SBIR-STTR Award

This Small Business Technology Transfer (STTR) Phase I project concerns R&D aimed at assessing the feasibility of applying statistical Machine Translation (MT) techniques to the problem of improving the productivity of human translators. Currently, human translators use translation memory tools, i.e., software packages that provide access to databases of previously translated sentences. Unfortunately, these tools do not provide significant help in translating previously unseen sentences and do not improve over time (with the exception of providing access to increasingly larger databases of previously translated material). Because automatic translation systems produce low quality translations that are not tailored to their genre and domain of interest, human translators refuse to use automatic translation systems. In this program, a prototype hybrid translation system and computer interface that will permit humans to translate text by exploiting both a translation memory and an automatic, statistical-based MT system will be built and the increase in text translation productivity that is enabled by the use of the hybrid tool will be measured. A hybrid translation tool such as that proposed here has the potential to reduce significantly the costs associated with human translation, and increase translation productivity.

The goal of this Small Business Innovation Research (SBIR) Phase II project is to bring radically new technology to the machine translation marketplace. While current systems are rule-based and difficult to extend, this company employs a statistical system that learns to translate by automatically analyzing large collections of previously translated material. This technology already outperforms rule-based systems, and it easily adapts to specific domains of interest, such as technical documentation generated by multinational corporations. The company has licensed (and co-developed) key software engines from the founders' research team at USC/ISI, a world leader in machine translation. In this project, they will extend their statistical engine in three ways, driven by customer needs -- they propose to build (1) a parallel, cluster-based training system for handling large text volumes (2) new capabilities for translating numbers, dates, personal names, locations, etc. ("named entities"), and (3) rapid customization tools that will assist with customized translation engines for specific customer domain requirements.