Rhetoric and Intentions in Discourse 
Robert Da.le 
rda©cogsci, edinburgh, ac. uk 
This brief note offers some views on the following questions: 
What is the precise relationship between rhetorical relations and intentions? Do they 
perform different (but possibility related) functions, or are rhetorical relations the re- 
alizations of intentions, or should rhetorical relations be discarded as a misconstrual of 
intentions proper? 
To make this more concrete, we'll focus on a specific pair of theories. We'll take Grosz and Sidner's 
Theory of Discourse Structures (henceforth GSDT) to be a good example of a theory whose notion 
of discourse structure is based on intentions; and we'll take Rhetorical Structure Theory (RST) to 
be a good example of a theory whose notion of discourse structure is based on rhetorical relations. 
On the face of it these theories are rather similar: both provide hierarchical characterisations of 
texts, and both seem concerned with the purpose of the text - for GSDT, the text realizes the 
intentional structure that underlies it; and for R.ST, the text is structured in terms of the rhetorical 
relations that hold between its parts. But these similarities are only skin deep. The kinds of 
trees used by the two theories are actually quite different: in GSDT, text can appear at both leaf 
nodes and internal nodes in the trees; whereas in RST, text can only appear at leaf nodes. So 
the structures might be isomorphic trader some mapping but that mapping remains to be worked 
out (at the very least, isonaorphism could only be achiew~d by annotating arcs iu an intentional 
structure tree with "rhetorical" labels; there is an interesting parallel here with the dilference between 
dependency grammar and phrase structure gramnaar). A more significant, difference is that, in RST, 
leaf nodes (the atomic text spans) are typically taken to be single clauses; but in GSDT, if our atomic 
elements (the bottom level discourse segments) were clauses, then, given the constraints the theory 
claims on pronominalization and reference, we'd rule out as anomolous a large space of perfectly 
legitimate pronominalizations (those where the antecedent and anaphor are in clauses at the same 
level of embedding in the discourse structure: recall that material in a sister discourse segment is 
inaccessible). Given this, it would be hard to argue that RST's text spans correspond to GSDT's 
discourse segments. 
This point is more important than it might: at first seem, since it draws attention to a relatively 
unnoticed difference between the two approaches: in RST, rhetorical relations are deemed to hold at 
all levels of discourse, including between individual clauses; whereas in GSDT, intentional relations 
can hold between individual clauses (precisely in those c~Lses where individual clauses correspond to 
individual discourse segments), but they need not do so - it's perfectly within the bounds of the 
5 
theory to have two clauses between which there is neither a DOM nor all SP relationship. This opens 
the possibility that there need not be an intention underyling each individual clause or sentence. 
RST's desire to postulate rhetorical relations all the way down to the level of clauses can be seen 
as tile root of a prol)lem that gives rise to another more frequenty voiced concern about the theory: 
that many of the so-called rhetorical relations are subject-matter relations, and very few are presen- 
tational. If, a.t the end of the day, we feel the need to postulate some relation between two sentences 
where the only obvious connection is that the event described by one follows the event described by 
the other, then it's hardly surprising that we feel the need to include a notion of SEQUENCE in 
our set of relations. But why stop there? Wc can ofl.en include information which might have been 
exl)rcssed in separate sentences using a noun phrase modifier instead: do we then want to say that 
there is a rhetorical relation between a head noun and each of its associated adjectives? I would 
suggest not; and that the first step we should take towards clearing up the confusion here is to banish 
those rhetorical relations which simply "mirror" underlying knowledge base relations. Doing this 
will also remove the problem that sometimes appears in the analysis of rhetorical structures, where 
the analyst is unsure which of several different but equally plausible relations should be postulated 
I)etweeu two elements: once we accept that all we are often really doing is identifying knowledge base 
relations, then it becomes unsurprising that two elements might equally well be related by means of 
a CAUSE relation or an EVIDENCE relation, since there can I)e any number of relations between 
two entities in the knowledge base (precisely because we assume there is no need for structures to 
be linea.r or tree-shaped: typically they are thought of as graphs). 
Taking this step will have two consequences. First, it will reduce substantially the number of textual 
relations we are willing to consider as rhetorical relations; and second, it will allow us to contemplate 
rhetorical structures which do not need to provide an analysis all the way down to individual clauses. 
This brings us closer to the thinking that lies behind GSDT, and answers some of its objections 
against RST; but we do not need to adopt as minimalist approach as GSDT when it comes to 
enumerating the possible relationships in a text.. Indeed, t.here is a sense in which GSDT falls 
foul of some of the same problems that affect RST: in particular, tieing the notion of intentional 
structure to the structure of the task that underlies a text leads us to postulate communicative 
intentions that mirror the goals and subgoals in the task itself, but this is little better than having 
rhetorical relations that mirror the underlying domain relations. A better way forward is to begin 
with those R,ST relations which are clearly *not* domain relations, such as MOTIVATION and 
JUSTIFICATION, and to elaborate their definitions in order to build towards a proper theory of 
communicative intention and purpose. 
