Statistical translation of spoken dialogues in the verbmobil system

This paper gives an overview of our work on statisti al ma hine translation in the framework of the Verbmobil proje t. The goal of the Verbmobil proje t is the trans-lation of spoken dialogues in the domains of appointment s heduling and travel planning. Starting with the Bayes de ision rule as in spee h re ognition, we show how the re-quired probability distributions an be stru tured into three parts: the language model, the alignment model and the lexi on model. We des ribe the omponents of the system and report results on the Verbmobil task. The experi-en e obtained in the Verbmobil proje t, in parti ular a large-s ale end-to-end evaluation, showed that the statisti-al approa h resulted in signi antly lower error rates than three ompeting translation approa hes: the senten e error rate was 29 % in omparison with 52 % to 62 % for the other translation approa hes. 1