Using pattern-action rules for the generation of GPSG structures from separate semantic representations

In many tactical NL generators the semantic input structure is taken for granted. In this paper, a new approach to multilingual, tactical generation is presented that keeps the syntax separate from the semantics. This allows for the system to be directly adapted to application-dependent representations. In the case at hand, the semantics is specifically designed for sentence-semantic transfer in a machine translation system. The syntax formalism used is Generalized Phrase Structure Grammar (GPSG). The mapping from semantic onto syntactic structures is performed by a set of pattern-action rules. Each rule matches a piece of the input structure and guides the GPSG structure-building process by telling it which syntax rule(s) to apply. The scope of each pattern-action rule is strictly local, the actions are primitive, and rules can not call each other. These restrictions render the production rule approach both highly modular and transparent.