ITRI-99-15 Generating Embedded Discourse Markers from Rhetorical Structure

To understand a discourse, the reader needs to recover the relations between the discourse elements as intended by the writer. Writers can, and very often do, help the reader along by providing explicit lexical signals of the intended discourse relations through the use of lexicalised discourse markers. In this paper we present a novel approach for generating texts containing multiple, embedded rhetorical relations, each of which is lexically marked. Our analysis integrates syntactic factors, ideas from rhetorical structure theory, notions of text-grammar, and findings from psycholinguistic research into a unified, feature-based account. The current implementation allows us to generate several good variants of text structures containing multiple discourse relations, while blocking dispreferred variants.