VERBMOBIL is a speech-to-speech translation system working in English, German and Japanese [Bub and Schwinn1996]. The first four years of the project resulted in a prototype being able to produce approximately 75% correctly translated contributions in the domain of appointment scheduling dialogues. In the second phase, different extensions in domain and functionality were to be implemented. The domain now includes travel planning and hotel reservations and one of the new features is the automatic generation of a dialogue summary or script [Alexandersson and Poller1998]. Since, in the second phase of VERBMOBIL, a possible scenario is that of a telephone server used by two participants with VERBMOBIL as a third party, a dialogue script does provide some kind of a status report where each participant can check what items have been agreed on already. A summary then lists all final decisions in a thematic order. Such documentation would reduce misunderstanding and the need for clarification (a dialogue partner forgetting the agreed on items). Lack of robustness can be compensated by such a transparent context component where the user can see and react to errors in speech recognition or analysis.
For translation in the VERBMOBIL system various fundamentally different methods are employed. We call them deep and shallow translation tracks [Bub et al. 1997]. The deep track works with syntactical and semantical analyses and a logical, DRT-inspired representation to conduct a semantic transfer. Each language has its own generator for content-to-speech generation. The shallow track uses finite state technology and statistics as well as rule and plan based techniques to produce an approximate translation of task-relevant content [Reithinger1999].
One of the features of VERBMOBIL is its ability to take into account contextual and pragmatical information which is gained, accumulated and distributed by the dialogue and context component. It was the idea of making transparent this information that led to the new feature of an automatic summary/script. This paper describes work in progress, explaining how the context information is gathered and encoded, and what extensions to the dialogue manager and our interfaces were necessary to support the summary feature. We will also give examples of other consumers of context information within the VERBMOBIL system. We conclude the paper with open problems and future work.
Figure 1: Dialogue acts hierarchy as employed in VERBMOBIL 2
