Practical, Template–Based Natural Language Generation with TAG
Tilman Becker · 2002
This paper describes a TAG–based (Joshi and Schabes, 1997), template–based approach to Natural Language Generation. It extends the idea of lexicalization to whole phrases, similar in style to the representation of idioms in a TAG grammar. In addition to this, there is a second type of templates: partial derivations. The first phase in the generator, driven by planning rules, produces a derivation tree which is then used in a straightforward realizer to create the derived tree. This tree is then the basis for the input to a Concept–To–Speech synthesis component (Schweitzer, Braunschweiler and Morais, 2002). There are two basic methods for constructing these two kinds of templates. The first and preferred one is based on an existing grammar where the templates represent shortcuts that would be used for reasons as simplicity, efficiency, or because the existing grammar does not contain an interface to the representation language, e.g., semantics, that is used as the generator‘s input. The second method which is used so far in the SMARTKOM project (Wahlster, Reithinger and Blocher, 2001) is necessary when no suitable grammar exists. It allows for an approach similar to rapid prototyping: only the templates that are needed are specified and the templates are initially kept as large as possible. Only in the following development steps, the templates are made smaller and smaller, eventually themselves becoming a grammar. The following sections briefly present the SMARTKOM project and the generator architecture. Then the use of fully specified templates is discussed in this context, including their use for concept–to–speech synthesis. Finally we present the current work, including some tools that are under development.