Generating discourses from networks using an inheritance-based grammar
Mark Seligman · 1992
This dissertation presents two separable designs: First, I describe an inheritance-based grammar which represents both syntactic and semantic structures as frames. (In a departure from previous designs, mapping structures which mediate between syntactic and semantic structures are frames as well.) While stressing representational neutrality between text analysis and generation, I focus on the latter. The approach to sentence generation features the merging of structures via inheritance and the tracking of expressive choices using multiple worlds. Second, I propose new semantic/pragmatic representations for paragraph-length discourses, and expression procedures based upon them. Discourse relations are to be discovered via judgment of synonymy, and are arranged in lattices (rather than exclusively in hierarchies). A small set of network traversal strategies permits repeated application to the sentence grammar to generate paragraphs expressing the whole net. (These include a bottom-up generative strategy, in which completed fragments are incorporated into larger structures in progress.) A facility for expressing situational formulas adds a pragmatic component to the system. Several variants are generated for two attested discourses.