Not all the innovation in the machine translation (MT) sector comes from the developers of that software. Language service providers (LSPs) with a huge stake in the productivity gains of their translations have gone beyond post-editing and workflow integration to conducting research into improving MT engines.

Their efforts are aimed at delivering on the promise of post-edited MT to increase translation volumes and decrease costs without degrading quality (see “The Business Case for Machine Translation,” Aug09). The Gerson Lehrman Group spoke with euroscript CEO Mark Evenepoel and Karina Martinez-Ferber, manager of Processes & Technology Language Services, about their company’s MT project.

euroscript Deutschland and the German Research Center for Artificial Intelligence (DFKI) created the taraXÜ project to develop an architecture that combines the leading approaches to MT (RBMT, SMT and template-based MT), a self-calibrating mechanism to refine the output, and automated source and target text evaluation that is an integral part of the workflow. The taraXÜ developers dubbed their architecture “tribrid” for its integration of these three elements. They created the name “taraXÜ” from the German verb Tarieren (to calibrate), “X” for “extensible” or “exchangeable” for the language combination, and “Ü” for Übersetzen (to translate). Other participants in the project are Acrolinx and Yocoy Technologies, both DFKI spin-offs.

Over the last 12 months, several MT developers have offered hybrid solutions that focused on different parts of the translation process, so they asked euroscript what comprised the “secret sauce” for taraXÜ? Evenepoel informed that taraXÜ views MT results from rules-based and statistical systems as just the first part of its analytic process. Then the self-calibration component applies controlled language technology and what Evenepoel characterized as a euroscript-developed “pragmatic methodology” that grounds the post-processing in real-world translation concerns. He said that “The evaluation test sets are combined of real-life source text material coming from euroscript’s translation projects and scientific text sets as they are used in research.” Read more.

See: Gerson Lehrman Group