well. When Haiti was devastated by an earthquake in January, aid teams poured in to the shattered island, speaking dozens of languages — but not Haitian Creole. How could a trapped survivor with a cellphone get usable information to rescuers? If he had to wait for a Chinese or Turkish or an English interpreter to turn up he might be dead before being understood. Carnegie Mellon University instantly released its Haitian Creole spoken and text data, and a network of volunteer developers produced a rough-and-ready machine translation system for Haitian Creole in little more than a long weekend. It didn’t produce prose of great beauty. But it worked.
The advantages and disadvantages of machine translation have been the subject of increasing debate among human translators lately because of the growing strides made in the last year by the newest major entrant in the field, Google Translate. But this debate actually began with the birth of machine translation itself.
The need for crude machine translation goes back to the start of the cold war. The United States decided it had to scan every scrap of Russian coming out of the Soviet Union, and there just weren’t enough translators to keep up (just as there aren’t enough now to translate all the languages that the United States wants to monitor). The cold war coincided with the invention of computers, and “cracking Russian” was one of the first tasks these machines were set.
The father of machine translation, William Weaver, chose to regard Russian as a “code” obscuring the real meaning of the text. His team and its successors here and in Europe proceeded in a commonsensical way: a natural language, they reckoned, is made of a lexicon (a set of words) and a grammar (a set of rules). If you could get the lexicons of two languages inside the machine (fairly easy) and also give it the whole set of rules by which humans construct meaningful combinations of words in the two languages (a more dubious proposition), then the machine would be able translate from one “code” into another.
Academic linguists of the era, Noam Chomsky chief among them, also viewed a language as a lexicon and a grammar, able to generate infinitely many different sentences out of a finite set of rules. But as the anti-Chomsky linguists at Oxford commented at the time, there are also infinitely many motor cars that can come out of a British auto plant, each one having something different wrong with it. Over the next four decades, machine translation achieved many useful results, but, like the British auto industry, it fell far short of the hopes of the 1950s.
Now we have a beast of a different kind. Google Translate is a statistical machine translation system, which means that it doesn’t try to unpick or understand anything. Instead of taking a sentence to pieces and then rebuilding it in the “target” tongue as the older machine translators do, Google Translate looks for similar sentences in already translated texts somewhere out there on the Web. Having found the most likely existing match through an incredibly clever and speedy statistical reckoning device, Google Translate coughs it up, raw or, if necessary, lightly cooked. That’s how it simulates — but only simulates — what we suppose goes on in a translator’s head.
Google Translate, which can so far handle 52 languages, sidesteps the linguists’ theoretical question of what language is and how it works in the human brain. In practice, languages are used to say the same things over and over again. For maybe 95 percent of all utterances, Google’s electronic magpie is a fabulous tool. But there are two important limitations that users of this or any other statistical machine translation system need to understand.
The target sentence supplied by Google Translate is not and must never be mistaken for the “correct translation.” That’s not just because no such thing as a “correct translation” really exists. It’s also because Google Translate gives only an expression consisting of the most probable equivalent phrases as computed by its analysis of an astronomically large set of paired sentences trawled from the Web.
The data comes in large part from the documentation of international organizations. Thousands of human translators working for the United Nations and the European Union and so forth have spent millions of hours producing precisely those pairings that Google Translate is now able to cherry-pick. The human translations have to come first for Google Translate to have anything to work with.
The variable quality of Google Translate in the different language pairings available is due in large part to the disparity in the quantities of human-engineered translations between those languages on the Web.
But what of real writing? Google Translate can work apparent miracles because it has access to the world library of Google Books. That’s presumably why, when asked to translate a famous phrase about love from “Les Misérables” Google Translate comes up with a very creditable answer, which just happens to be identical to one of the many published translations of that great novel. It’s an impressive trick for a computer, but for a human? All you need to do is get the old paperback from your basement.
But the program can be very patchy. For works that are truly original — and therefore worth translating — statistical machine translation hasn’t got a hope. Google Translate can provide stupendous services in many domains, but it is not set up to interpret or make readable work that is not routine — and it is unfair to ask it to try. After all, when it comes to the real challenges of literary translation, human beings have a hard time of it, too.
____________________________________________________________
David Bellos is the director of the Program in Translation and Intercultural Communication at Princeton. The New York Times (abridged).