Veins Theory: A Model of Global Discourse Cohesion and Coherence

Veins Theory: A Model of Global Discourse Cohesion and Coherence Dan CRISTEA Dept. of Computer Science University <<A.I. Cuza>> la~i, Romania dcristea@infoiasi.ro Nancy IDE Dept. of Computer Science Vassar College Poughkeepsie, NY, USA ide@cs.vassar.edu Laurent ROMARY Loria-CNRS Vandoeuvre-Les-Nancy, France romary @ loria.fr Abstract In this paper, we propose a generalization of Centering Theory (CT) (Grosz, Joshi, Weinstein (1995)) called Veins Theory (VT), which extends the applicability of centering rules from local to global discourse. A key facet of the theory involves the identification of <<veins>> over discourse structure trees such as those defined in RST, which delimit domains of referential accessibility for each unit in a discourse. Once identified, reference chains can be extended across segment boundaries, thus enabling the application of CT over the entire discourse. We describe the processes by which veins are defined over discourse structure trees and how CT can be applied to global discourse by using these chains. We also define a discourse <~smoothness>> index which can be used to compare different discourse structures and interpretations, and show how VT can be used to abstract a span of text in the context of the whole discourse. Finally, we validate our theory by analyzing examples from corpora of English,French, and Romanian. Introduction As originally postulated, Centering Theory (CT) (Grosz, Joshi, and Weinstein (1995)) accounts for references between adjacent units but is restricted to local reference (i.e., within segment boundaries). Recently, CT-based work has emerged which considers the relation of global discourse structure and anaphora, all of which proposes extensions to centering in order to apply it to global discourse. We approach the relationship between global structure and anaphora resolution from a different, but related, perspective. We identify domains of referential accessibility for each discourse unit over discourse structure trees such as those defined in Rhetorical Structure Theory (RST ; Mann and Thompson (1987)) and show how CT can then be appfied to global discourse b~( using these domains. As such, our approach differs from Walker's (1996), whose account of referentialit), within the cache memory model does not rely on discourse structure, but rather on cue phrases and matching constraints together 281 with constraints on the size of the cache imposed to reflect the plausible limits of the attentional span. Our approach is closer to that of Passonneau (1995) and Hahn and Strtibe (1997), who both use a stack-based model of discourse structure based on Grosz and Sidner's (1986) focus spaces. Such a model is equivalent to a dynamic processing model of a tree-like structure reflecting the hierarchical nesting of discourse segments, and thus has significant similarities to discourse structure trees produced Ib/o RST (see Moser and Moore (1996)). wever, using the RST notion of nuclearity, we go ,beyona previous work by revealing a 'hidden' structure in the discourse tree, which we call veins, that enables us to determine the referential accessibility domain for each discourse unit and ultimately to apply CT globally, without extensions to CT or additional data structures. In this paper, we describe Veins …

Content truncated to 500 words