Tower of Google uses stats for translation
Number crunching our way to understanding
Google dreams of a world where hundreds of languages can be simultaneously translated by machines which compare texts using statistics rather than applying grammatical rules.
Statistical machine translation uses a computer to compare two documents - one in the original language and one translated by a human. It finds patterns and links between the two and uses them to create its own future translations.
Google has used documents from the European Commission and United Nations to feed its machines.
Franz Och, who runs Google's translation team, told Reuters that early efforts impressed people with experience of machine-run translation systems.
Och said: "The more we feed into the system the better it gets."
Google already offers statistical translation of Arabic, Chinese and Russian. Other language translations are provided by third parties. Och repeated the Google mantra that the focus was on improving the software and that once it was working well they would look at making money from it.
Dr Miles Osborne, a lecturer at the University of Edinburgh who spent last year on sabbatical at Mountain View working on the system, told the Reg: "This is quite a recent move by Google - they hired Franz Och one of leading lights in statistical translation. What you see with this system is what an academic would make if they had lots of money, support, and access to lots of machines. They have one of the world's best translators, especially for Arabic and Mandarin."
Osborne said the development was important for Google because documents on the web are increasingly in languages other than English. To continue to improve its core search engine, Google would need translation software.
Asked why Arabic and Mandarin were the first languages chosen, Osborne said it was down to US paranoia and homeland security. He said the US military put cash into research for translating languages from areas they considered a threat.
Osborne said the US Army preferred computer-based systems because they distrusted human translators.
Reuters' story is here.
Google's language tools are here. ®
Sponsored: Benefits from the lessons learned in HPC