Dialogue System Architectures

What are the typical modules of a dialogue system, how are they connected, and how do they communicate. We want to compare various options from the viewpoint of what consequences particular choices have w.r.t. to handling dialogue phenomena. For example, in order to allow for flexible turn-taking (i.e., barge-in), the system must be able to listen (and interpret) while talking, interrupt talking (only) when appropriate, and keep track of what output it actually has produced before the interruption. Another issue is speed: the typical pipe-line architecture results in slow system response, which may be disturbing for the user. Regarding modularity, what functionalities do modules have and how can they be reused for different tasks in different domains? Can we seprate generic and task/domain-specific functionality?

Just some references: Matthew Stone's issue-raising discussion [Stone2003]; Communicator project: Galaxy; TRIPS/TRAINS [Allen
2#2
2001]; BEETLE [Zinn
2#2
2002]; TrindiKit/GoDiS; CLT system (?), SmartKOM and related (e.g., [Pfleger
2#2
2003];