Says who? On the treatment of speech attributions in discourse structure

In this paper, we present what we think is an elegant solution to some problems in the discourse-structural modelling of speech attribution. Using mostly examples from the Wall Street Journal Corpus, we show that the approach proposed by Carlson and Marcu (2001) leads to irresolvable dilemmas that can be avoided with a suitable treatment of attribution in an underspecified representation of discourse structure. Most approaches to discourse structure assume that textual coherence can be modelled as trees. In particular, it has been shown that coherent discourse follows the so-called rightfrontier constraint (RFC), which essentially ascertains a hierarchical structure without crossed dependencies. We will discuss putative counterexamples to these two assumptions, most of which involve reported speech as in (1) (cited in Wolf and Gibson 2005): (1) “Sure I’ll be polite,” promised one BMW driver who gave his name only as Rudolph. “As long as the trucks and the timid stay out of the left lane.” In (1) the second part of the quote should be linked to the first part (and not the whole first sentence) by a condition relation. If we were to analyse the parenthetical speech reporting clause (“promised one BMW driver ...”) as the nucleus of its host clause (i.e., the quote), the RFC would prevent linkage between the two parts of the quote. If the attribution is analysed as a satellite of the quote, as in Carlson and Marcu (2001), Wolf and Gibson argue, it should be a satellite to both parts of the quote, thus violating treeness. In this paper, we will explore the problems arising from this type of construction and propose a treatment of speech report attributions that we will argue allows us to preserve both, treeness and the RFC in building discourse structures. 1 The (non-)treatment of speech attribution in classic Rhetorical Structure Theory (RST) In ‘classic’ Rhetorical Structure Theory (RST; Mann and Thompson 1988), the problems in accommodating speech report attributions do not arise, because classic RST does not separate complements of verbs and parenthetical speech reporting clauses from their host clause. Leaving speech attribution implicit is in line with the general ‘philosophy’ of RST, which aims to represent not all possible links, but the most plausible structure licensed by 1 Markus Egg is now at Humboldt University, Berlin. This manuscript is a slightly updated (2009) version of our paper in the Proceedings of the Workshop on Constraints in Discourse, Maynooth, Ireland 2006 (http://www.constraints-in-discourse.org/cid06/).