A Goal-Based Grammar of Rhetoric 
Chrysanne DiMarco, Graeme Hirst, and Marzena Makuta-Giluk * 
University of Waterloo 
and 
University of Toronto 
May 28, 1993 
1 An advocacy of rhetoric 
In many applications in computational linguistics, such as machine translation, we want to un- 
derstand why a writer has used language in a particular way, what specific effects she intended 
to convey, and which linguistic choices were made to achieve those goals. This idea, that a writer 
has specific communicative goals or intentions in mind when she composes text, goes back to the 
study of classical rhetoric, in which the rhetorical form of a writer's text was considered to reflect 
his thoughts and intentions. The rhetorical form was realized through the appropriate choice and 
organization of words and syntactic structures meant to "evoke the desired response" (Corbett 
1971). Content and form were recognized as inseparable, and as exerting a reciprocal influence 
upon each other. 
In our work, we have integrated the theory of classical rhetoric with conteml)orary work in 
knowledge rel)resentation to develop an approach to relating high-level intentions, through a hi- 
erarchy of increasingly detailed linguistic knowledge, to specific syntactic choices. We have con- 
structed a multi-level grammar of rhetoric that ties rhetorical goals, such as formality and force, 
to syntactic choices, through the intermediary of stylistic goals such as clarity and concreteness. 
When a writer sets out to produce a piece of text, he does so with a specific pragmatic goal, 
such as informing or persuading, in mind. In order to achieve this goal, he must make particular 
linguistic choices. The mechanism that we propose for achieving the desired pragmatic effect is 
the use of lower-level rhetorical goals. For example, if the writer wishes to be I)ersuasive, he must. 
present the argument in such a way as to make the reader receptive to his message. Thus, he 
might use a combination of the rhetorical goals of force and amity to convey both a sense of 
authority and a desire to communicate. He must then decide on the specitic linguistic choices 
that will realize these rhetorical goals. Style is the medimrn that enables him to do this. In order 
to model formally the linguistic choices of a writer, we decompose complex rhetorical goals into 
simpler stylistic goals that can then be related to lexical, syntactic, and semantic choices. 
*l.)iMarco and Makuta-(iiluk: I)epartnmnt of C, onlputer Science, University of Waterloo, Waterloo, ()ntari*J, 
(',anada N21, 3(i;I; e-mail: cdimarco(o)logos.uwaterloo.ca; mlmlakuta(@logos.uwaterloo.ca. Ilirst: \[)epari.nmnt of 
(',OUnl)U(.er Science, I lniverslty of Toronto, Toronto, Ontario, Canada M5S 1A4; e-mail: gh(t~cs.|.orontto.edu. 
15 
2 A multi-level grammar of rhetoric 
()ur grammar of rhetoric allows us to recognize, at the top level, how the overall structure of a text 
works to achieve a certain communicative goal, and, at the lower level, how the individual pieces 
of text fit, together to produce subtle stylistic effects. Thus, through multiple levels of abstraction, 
we tie together rhetorical forms characteristic of high-level intentions and syntactic relationships 
associated with stylistically significant effects. Thus, we have four levels, the pr, gm~ttir, the 
rht'loric~d, the stylistic, and the syntttetic. 
In DiMarco and Hirst (1993), DiMarco et al (1992), Green (1992), and tloyt (forthcoming), we 
describe the construction of a syntactic stylistic grammar that relates stylistic goals to abstract 
stylistic properties, and then relates these abstract properties to low-level syntax. The foundations 
of the grammar draw on the work of Halliday (1985) and Halliday and Hasan (1976); we consider 
the primitive stylistic effects of a sentence to be correlated with its underlying cohesive and hi- 
erarchical syntactic structure. We assign each type of sentence component a primitive-element 
cl;L~sitication on the basis of the nature of these fundamental structures. 
We then compose these primitive stylistic elements into higher-level, abstract stylistic proper- 
tit.s,or, elements. In I)iMarco and Hirst (1993), we proposed a set of stylistic terms that made 
¢~xplicit these kinds of abstract properties. Here are three examples: 
Het(¢ropoise: A sentence in which one or more parenthetical components are syntactically 'de- 
tached' and dissimilar from the other components at the same level in the parse tree. 
