The machines do the translating

Posted by Franz Och, Research Scientist

This year, Google entered the National Institute of Standards and Technology (NIST) Machine Translation evaluation, an event that has taken place annually since 2001 to support research in this area and to improve the state of the art. Participants included university labs, industrial labs, government labs and commercial machine translation companies from all over the world.

Our approach was to use statistical translation models learned from parallel text, that is, sets of documents and their translations. The system learns a model automatically from the parallel data. This approach differs from the rule-based approach used by many existing commercial machine translation companies which is based on large sets of handwritten translation rules.

We're very pleased with the results of this evaluation. Our computing infrastructure allows us to do a lot of experiments and work with huge data sets very easily.