Maite Taboada
Simon Fraser University

Maite Taboada

Simon Fraser University
Canada

Finding discourse relations

In this talk, I will do two things. First, I will discuss the space that discourse or rhetorical relations occupy in language, and the issue of consensus on a common taxonomy of relations. Second, I will review the issue of signals for coherence relations and describe our corpus annotation of a broad set of signals.

First, in terms of the space that rhetorical relations occupy and their classification, I propose a top-down approach, that is, one that views relations between propositions in discourse as relations that help create coherence. I will review different approaches to rhetorical, coherence and conjunctive relations, and explain where Rhetorical Structure Theory (Mann and Thompson, 1988) fits in with other proposals. Coherence is part of texture, and thus related to entity-based coherence or cohesion (Halliday and Hasan, 1976) and to general properties of discourse. I will argue that there is a cline of grammaticalization of rhetorical relations, from discourse to syntax, and that differences across theories are sometimes rooted in where in that cline the theory positions itself. For instance, RST is at the end of the cline closer to discourse, and does not make strong claims about the syntactic realization of rhetorical relations. The conjunctive relations of Halliday and Hasan (1976) and Martin (1992), on the other hand, are more clearly syntactic, and have lexical elements as signals of the relation. My optimistic view is that, in this broad space of rhetorical relations, we can map relations across different theories if we bear in mind that they may be more or less abstract versions of each other.

In the second part of the talk, I will discuss signalling. In this sense of rhetorical relations as relations of coherence, the relations are present whether signalled by a particular device or not. This is the long-held view within Rhetorical Structure Theory. The concern in RST has been to explain how coherence, and the impression of coherence, is achieved when relations are apparently not signalled. Signalling has traditionally been taken to refer to conjunctions or discourse markers which link propositions. I will propose that signalling is actually quite prevalent, if we broaden our definition of signalling devices. I will report on the results of our annotation (Das and Taboada, 2018a, 2018b) of the RST Discourse Treebank (Carlson et al., 2002), which shows that the vast majority of relations are signalled by at least one device, according to our annotation of the RST-DT, available through the Linguistic Data Consortium (Das et al., 2015). I will describe the annotation process, the taxonomy of signalling devices, and will provide detail on the types of signalling devices found for various relations.