Discussion and Future Work

We have presented the work in progress on the VERBMOBIL dialogue component which had to adapt to an extended domain and the new functionality of summary/script generation in the second phase of the project. An overview of the shallow processing track was given, demonstrating the advantages of statistical and shallow finite state technology in a speech-to-speech system which requires a high level of robustness. We explained the structure of our representation of dialogue context and how to manage topic, focus and agreement in a negotiation dialogue so that a summary can easily be generated, again taking into account the requirements for robustness.

There are many open questions to think about and work on. Here we list some of them:

• how far do we get with this approach? Will we be able to extend the coverage of the summary/script with this representation? How does it transfer to other domains?
• how do we evaluate the dialogue manager? We still have to check how decisions about negotiation agreements can be made more reliable.
• is there an overall framework? We have to compare our approach with related systems/techniques.
• do the results apply for interactive systems? We believe so but what extensions would we need?

Our future work in the further course of the VERBMOBIL project will focus on the following items:

• multi-speaker: do we need other dialogue acts? How do we know who's talking to whom?
• combining deep and shallow information: how do we handle different segmentation?
• extending coverage: so the summary can include timetables, excuses, reasons, social formulae