C(,,ntroschematic: A sentence with a central, dominant clause with one or more of the follow- 
ing optional features: coml)lex phrasal subordination, initial dependent clauses, terminal 
(lel)etl(lent clauses. 
R.(~.s.hation: A shift in stylistic effect that occurs at the end of a sentence and is a move front a 
relative discord (an incongruity) to stylistic concord (normal usage). 
Each abstract stylistic element is defined as a composition of primitive stylistic elements. 
Next, the abstract stylistic elements are composed into stylistic goals. Stylistic goals, such 
as clarity and concreteness, are elusive qualities that were traditionally defined by stylists only 
I)y means of examples and informal rules. However, with our grammar, we can abstract from a 
plethora of low-level syntactic rules that stylists have used and can define formal rules for specitic 
stylistic goals. 
For example, the goal of concreteness is associated with heteropoise, a stylistic element that 
char;u:terizes the kinds of cohesive (and non-cohesive) syntactic interruptions that create forms of 
stylistic emphasis. In the grammar, concreteness is defined as various forms of stylistic highlighting, 
either emphasis (heteropoise, dissolution) or deviation from established usage (discord). 
An example of a concrete sentence that is a stylistic heteropoise, beginning with a canonical 
structure, but then emphatically interrupted, is this: Your writing, if I may say so without offence, 
is immatu~v. 
3 The level of rhetorical goals 
Ill Makuta-(~iluk (1991) and Makuta-Giluk and DiMarco (1993), we describe the development 
of a rhetorical grammar that is built upon our stylistic grammar and composes rhetorical goals 
from combinations of stylistic goals. Where goal-directed style accounts for the stylistic choices 
that will express a certain effect, goal-directed rhetoric considers the higher-level linguistic choices 
16 
associated with specific rhetorical effects that also express the communicative goals of a text. 
Thus, we have formalized some of the syntactic aspects of the rhetorical structure of texts. 
There rnay be many reasons why an author writes a text and why she chooses to express it in 
a particular form. These reasons are pragmatic goals (of. Hovy (1988)). Informing or persuading 
are examples of such goals. Once a writer commits herself to a specific pragmatic goal, she must 
determine apl)ropriate content and linguistic realization. Both these issues involve a number of 
rhetorical options. A possible mechanism for achieving tile desired pragmatic effect is using a set 
of rhetorical goals. Once the writer knows which rhetorical effects she wants in tim text., she now 
needs to make the choices that realize these goals. We decompose complex rhetorical goals into 
simpler entities, the stylistic goals described above, such as simplicity or clarity. In Makuta-(~iluk 
(1991) and Makuta-(iHluk and DiMarco (1993), we pointed out that a stylistic goal can be used 
to realize more than one rhetorical goal, and a rhetorical goal can be achieved in several different 
ways. The author must choose a set of interrelated stylistic goals to realize her particular set 
of rhetorical goals. The stylistic goals taken separately do not determine the rhetorical effect of 
the text; it is their interplay that makes it possible to express a wide range of rhetorical goals. 
In our gramrnar, we have defined goals such as force/ineffectiveness, formality/informality, and 
amity/distauce. For example, we define force to be characteristic of sentences that display one ()f 
tile following combinations of stylistic goals: 
force 
directness and conciseness and concreteness 
dynamism and simplicity 
An example of the first kind of forceful sentence is: Frankly, my dear, I do)l't give a da)lt)t. 
We associate ineffectiveness, the dual of force, with diffusive and garrulous communication: 
ineffectiveness 
complexity 
obscurity and verbosity 
obscurity and obliqueness 
Tile following example of ineffectiveness is from a textbook of rhetoric: The sequence of develop- 
me)tt is fortuitous and even implausible, for the treatment of rhetoric becomes more perfunctory as 
eruditiou in the works of rhetoriciaus increases, and rhetoric disappears abruptly when knowledge 
of it is at a maximum, particularly from the works of the authors who acknowledge the influe)~cc 
of Cicero and Quintilian. We define other rhetorical goals in an analogous manner. 
