This thesis presents a computational model of some of the knowledge and processing capabilities required by a speaker to produce fluent, coherent discourse in a natural language. The main purpose of the work is to study the representation of linguistic knowledge and the way in which this knowledge is used to produce utterances which achieve a communicative task in a given discourse situation. The model defines a simple world in which two actors communicate with each other in the course of attempting to achieve practical goals. The actors' behaviour is organised at three levels the level of non-linguistic planning and action, the level of the organisation of self-other interactions (turn-taking), and the level of the construction of individual utterances. Actions at lower levels gain functional relevance by their contribution to the achievement of goals defined at higher levels. At all levels, the model performs principled computations based on explicitly represented knowledge, thus embodying an account of how a non-linguistic goal might give rise to a particular form of words under appropriate circumstances