A Distributional View of Discourse Encapsulation: Multifactorial Prediction of Coreference Density in RST

Early formulations of discourse coherence constraints postulated a connection between coreference likelihood and distance within a discourse parse, e.g. in the framework of Veins Theory (Cristea et al. 1998%CristeaIdeRomary1998), which proposes that coreference is expected to be encapsulated within tightly linked areas of discourse parses, called Domains of Referential Accessibility (DRAs). Using an RST dependency representation, this paper expands on previous work showing the relevance of DRAs to coreference likelihood. We develop a multifactorial model using both rhetorical and surface distance metrics, as well as confounds such as unit length and genre, and direct versus indirect rhetorical paths. We also explore coreferential accessibility as it applies to less studied types of coreference, including bridging and lexical coreference. The results show that rhetorical and surface distance, as well as direct linking, all influence coreference likelihood, and should not be treated as mutually exclusive or redundant metrics. Finally, we incorporate RST relation-specific tendencies that offer a more fine-grained model of coreference accessibility.