Machine translation

Researchers Are Cracking the Code of Machine Translation

University of Southern California scientists Sujith Ravi and Kevin Knight have developed a machine-translation technique that treats translation as a cryptographic challenge. Current machine translation methodology uses statistical analysis of the same text in two different languages to produce translations. Ravi and Knight have written software that calculates the likelihood of a foreign word matching an English word based on the frequency of its occurrence within the text. The pair then use a second computer program to evaluate the quality of the translation and guarantee that it makes sense in English. This assessment information can then be used to further refine the translation software. Ravi explains that their approach was designed to avoid the need for parallel data—the same text in two different languages—which frequently stumps machine translation in less commonly used languages, such as Farsi or Pashto. There are often too few bilingual sources available in these languages for translation software to be successful. Ravi and Knight’s system has been tested on a collection of short phrases and movie titles. The results, known as monolingual translation, compared favorably to standard machine translation. Johns Hopkins University researcher Chris Callison-Burch says that Ravi and Knight’s technique holds promise, but has yet to prove itself. His team is also developing translation software that avoids parallel data. Their program “crawls” online texts and compares disparate texts from different languages. For example, when comparing Spanish blogs to English news articles following the 2004 Indian Ocean earthquake, the use of the word “tsunami” increased sharply in English as did the word “maremoto” in Spanish, suggesting that the two words mean the same thing. Phil Blunsom, a machine translation researcher at the University of Oxford, says neither technique is likely to make waves in the immediate future. “They’re trying to do something very ambitious,” Blunsom says. “It’s not something you’re going to see popping up in commercial systems any time soon.”

From «Cracking the Code of Machine Translation»
New Scientist (United Kingdom) (06/20/11) Aron, Jacob