Negotiation Strategies in an Integrated Natural Language Generation System

This paper describes negotiation strategies in an integrated generation system (INLGS) based on the formalism of Schema–Tree Adjoining Grammars with Unification (SU–TAGs). Integrated or uniformmeans that all knowledge bases are specified in the same formalism and run the same processing algorithm. In our project a reversible parser/generator runs knowledge bases in the formalism of Schema– Tree Adjoining Grammars with Unification. SU–TAGs comprise a condensed grammar representation and direct parsing/generation deals with partially unspecified schemata. Parametrizing the knowledge bases means that the user specifies a hierarchical system where the rules are classified according to their generality. This is done for all knowledge bases (i.e. components). The decision making of the generation system always tries to find a solution that satisfies as much rules of the most highest generality level. Backtracking of dead-ends which can e.g. occur in incremental generation is also guided by this method. First the most fine-grained rules are revised before more general decisions are withdrawn. A main advantage of this architecture is that negotiation between generation components can easily be imposed on the system. We illustrate this on how to resolve local dead–ends in a parametrized manner.