The Statistical Translation Module in the Verbmobil System
Stephan Vogel, Franz Josef Och, Hermann Ney · 2000
In the VERBMOBIL project several translation methods are used concurrently to improve the overall performance of the system. One of the translation modules is based on a statistical approach. In this paper, we give an overview of this approach and how it is embedded in the VERBMOBIL system. Results from a comparative evaluation are given to compare the statistical approach to the other translation methods. 1 Introduction Several features make the statistical approach attractive not only for NLP in general but for machine translation in particular. Quick prototyping: As more and more bilingual corpora become available, prototyping of a translation system for a new domain within a very short period of time is possible. Adaptation to task: as the system is trained on given data, specific characteristics of the domain are captured automatically. Robustness: Using statistical methods no hard decisions are necessary. Whereas rule based systems often fail on ungrammatical input, a ...