| When people talk with computers, they interact very differently from how they interact in conversation with other people. Some of these differences come from people's adapting to limitations in the computer's abilities to interact naturally. If we can make computers interact in a more realistic way, we can reduce the cognitive load on users of dialogue systems.; In this dissertation, we explore how to design generation systems that can not only produce coherent, informative and responsive dialogue contributions, but also explicitly model human styles of interaction. This exploration includes two approaches. First, we report on the in-depth analysis of a corpus of human-human task-oriented dialogues, the Monroe corpus. We use this analysis to examine the relative contributions of the different levels of conversation acts: turn-taking, grounding, speech acts and argumentation acts. We also examine the effect of initiative and dialogue levels on task-oriented conversation.; Second, we describe a computational model of generation based on conversation acts theory. This model has been implemented in a generation system for the TRIPS system at the University of Rochester. The generation system is flexible and efficient. It includes components to plan dialogue contributions, perform surface generation for different media, and coordinate output production. We discuss our implementation, and describe an evaluation of the model. |
