Submitted to Eurospeech 2003 (Geneva, Switzerland, September 2003)
ABSTRACT: Widely used declarative approaches to generation in which generation speed is a function of grammar size are not optimal for real-time dialogue systems. We argue that a procedural system like the one we present is potentially more efficient for time-critical real-world generation applications as it provides fine-grained control of each processing step on the way from input to output representations. In this way, the procedural behaviour of the generator can be tailored to the task at hand. During the generation process, the realizer generates flat deep structures from semantic-pragmatic expressions, then syntactic deep structures from the deep semantic-pragmatic structures, and from these syntactic deep structures surface strings. Nine different generation levels can be distinguished and are described in the paper.
To access the full content of this article, please fill in the form below.