Publication
Putting Semantic-Head-Driven Generation to the Limits: Experiments with Multi-Purpose Semantic Representations
Stephan Busemann
In: Tilman Becker; Stephan Busemann; Wolfgang Finkler (Hrsg.). DFKI Workshop on Natural Language Generation. DFKI Workshops, August 23, Saarbrücken, Germany, Pages 8-14, DFKI Documents (D), Vol. 97-06, DFKI, 1997.
Abstract
Constraint-based grammars allow for explicitly interfacing syntactic and semantic layers of linguistic sentence representation. They can, in principle, serve as the major linguistic knowledge source for both parsing and generation. While the input layer for parsing is close to trivial (usually a list of morphologically analyzed words), semantic representations, which serve as input to generation, may vary according to the theory adopted and the purpose followed by the grammar writer. As a consequence, not all semantic representations can be processed equally well by some generation algorithm.
We show that linguistically plausible semantic analyses can cause severe problems for generation. We use a variant of the semantic-head-driven generation algorithm and the PAGE grammar of German, both developed at DFKI, as our source of examples. We then discuss possible ways out, and describe in detail the approach taken and implemented within the SeReal system. It consists in defining a suitable reorganisation of the semantics within the grammar and using it as input representation for the generator.