Within the natural-language research community it has long been acknowledged that the conventions and pragmatics of natural-language communication often oblige dialogue systems to consider and address the underlying purposes of queries in their responses rather than answering them literally and without further comment or elaboration. Such systems cannot simply translate their users' requests into transactions on database or expert systems, but must apply many more complex reasoning mechanisms to the task of selecting responses that are both appropriate and useful. This idea has given rise to a broadly-defined program of research in cooperative response generation (CRG). Research in CRG carried on over more than a decade has yielded a substantial body of literature. Analysis of that literature, however, shows that investigators have focused primarily on modeling manifestations of cooperative behavior without directly considering the nature and motivations of the behavior itself. But if we want to develop natural language dialogue systems that are truly to function as cooperative respondents instead of serving only as models of particular kinds of cooperative responses, a different approach is required. I identify two opposing perspectives on the process of cooperative response generation: the questioner-based and the respondent-based perspectives. I argue that past research efforts have largely been questioner-based, and that this view has led to the development of theories that are incompatible and cannot be integrated. I propose the respondent-based view as an alternative, and provide evidence that taking such a perspective might allow several interesting but otherwise poorly-understood aspects of cooperative response behavior to be modeled. The final portion of the dissertation explores the computational implications of a respondent-based perspective. I outline the architecture of a Cooperative Response Planning System, a dialogue system that raises, reasons about, and attempts to satisfy high-level cooperative goals in its responses. This architecture constitutes a first approximation to a theory of how a system might reason from the beliefs it derives from a questioner's utterances to choose a cooperative response. The processing of two sample responses in this framework is described in detail to illustrate the architecture's capabilities.