On the necessity of intentions and (at least) the usefulness of rhetorical 
relations: A position paper 
Vibhu O. Mittal and C~cile L. Paris 
Inh~rmation Sciences Institute and Department of Computer Science 
University of Southern Calitt~rnia 
4676 Admiralty Way, Marina del Rey, CA 90292 
{mittal, paris} @isi.edu 
In this position paper, we argue Ibr the need for a generation system to represent communicative goals 
(i.e., goals that express the speaker's intent to aflect the bearer's mental attitudes), at least if certain tasks 
are to be achieved. We lurther argue that, wlfile rhetorical relations might be recoverable ttom other 
Iactors, they appear to be a useful level of abstraction to maintain about the discourse structure to avoid 
costly reasoning. However, we believe a source of conliasion comes Ii'om the Ihct that the term 'rhetorical 
relations' encompasses various diftbrent constraints about a discourse structure, and that there is a need to 
be clear about what a 'rhetorical relation' is. 
1 Need for communicative goals 
Typically, a speaker has some goal in mind when producing an utterance. Tiffs can be a very concrete 
goal such as "describe an object", "motivate the user to perlt~rm an action", or a much vaguer one, such as 
discussing an event or simply providing information about a subject. We believe that communicative goals 
should be represented explicitly in a generation system Ik~r several reasons: 
• To select the facts to be expressed, as typically, one does not utter everything one knows, and one's 
comnmnicative intentions influences strongly what is IO be said. 
• To determine the ordering of lacts to be expressed, in the event thai these facts have been determined 
ahead of time. Indeed, typically, several orderings will be possible, but only a subset of these will be 
appropriate Ik~r the speaker's intention(s). 
• To handle dialogue (feedback): it has already been shown that a record of the rhetorical relations 
holding in an utterance is not enough to be able to participate in a dialogue (specifically to recover from 
communication failures) \[6\]. This is because, as there is not always a one to one relationship between 
rhetorical relations and communicative goals, communicative goals cannot always be recovered Iiom 
a record of the rhetorical relationships and the communicative goal is necessary to handle fi~llow-up 
questions. 
• To take into account any non-textual phenomena: li.~r instance, there may be many ways to elaborate 
on a specific aspect in a presentation that includes both text and graphics: the specific manner of 
elaboration will depend upon the communicative goal to be achieved (e.g., providing inli:~rmation 
about the position of a knob in the control panel may result in a picture of the whole panel, while 
providing inlk~rmation in order to identity the knob may result in a descriptive phrase such as "the 
hexagonal knob"). 
We thus argue that communicative goals are important to represent in a generation system and are an 
essential part of discourse structure (which is not to say they are suMcient, however). 
Some researchers have argued that colnmunicative goals are not necessary l~lr coherent text generation, 
and have cited architectures such as schemas and systems based on "domain knowledge" to illustrate their 
point. In our view, this simply illustrates how one can compile out the intentional knowledge necessary 
90 
to achieve coherent discourse within a specific domain and context, not that intentions are not an essential 
part of the discourse structure. It is important to realize that, while Ibr some generation tasks, schemas that 
compile out intentional information are enough, systems based on such generation schemes are then unable 
to handle other discourse phenomena such as dialogue. 
Finally, some researchers have conflated the notion of communicative goals with that of rhetorical relations, 
arguing that it is sufficient to reason about rhetorical relations that hold between certain facts to construct a 
coherent text, and Ihrther arguing that the choice of a rhetorical relation indicates a specific communicative 
goal. We do not agree with these arguments, however, at least not with respect to the current theories of 
rhetorical relations. Indeed, in general, there is not a one-to-one mapping between intentions and rhetorical 
relations \[6\]. Consider lbr example the Ibllowing dialogue, taken from \[6\]: 
S: (1) Remove the cover. (2) You'll need the Phillips screwdriver. (3) It's in the top drawer of 
the toolbox. Do you have it'? 
H: No. 
