Developing Deployable Spoken Language Translation Systems given Limited Resources

Automatic machine translation systems in research institutions have reached a considerable level of performance. This is especially true for Statistical Machine Translation. Since its introduction in the 1990s, it has outperformed earlier approaches and produces a translation quality that seemed impossible only a short time ago. Particularly for limited domains like tourism dialogs, medical relief or force protection the translations could already be very useful for real applications outside of laboratories or demonstrations. The question here is no longer how to get the technology to an acceptable level, but how the existing technology can be globally deployed. How can a system be developed that tourists, health professionals or soldiers can actually use and benefit from? The following lists the issues which need to be adressed in order to make a current research system deployable to real users: Research system Deployable system Languages Limited number Many Language Pairs Hardware High-End Server Notebook, Mobile Device Application Evaluations, Demonstrations Actual Communication Evaluation BLEU/NIST Scores Communication success Users Expert researchers Tourists, Medical users, Military users Interface Complex, command line Easy, interactive The first difference between a research system and a deployable system is the number of generally supported languages. In research most groups focus on a few language pairs for which large training corpora are available. Systems for tourists and especially for medical relief or military uses will have to support a much larger number of languages. It is especially necessary to rapidly support specific language pairs if a sudden demand develops. In research labs the translation systems are usually run on high-powered, expensive servers to be able to use larger and more complicated models and to realize even small performance improvements. The users are researchers and