Structural translation with synchronous tree adjoining grammars in VERBMOBIL

Abstract

The VERBMOBIL project is developing a translation system that can assist a face-to-face dialogue between two non-native english speakers. Instead of having continiously speak english, the dialogue partners have the option to switch to their respective mother tongues (currently german or japanese) in cases where they can't find the required word, phrase or sentence. In such situations, the users activate VERBMOBIL to translate their utterances into english. A very important requirement for such a system is realtime processing. Realtime processing is essentially necessary, if such a system is to be smoothly integrated into an ongoing communication. This can be achieved by the use of anytime processing, which always provides a result. The quality of the result however, depends on the computation time given to the system. Early interruptions can only produce shallow results. Aiming at such a processing mode, methods for fast but preliminary translation must be integrated into the system assisted by others that refine these results. In this case we suggest structural translation with Synchronous Tree Adjoining Grammars (S-TAGs), which can serve as a fast and shallow realisation of all steps necessary during translation, i.e. analysis, transfer and generation, in a system capable of running anytime methods. This mode is especially adequate for standardized speech acts and simple sentences. Furthermore, it provides a result for early interruptions of the translation process. By building an explicit linguistic structure, methods for refining the result can rearrange the structure in order to increase the quality of the translation given extended execution time. This paper describes the formalism of S-TAGs and the parsing algorithm implemented in VERBMOBIL. Furthermore the language covered by the german grammar is described. Finally we list examples together with the execution time required for their processing.