August 01, 2009, Ottawa, Ontario
Canada can compete with global leaders in the worldwide language industry now that NRC has launched PORTAGE — a statistics-based software system that yields far better results than earlier attempts to automate the highly nuanced art of translation.
The PORTAGE system was developed by the NRC Institute for Information Technology (NRC-IIT). “PORTAGE is the outcome of our research to create a robust and adaptable system built on machine learning techniques,” says Michel Mellinger, a business development officer for NRC-IIT . “Although other countries have already developed their own statistical machine translation systems, PORTAGE is the only such Canadian-made system. And, in international competitions, PORTAGE has come out near the very top.”
Research into machine translation goes back to the 1950s. But, until recently, all such systems depended on human experts creating the language rules for each paired set of languages, which were then programmed into the software. These “rule-based” systems had the reputation of creating real bloopers — the kind you sometimes see on funny English signposts in other countries. These systems are also expensive to build and maintain and, as the number of manually generated rules grows, they yield unexpected interactions that ensure constant work for software engineers.
PORTAGE goes a long way to avoid this sort of outcome. NRC researchers and software engineers have “trained” the system with a vast body of paired translated texts (French-English, Chinese-English, Finnish-English, and many more combinations). The quality of the training is one reason why PORTAGE has performed so well in competitions.
Despite some 40 years of R&D on automated translation systems, it was only in the 1990s that researchers began looking at creating programs that could “learn” how to translate through the statistical analysis of vast bodies of paired translations. This approach doesn’t require the onerous step of integrating and programming grammar rules and terminology for each language. Instead, the system uses a combination of statistical models about how to go from one language to another, the properties of the target language and the kind of text being translated. One of the biggest advantages of this new approach is that it provides several possible translations of the same sentence, with statistical analysis indicating which one has the highest probability of being the best match based on several criteria.
So, does this advance mean we won’t need translators any more? “Because the quality of human-generated translation will always be superior to machine-generated translation, PORTAGE will never replace people,” says Mellinger. “It will simply help them increase the speed and accuracy of translation to meet increasing demand. In the near future, most of the translator’s job will be to revise a high-quality machine-generated text.”
NRC is offering a fully operational distribution version of its system — called PORTAGEshared — to Canadian universities interested in conducting R&D on statistical machine translation. By making PORTAGE available to university researchers, NRC is contributing to the development of a Canadian community of expertise that could help take statistical machine translation to new levels of performance. NRC is also offering evaluation licences for language companies to test the PORTAGE system and determine whether they can develop specialized commercial applications for it.
At the end of August, PORTAGE will be on display in a technology showcase organized by the 12th international Machine Translation Summit in Ottawa. The NRC team will share how they package a “trained” PORTAGE system in any computer environment through a technique called virtual machines. The result, called PortageLive, will make it easy for clients (e.g. translation firms) to start using the PORTAGE system in their businesses.