Using an RST-type analysis \[4\], (3) is a CIRCUMSTANCE to (2). In terms of intentions, (2) is uttered 
in an attempt to make the hearer idenli\[y and find the screwdriver. Suppose the hearer is not able to 
find the screwdriver, because he or she cannot identit3, it. At this point, the speaker needs to recognize 
that the intention to make the hearer identiI3, the screwdriver has not been achieved. However, the 
speaker cannot recover this intention simply from knowing that there is a CIRCUMSTANCE relation between 
(2) and (3). CIRCUMSTANCE could have been used to achieve a variety of other intentions. Similarly, 
the communicative goal could have been achieved using other rhetorical relations such as, Ior example, 
ELABORATE-0BJECT-ATTRIBUTE. (For more details, see \[6\]). 
2 Rhetorical relations 
Doing text analysis, researchers have tbund that coherent texts tend to exlfibit patterns in their organization, 
with a limited set of relationships between their components. As a result, people have attempted to 
characterized these relationships in terms of "rhetorical relations" (or schemas of rlaetorical predicates). 
Typically, the relationships identified above arise because of various filctors, including: 
• There is a semantic relation that llolds between tile various semantic units (or sets of semantic units 
taken as a whole) - for example, given the Ihcts: <John has a black car> <John's car is a Honda>, 
since 'being a Honda' is an attribute of John's car, an item introduced in the first fact, we could say: 
"John has a black car. It is a Honda" and draw an "elaborate-object-attribute" between the two clauses. 
• Intentions also sometimes have a canonical clustering: tot example, the goal to provide evidence 
Ibr a claim often appears when a claim is given. In the resulting text, the relation "evidence" can be 
drawn between the two spans of text that result IYonl tile two intentional goals. 
• A recognized pattern of thematic progression is present in the text, giving rise to what has been called 
to a "textual relation" between units of text. 
Essentially, these standard patterns (or standard sets of constraints) have been given names (the names 
of the "rhetorical relations", e.g., "elaboration", "evidence", or "additive"). These labels seem usetul to 
perform certain types of reasoning on a text, such as the choice of syntactic realization (including deciding 
where a sentence break should occur, or the choice of a cue phrase), e.g., \[6,10\], without reasoning explicitly 
about the actual constraints that hold, and which the speaker wanted to lfighlighl. (Note that this is not to 
say that there is a clear unambiguous mapping between tile rhetorical relation and the resulting syntactic 
realization). 
91 
For example, if a system knows that a unit of lext U 1 is in PURPOSE relation to another unit U2, then 
the realization component can select an appropriate syntax It verbalize U 1 and U2 (e.g., "in order to..."), 
without re-reasoning about the constraints that hold between the inll~rmation presenled in U 1 and U2. 
Similarly, when several constraints hold and the speaker chooses to highlight one, tiffs can be indicated in 
the discourse structure by the appropriate preferred label, which then constrains the realization component 
to choose the appropriate syntactic structure allowing the hearer to recognize the prelerred interpretation. 
For example, consider the propositions "turn on the light" and "flip the switch" (taken lYom \[9\]). These 
propositions stand in several semantic relations with respect to each other: e.g., at least, CAUSE-EFFECT 
and PURPOSE. Given the speaker's intention to make the hearer know how to turn the light, depending, for 
example, on the focus the speaker wishes to have, several texts can be generated, each would highlight a 
different relation, and each would be analyzed with a ditl~rent RST-relation between the two clauses (e.g., 
"to turn on the light, flip the switch" or "flipping the switch will cause the light to be turned on"). If the 
discourse structure is annotated with the interpretation the speaker wishes to preI~r, the appropriate sentence 
can be generated. 
If, then, when generating text, one keeps track of these rhetorical relations to indicate the reasoning that 
has already taken place, the constraints that are already known to hold between various units, and which 
the speaker wishes to highlight, one avoids having to perl~rm that reasoning again, t~r example, to choose 
a cue phrase. While this might not be imporlant when the relation is a simple semantic one (e.g., <x> is the 
color of <y>) as tiffs check can be done easily, it is uselul (computationally speaking) when the reasoning 
that took place is complex (even on a semantic level, lot example, to ligure out that <x> is a cause of 
<y>, when no causal link exists in the knowledge base). It is also important when several relations exist, 
diflerent syntactic realization highlight different relations, and one was prelerred by the speaker. In this 
sense, current relations can be considered as 'macros' that represent a set of constraints that hold between 
various units (semantic units, textual units or intentions). As such, they might turn out to be an essential 
level of abstraction to represent as part of a discourse structure in a generation system to control realization. 
