Rhetoric as Knowledge 
Owen Rambow* 
University of Pennsylvania and CoGenTex, Inc. 
A proper assessment of the relation between discourse structure and speaker's communicative inten- 
tions requires a better understanding of communicative intentions. This contribution proposes that 
there is a crucial difference between intending the hearer to entertain a certain belief (or desire, or 
intention), and intending to affect the strength with which the hearer entertains the belief (or desire, 
or intention). Rhetoric, if defined as a body of knowledge about how discourse structure affects the 
strength with which a discourse participant entertains beliefs, desires, and intentions, can be seen 
to play a precise and crucial role in the planning of discourse. 
1 Communicative Intentions 
It is commonly assumed that in engaging in communication, tile speaker 1 (S) has a specific com- 
municative intention. Communicative intentions relate to that domain that the use of language can 
affect, namely the mental state of the hearer (H), i.e., H's beliefs, desires, and intentions. Of course, 
S may have other intentions that relate to tile world at large (such as to get H to open a window), 
but these are not properly speaking eonnnunicative intentions: they can only be achieved by use of 
language if language first produces some appropriate change in H's mental states. 
In discussing the ways in which S's use of language can affect H's mental states, it is important 
to make a distinction between the fact that H is entertaining a belief, desire, or intention, and the 
degree or strength with which it is entertained. This distinction has not always been made in the 
text generation literature; presumably, this is because for the types of texts whose generation has 
been studied, namely reports, documentation, and manuals of various kinds, this distinction is not 
relevant. This is because these texts (weather reports, military reports, instruction manuals, and so 
on) are "authoritative text.s": if the text makes H entertain a certain belief or intention, then H will 
do so with a sufficient strength to satisfy S's communicative goals. 
I claim that the interesting aspects of rhetorical relations (or "rhetoric" for short), and the intersting 
role that rhetoric can play in discourse generation, have not been studied because the types of text 
analyzed and generated have been one-sided in significant respects. For the sake of concreteness, let 
us assume the following definition: 
*Department of CIS, University of Pennsylvania, Philadelphia PA 19104. ramb0u@unagi.cis.upenn.edu. This 
work was partially supported by the following grants: ARO DAAL 03-89-C-0031; DARPA N00014-90-J-1863; NSF 
ll~I 90-16592; and Ben Franklin 91S.3078C-I. 
! I use the terminology fi'om spoken language; these remarks apply equally well to written language. 
102 
Rhetoric is S's knowledge of how text structure manipulates tile strength of beliefs, 
desires, and intentions already entertained by H. 
In order to argue that this definition makes sense, I will discuss three questions: is rhetoric necessary, 
is it trivial, and is it useful? 
2 Is Rhetoric Necessary? 
More precisely, I will discuss whether S must necessarily have access to rhetoric in order to com- 
nmnicate effectively. S can only affect H's mental state through the use of a text (the sequence of 
utterances which constitute an act of communication), which means that it is in the act of decoding 
the text that H is affected. Thus S's goal is that H be affected by the text. In order to achieve this 
goal, S nmst know about the process of decoding that I1 will use, otherwise S carmot purposefully use 
language. Does text structure contribute to affecting H? Consider the following discourses (uttered 
in a context where S knows that H does not like paying taxes under any circumstance). 
(1) Vote for Bush. Clinton will raise taxes. Bush will not raise taxes. 
(2) Clinton will raise taxes. Vote for Bush. Bush will not raise taxes. 
(2) is arguably less convincing than (1) since tile the two sentences about tim candidates' attitude 
towards taxation are not juxtaposed, thus depriving them of their contra.stive force. Since text 
structure participates in affecting H, then S must have knowledge about tile mechanism; i.e., S must 
have access to rhetoric. 
3 Is Rhetoric Trivial? 
It has often been observed that many of tile definitions of rhetorical relations are tautologous. 
Consider, for example, RST's definition of EVIDENCE \[Mann and TholnpSOlL 1987, p.10\]. 'Nuc' 
refers to the nucleus, that for which evidence is being contributed, and 'Sat' refers to the satellite, 
tile evidence. Nuc and Sat, are assumed to be juxtaposed. 
relation name: 
constraints on N: 
constraints on S: 
constraints on the 
N+S combination: 
the effect: 
EVIDENCE 
H might not believe Nuc to a degree satisfactory to S 
The reader believes Sat or will find Sat credible. 
H's eonlprehending Sat increases H's belief of Nuc 
II's belief of Nuc is increased 
Tile effect is a mere restatement, of the constraint on tile combination of Nuc and Sat - the definition 
of the relation does not seem to be very informative. But consider a possible discourse relation 
COUNTER-EVIDENCE: 
103 
relation name: 
constraints on N: 
constraints on S: 
constraints on the 
NTS combination: 
the effect: 
COUNTER-EVIDENCE 
H might believe Nuc more strongly than S intends H to 
The reader believes Sat or will find Sat credible. 
