LANGUAGE TRANSLATION is a tricky problem, not only for a piece of software but also for the human mind. A single word in one language, for example, may map into three or more in another. Carbonell likes to cite bank, with its utterly divergent uses for the place you keep your money, the edge of a river, and what an airplane might do. Then there are the dramatic differences in grammar and structure across languages. Arabic, for example, uses very little punctuation compared with English; Chinese contains no conjugations or plurals. For human translators, these problems are most often resolved through context or personal experience. There’s no rule that says “between a rock and a hard place” isn’t literal. We just know.
Machine translation is even trickier, and Carbonell’s “interesting errors” line is a good encapsulation of its history. Perhaps no technological endeavor has been more defined by its failures than the attempts over the last 60 years to use computers to convert one language into another. “It’s one of the earliest computer science problems to be attacked, and it has proven to be the one that’s most difficult,” says Nizar Habash, a research scientist at the Center for Computational Learning Systems at Columbia University.