4 Multiple levels of rhetorical relations 
The notion of multiple levels of rhetorical analysis is intrinsic to our formalization: communicative 
goals are represented at several levels of abstraction, and each level is composed of elements from 
the level below. We believe that this idea of stratified levels is applicable not only to syntactic 
aspects of rhetoric, but to lexical and semantic aspects as well, and have begun to apply our 
approach to studying \]tow lexical choices realize particular intentional goals (DiMarco, Hirst, and 
Sl,c(lc 1993). 
17 
Eventually, we see lexis, syntax, and semantics being represented by separate primitive-level 
r(q)resentations that act together to determine the realization of communicative goals at the stylis- 
tic, rhetorical, and pragmatic levels; our formalism is therefore both stratified and branching. 
Thus, the ways in which intentional relations interact with ideational, or informational, relations 
(Moore and Pollack 1992) can be accounted for nicely by our model of rhetoric, which integrates the 
ell~cl,s of lexis, syntax, and semantics on rhetorical strncture within a single cohesive framework. 
5 Conclusion 
'lb construct a fidl computational theory of rhetoric, we will need to first develop complete formal- 
izations of the lexical and semantic aspects of style, and then integrate these representations with 
our current syntactic grammar. The syntactic theory formalizes some significant aspects of style 
and rhetoric. While it has limitations, it does explain several aspects of the writing process and 
how intentional goals can be realized through several interrelated levels of rhetorical and stylistic 
goals. 

References 
(:orbctt, Edward F).J. Classical rhetoric for the modern student. Oxford: Oxford University Press, 
1971. 
DiMarco, Chrysanne, and Hirst, Graeme. "A computational theory of goal-based style in syntax." 
To appear in: Computational Linguistics, 1993. 
l)iMarco, Chrysanne; Hirst, Graeme; and Stede, Manfred. "The semantic and stylistic differenti- 
ation of synonyms and near-synonyms." Proceedings, AAA I Spring Symposium on Building 
Lexicons for Machine Translation, Stanford, March 1993. 
I_)iMarco, Chrysanne; Green, Stephen J.; Hirst, Graeme; Mah, Keith; Makuta-Giluk, Marzena; 
and Ryan, Mark. "Four papers on computational stylistics." Research report CS-92-35, 
Faculty of Mathematics, University of Waterloo, June 1992. 
Green, Stephen. A fu~zctional theory of style for natural language generation. MMath thesis, 
Department of Computer Science, University of Waterloo, 1992 \[published as research report 
CS-92-48, Faculty of Mathematics, University of Waterloo\]. 
Italliday, M.A.K. Ai~ introduction to functional grammar. London: Edward Arnold, 1985. 
Halliday, M.A.K., and Hasan, Ruqaiya. Cohesion in English. London: Longman Group Limited, 
1976. 
liovy, Eduard H. Generating natural language under pragmatic constraints. Hillsdale, N.I: Lawrence 
Erlbaum Associates, 1988. 
Hoyt, Pat. An efficient functional-.based stylistic analyzer. (working title) MMath thesis, Depart- 
ment of Computer Science, University of Waterloo, forthcoming. 
Makuta-(liluk, Marzena. A computational rhetoric for syntactic aspects of text. MMath thesis, 
Department of Computer Science, University of Waterloo, 1991 \[published as research report 
CS-91-54, Faculty of Mathematics, University of Waterloo\]. 
Makuta-Gihlk, Marzena and DiMarco, Chrysanne. "A computational formalism for syntactic 
aspects of rhetoric." Proceedings, First ConfeTvnce of the Pacific Association for Computa- 
tional Linguistics, pages 63-72, Vancouver, Canada, April 1993. 
Moore, Johanna D. and Pollack, Martha E. "A problem for RST: The need for multi-level discourse 
analysis." Computational Linguistics, pages 537-544, Volume 18, Number 4, December 1992. 
