Testing SDRT’s Right Frontier

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1–9, Beijing, August 2010 Testing SDRT’s Right Frontier Stergos D. Afantenos and Nicholas Asher Institut de recherche en informatique de Toulouse (IRIT), CNRS, Université Paul Sabatier {stergos.afantenos, nicholas.asher}@irit.fr Abstract The Right Frontier Constraint (RFC), as a constraint on the attachment of new con- stituents to an existing discourse struc- ture, has important implications for the in- terpretation of anaphoric elements in dis- course and for Machine Learning (ML) ap- proaches to learning discourse structures. In this paper we provide strong empirical support for SDRT’s version of RFC. The analysis of about 100 doubly annotated documents by five different naive annota- tors shows that SDRT’s RFC is respected about 95% of the time. The qualitative analysis of presumed violations that we have performed shows that they are either click-errors or structural misconceptions. 1 Introduction A cognitively plausible way to view the construc- tion of a discourse structure for a text is an incre- mental one. Interpreters integrate discourse con- stituent n into the antecedently constructed dis- course structure D for constituents 1 to n − 1 by linking n to some constituent in D with a dis- course relation. SDRT’s Right Frontier Constraint (RFC) (Asher, 1993; Asher and Lascarides, 2003) says that a new constituent n cannot attach to an arbitrary node in D. Instead it must attach to ei- ther the last node entered into the graph or one of the nodes that dominate this last node. Assuming that the last node is usually found on the right of the structure, this means that the nodes available for attachment occur on the right frontier (RF) of the discourse graph or SDRS. Researchers working in different theoretical paradigms have adopted some form of this con- straint. Polanyi (1985; 1988) originally pro- posed the RFC as a constraint on antecedents to anaphoric pronouns. SDRT generalizes this to a condition on all anaphoric elements. As the at- tachment of new information to a contextually given discourse graph in SDRT involves the reso- lution of an anaphoric dependency, RFC furnishes a constraint on the attachment problem. (Webber, 1988; Mann and Thompson, 1987; 1988) have also adopted versions of this constraint. But there are important differences. While SDRT and RST both take RFC as a constraint on all discourse at- tachments (in DLTAG, in contrast, anaphoric dis- course particles are not limited to finding an an- tecedent on the RF), SDRT’s notion of RF is sub- stantially different from that of RST’s or Polanyi’s, because SDRT’s notion of a RF depends on a 2- dimensional discourse graph built from coordinat- ing and subordinating discourse relations. Defin- ing RFC with respect to SDRT’s 2-dimensional graphs allows the RF to contain discourse con- stituents that do not include the last constituent entered into the graph (in contrast to RST). SDRT also allows for multiple attachments of a con- stituent to the RFC. SDRT’s RFC has important implications for the interpretation of various types of anaphoric …

Content truncated to 500 words