Generierung natürlicher Sprache mit Generalisierten Phrasenstruktur-Grammatiken

This dissertation describes NL generation with generalized phrase structure grammars (GPSG). It thoroughly discusses the theory of GPSG and argues that it can, in its 1985 version, not efficiently be implemented. Therefore, some modifications are suggested that overcome this problem. Using the modified formalism, two different approaches to GPSG-based generation are presented. The grammar-driven approach is shown to suffer from indeterminism, whereas the structure-driven approach can be efficiently implemented. The latter method has been explored within the Berlin machine translation system, where the generator starts from sentence-semantic input structures designed for transfer. Since the semantics is not integrated into the grammar, an explicit mapping from partial semantic onto syntactic structures is necessary, which is accomplished by using techniques known from AI production systems. Both generators allow for multilingual generation, which is demonstrated for English and German.