Integrated natural language generation is an approach to circumvent the problem of communication between different components of a language generation model. In order to achieve efficient integrated generation architectures, this dissertation tackles three tightly related questions: Formalism - a thorough analysis of appropriate formalisms with respect to power, expressiveness and complexity is given. Knowledge - based on the formalism, the question of how to transform and represent linguistic knowledge as well as application domain specific knowledge is addressed. Algorithm - step by step, a reversible and direct parser/generator based on simplified Schema-Tree Adjoining Grammars is developed to provide the core of an integrated natural language generation architecture and its prototypical implementation.