How Complex is Discourse Structure?

This paper contributes to the question of which degree of complexity is called for in representations of discourse structure. We review recent claims that tree structures do not suffice as a model for discourse structure, with a focus on the work done on the Discourse Graphbank (DGB) of Wolf and Gibson (2005, 2006). We will show that much of the additional complexity in the DGB is not inherent in the data, but due to specific design choices that underlie W&G’s annotation. Three kinds of configuration are identified whose DGB analysis violates tree-structure constraints, but for which an analysis in terms of tree structures is possible, viz., crossed dependencies that are eventually based on lexical or referential overlap, multiple-parent structures that could be handled in terms of Marcu’s (1996) Nuclearity Principle, and potential list structures, in which whole lists of segments are related to a preceding segment in the same way. We also discuss the recent results which Lee et al. (2008) adduce as evidence for a complexity of discourse structure that cannot be handled in terms of tree structures.