We do not keep an explicit dialogue model with a range of legal states and transitions. Instead we have a set of internal actions triggered mainly by the segment's dialogue act and using different sorts of information as arguments (topic, focus, stored suggestions etc.). It is certainly possible to formulate a dialogue model based on a formal definition of dialogue states but this may be of more importance in an interactive system where one wants strict control over the system's (visible) behaviour.
The dialogue manager needs to handle content: anaphoric references and ellipses in suggestions have to be resolved. The dialogue manager also keeps track of the speakers' attitudes towards the content objects (accept/reject etc.) and handles topic (see above) as well as focus. Each new segment/utterance is processed according to its dialogue act:
The annotation of the speakers' attitudes serves as evidence for the summarization (see section 4.4).
Question-answer pairs are dealt with by pushing the question item to a temporary short-term storage and waiting for a reply. The reply then triggers the treatment of the content data. We distinguish two types
Figure 4: principal temporal units (capital letters) and their possible specifications