In this dissertation I present a model for the determination of intonation contours from context and provide two implemented systems which apply this theory to the problem of generating spoken language with appropriate intonation from high-level semantic representations. The theory and implementations presented here are based on an information structure framework that mediates between intonation and discourse, and encodes the proper level of semantic information to account for both contextually-bound accentuation patterns and intonational phrasing. The structural similarities among these linguistic levels of representation are the basis for selecting Combinatory Categorial Grammar (CCG,Steedman1985, 1990a) as the model for spoken language production. This model licenses congruent syntactic, prosodic and information structural constituents and consequently represents a simplifcation over models of prosody developed in syntactically more traditional frameworks. The previous mention heuristic, which has been widely used as a model for determining intonation contours, is shown to be inadequate for handling a broad range of examples involving semantic contrasts, which require pitch accents to be allocated based on their ability to discriminate among available entities in the discourse model. To address this problem, I introduce a model that determines accentual patterns based on sets of alternative entities in the knowledge base. The algorithms for building the information structural representations that encode the semantics of intonation supply the foundation for two computational implementations. These implementations demonstrate how the theoretical model applies to the problem of producing contextually-appropriate spoken output in a natural language generation framework and provide a platform for incrementally testing and refning the underlying theory.