Dependency-directed Text Planning Keywords: Text Planning Data Dependencies Constraint-based Processing

Flexibility and EEciency. These are the two, connicting, requirements for a multilingual text planning component. Generating a natural language text requires a great amount of exibility in the rule set employed and, subsequently, in the control mechanism. For any local input semantic structure, there are, in general, many ways to implement it. Each choice produces ripples that move out and aaect other parts of the text. A choice that is locally optimal may, in the big picture, constrain text to global choices that are less than optimal. The interplay of constraints, dependencies, and local and global optimization can create a maze of possibilities that must be carefully planned for quality text to be produced in a reasonable amount of time. Which leads to EEciency. MLG applications such as database interfaces and, in the future, spoken language interfaces demand speed. In the past, this requirement has often been met by simplifying the semantic input structure as well as the range of language outputs produced. Simplifying semantics leads to misunderstanding and misrepresentation. Simplifying language leads to sterile text at best, and often to incorrect assumptions that again produce misunderstanding. This paper presents an approach to controlling an MLG text planning component that recognizes and takes advantage of the complex inter-dependencies present in a realistic communication situation. It describes how dependencies can be tracked and then exploited by \opportunistic", \island driven", and \sound" processing techniques. Furthermore, it discusses the typology of rules necessary for the text planning component , and how these aaect the style of processing.