It's COlnprehending Sat decreases H's belief of Nuc 
It's belief of Nuc is decreased 
The definition of COUNTER-EVIDENCE is entirely similar to (and as tautologous as) that of EV- 
IDENCE, but it does not yield a coherent discourse. Consider the following, uttered in the same 
context as (1) and (2) above: 
(3) Clinton will lower taxes. No democratic president has ever lowered taxes. 
There is no coherent reading for (3) under which S wants to decrease H's belief in the first sentence. 
In order to achieve this effect, S needs to explicitly negate the first clause, which results in an 
EVIDENCE relation. 
(4) Clinton will not lower taxes. No democratic president has ever lowered taxes. 
Thus, the contribution of rhetoric lies in detailing what sort of effects (on It's beliefs, desire, and 
intentions) mere juxtaposition can achieve. The fact that not all possible effects can be achieved 
without explicit lexical and syntactic means is what makes rhetoric an important body of knowledge 
in discourse processing. 
4 Is Rhetoric Useful? 
More specifically, can rhetoric be used in discourse planning? From what has been said, it would 
follow that rhetoric is essential in discourse planning, since without rhetoric, S would have no 
knowledge of the effects of the discourse structure on H. But can rhetoric, as defined here, in fact 
be used for discourse planning? Text planning architectures using knowledge about the relationship 
between communicative intentions and the juxtaposition of discourse segments have been developed 
at ISI \[Hovy, 1988\]. 1 conjecture that tiffs type of architecture is well suited for planning texts (or 
those aspects of texts) that manipulate the strength with which H entertains beliefs, desires, and 
intentions. Such a plalmer would require a more sophisticated representation of mental states: 
1. A logic of desire and a logic of intentions are required. Tile logic of desire would most likely 
he a modal logic; the logic of intentions wouhl, presmnably, require a representation of action. 
2. For beliefs, desires, and intentious, the strength with which they are entertained must also be 
represented. Strength can be represented by discrete indications of quality, as has been done 
by \[Walker, 1993\] in conversational models. 
Ideal applications of such a planner would include texts whose primary goal is to increase H's desire 
to perform a certain action, such as advertising texts of tile types given above. Furthermore, dialog 
104 
planning is more likely to yield interesting instances of the use of rhetoric, since in monologic genres 
the modeling of the strength of H's beliefs, desires, and intentions must remain conjectural, while in 
dialogic genres, H's feedback can contribute to S's assessment of H's mental state. 
Rhetoric is not useful for texts whose goal is mainly or exclusively to convey information, and for 
which H is assumed to stroTtgly believe any proposition the text makes H entertain. This is typically 
the case for reports, including those generated by the original ISI planner. The use of a rhetoric- 
based top-down planner for planning such texts cannot be successful \[Kittredge et al., 1991\], since 
here the decomposition of the communicative goal nmst refer to the domain structure, and therefore 
is too unconstrained to be handled by a domain-independent body of knowledge. 
5 Types of Intentions, Types of Relations, Types of Texts 
The distinction that has been made throughout this paper, between making H entertain a be- 
lief, desire, or intention, and increasing the strength with which it is entertained, corresponds to 
the one between intentional/presentational and informational/subject matter relations made by 
\[Mann and Thompson, 1987\], and recently discussed by \[Moore and Pollack, 1992\]. Rhetoric relates 
to the intentional level - note that the effect of RST's presentational relations are all formulated 
in terms of increasing the strength of some aspect of H's mental state, while the subject matter 
relations all have the effect of making H entertain ("recognize") new beliefs. Texts that do not have 
a meaningful intentional level, such as reports, cannot be planned using a rhetoric-based planner. 
Texts that only have an intentional level of structure, such as (1) above, are ideally suited for such 
planners. However, as Moore and Pollack argue, many texts simultaneously have both types of 
structure. For such texts, new planning architectures must be found. They will require a better 
understanding of comlnunicative goals along the lines argued for in this paper. 

References 
\[Hovy, 1988\] ltovy, E. H. (1988). Planning coherent multisentential text. Ill 261.h Meeting of lhe 
Associaliou for Compulalional Liuguislics (ACL '88), pages 163-169, Buffalo. ACL. 
\[Kittredge et al., 1991\] Kittredge, R., Korelsky, T., and Rainbow, O. (1991). On the need for domain 
communication knowledge. Compulational Intelligence, 7(4). 
\[Maim and Thompson, 1987\] Mmm, W. C. and ThonH~son, S. A. (1987). Rhetorical Structure The- 
ory: A theory of text organization. Technical Report ISI/RS-87-190, ISI. 
\[Moore and Pollack, 1992\] Moore, J. D. and Pollack, M. E. (1992). A problem for RST: The need 
for multi-level discourse analysis. Computational Linguistics, 18(4). 
\[Walker, 1993\] Walker, M. (1993). lnfornaation and deliberation in discourse. In this volume. 
