Discourse Structure and Anaphora: An Empirical Study

ne of the main motivations for studying discourse structure is its effect on the search for the antecedents of anaphoric expressions. We tested the predictions in this regard of theories assuming that the structure of a discourse depends on its intentional structure, such as Grosz and Sidner?s theory. We used a corpus of tutorial dialogues independently annotated according to Relational Discourse Analysis (RDA), a theory of discourse structure merging ideas from Grosz and Sidner?s theory with proposals from Rhetorical Structure Theory (RST). Using as our metrics the accessibility of anaphoric antecedents and the reduction in ambiguity brought about by a particular theory, we found support for Moser and Moore?s proposal that among the units of discourse assumed by an RST-like theory, only those expressing an intentional ?core? (in the RDA sense) should be viewed as constraining the search for antecedents; units only expressing informational relations should not introduce separate focus spaces. We also found that the best compromise between accessibility and ambiguity (?perplexity?) reduction is a model in which the focus spaces associated with embedded cores and embedded contributors remain on the stack until the RDA-segment in which they occur is completed, and discuss the implications of this finding for a stack-based theory.