| When computers communicate with humans to facilitate a decision process, they must do so in an unambiguous and concise manner which reflects the needs and preferences of the user, and takes into account the situation in which the system performs. This dissertation explores the fulfillment of these requirements of decision support in the context of systems that receive information from multiple, independent sources. In the first part of this thesis I present my hypothesis that independent text plans, analyzed in terms of their rhetorical structure [MT87], can be integrated to enhance conciseness and coherence. An implemented system, RTPI, demonstrates the application of this hypothesis to improve message sets generated by an existing decision support system through the integration of related messages. An evaluation of RTPI shows that it reduces repetition and overall message length, and human subjects prefer its integrated messages to the original sets of messages.; The second part of this thesis explores the additional challenges presented when decision support information is obtained and integrated from multiple sources in a dynamic information environment. In this environment, information attributes such as quality, cost, length, and production time can vary over time; in addition, user preferences regarding those attributes can also change. I present my hypothesis that this problem can be computationally addressed, by a flexible, responsive, agent-based system that takes into account both user preferences and information attributes that change over time. The implemented system MADSUM demonstrates a multi-agent architecture that addresses all of these issues. An evaluation of the system's output shows that human subjects, when presented with a scenario including user preferences and information, agree with MADSUM's selection and ordering of information from a proposition set. |
