IBM's Two Paths to Machine Translation

Article Details

Global Watchtower

Common Sense Advisory Blogs

Posted by Donald A. DePalma on September 7, 2005 in the following blogs: Technology, Translation and Localization
Pages \| \|

In August Google had yet another 15 minutes of fame as word of its performance in the NIST machine translation shootout spread. On the heels of that we spoke with Language Weaver and then IBM. Brian Garr, ex-CTO of 1990s player Globalink, talked about the NIST test and IBM's MT offerings.

On the subject of Google having bested all comers in the NIST test, Garr told us what we had heard from Language Weaver and other MT suppliers -- participation is voluntary. Although IBM didn't participate, Garr noted that IBM has won several other Arabic MT bake-offs, including one at an unnamed federal agency that has been doing MT for 50 years. Incidentally it was IBM that created the BLEU (Bilingual Evaluation Understudy) benchmark used by NIST.

WebSphere Translation Server (WTS), IBM's commercial MT offering, is a rules-based system that IBM uses at its own online support center. 60 companies use WTS directly for outward-facing translation, while many more use it for internal business-to-employee applications such as e-mail. IBM includes a 1-CPU license with every portal license it sells, thus limiting its insight into how widely it is used. Garr said companies bring WTS inside their firewall to avoid having employees use free online translation services for sensitive documents.

IBM's statistics-based future entry in the NIST competition is under development at the T.J. Watson center in Yorktown Heights, NY (USA). Starting with a slot grammar invented by one of the researchers, the company is taking the path of many MT developers by melding its statistical MT with its rules-based research, adding more corpora for accuracy, and using its OmniFind project to search for word meanings. The team is tasked with determining whether more corpora or more rules will improve the quality of output.

What will happen with WTS? Garr says that IBM is happy with its scalability, using it both internally and externally. The company feel that it has the right languages for the markets it targets and will continue selling it until its research effort can visibly increase the quality of MT output. Will WTS meet your company's needs? As companies become more comfortable with MT and its limitations, NIST and other organizations will craft benchmarks to use-case workloads, processor loads, and language pairs, thus giving you a more scientific place to start your own evaluations. But as we have noted before, you won't know how any MT solution works until you try it out with your own content, workflow, and users.

Categories