In general, however, there are serious problems with current theories of rhetorical relations such as RST. 
Several problems stand out: 
• It is not always clear what the label actually stands lbr and what exactly it is used tbr. The constraints 
they represent are not well defined, may even differ depending on tile surrounding context, and may 
encompass several different aspects of the discourse. 
• As already pointed out, these relations represent difl~rent types of constraints. This has also been 
discussed by other researchers \[ 1-3, 5-8\]. Conllating all these faclors by using one term "rhetorical 
relations" leads to contusion. 
• Few theories allow Ik~r several relations to hold at the same time. ~ Yet. semantic constraints, relations 
between communicative goals and textual relations usually co-occur \[2, 7\]. In fact, not only do they 
co-occur between two units of text, but they may also give rise to different non-isomorphic structures. 
However, all these relations (or constraints) contribute to generating an appropriate text and all should 
be represented and taken into account. Furthermore, several constraints of the same type might be 
present. 
• To do an analysis in terms of rhetorical relations such as RST, one is often implicitly also doing an 
intentional analysis. For example, consider the sentence: 
(1) if you cook tonight, (2) I'll take you out to the movies. 
I Notable exceptions ale theories based on Systemics Linguistics. e.g.. \[5\]. 
92 
Based purely on semantic ground, the sentences would be analyzed as having a condition between 
their two clauses. The other interpretation of MOTIVATION requires the analyzer to also do the 
intentional analysis. In truth, of course, both relations co-exist. (This is also discussed in \[7\].) 
3 Conclusions 
Communicative goals are necessary to build a generation system capable of both generating coherent dis- 
course and being able to respond to feedback and tbllow-up questions. Furthermore, rhetorical relations are 
a useful computational tool to represent constraints we currently don't totally understand, avoid duplicating 
reasoning from lirst principles, and provide an appropriate level of interPace with the realization component 
of a system. A theory of rhetorical relations, however, is far from complete and would gain l?om identifying 
more clearly what the relations stand lbr and how they are used. 

References 
HovY, E. H., LAVID, J. L., MATER, E., MtTrAL, V. O., AND PARTS, C. L. Employing Knowledge 
Resources in a New Text Planner Architecture. In Aspects of Automated Natttral Lzmguage Generation, 
R. Dale, E. Hovy, D. R6sner, and O. Stock, Eds. Springer-Verlag, Berlin, 1992, pp. 57-73. 
MAIER, E. A. The Extension of a Text Planner Ibr the Treatment of Multiple Links between Text Units. 
In Proceedings of the 4th European Worlc~hop on Text Generation (Pisa, Italy, 1993), pp. 103-114. 
MAIER, E. A., AND HOVY, E. H. Organizing Discourse Structure Relations using Meta-Functions. 
In New Concepts in Natural Language Processing, H. Horacek and M. Zock, Eds. Pinter Publisher, 
London, 1993, pp. 69-86. 
MANN, W. C., AND THOMPSON, S. A. Rhetorical Structure Theory: towards a functional theory of text 
organization. Text 8 (1988), 243-281. 
MARTIN, J. R. English Text: System and Structure. Benjamins, Amsterdam, 1990. (In Press). 
MooRE, J. D., AND PARIS, C. L. Planning Text lor Advisory Dialogues: Capturing Intentional, 
Rhetorical and Attentional Intormation, 1992. Technical Report lhml the University of Pittsburgh, 
Department of Computer Science (Number 92-22) and li'om USC/IS1, # RS 93-330. 
MOORE, J. D., AND POLLACK, M. E. A problem Ibr RST: The need for multi-level discourse analysis. 
Computational Linguistics 18, 4 (1992). 
REDEKER, G. Ideational and pragmatic markers of discourse structure. Journal of Pragmatics 14 
(1990), 367-381. 
SCOTt, D. Introduction to the Session on Intersegment Relatedness, 1993. Presented at the Burning 
Issues in Discourse Workshop, Maratea, Italy, 1993. 
ScoTr, D., AND DE SOUZA, C. Getting the Message across in RST-Based Text Generation. In Current 
Research in Natural Language Generation, R. Dale, C. Mellish, and M. Zock, Eds. Academic Press, 
Boston, 1990, pp. 47-74. 
