The Multifaceted Organization of Discourse 
Dan Suthers 
Learning Research and Development Center 
3939 O'Hara Street 
University of Pittsburgh 
Pittsburgh, PA 15260 
suthers+@pitt.edu 
The organization of discourse is multifaceted. Theories that take a single perspective on 
discourse structure, examining for example only goal satisfaction or textual coherence or ignoring 
the context in which a text was generated, will fail to identify the full functional significance of 
the content and organization of text. The task of generating coherent text to meet some commu- 
nicative intentions requires consideration of a number of constraints on content and organization 
that manifest in different ways in the end-product. Therefore the task of analyzing a text with 
respect to its coherence and intended functionality requires factoring distinct influences on the 
organization of the text itself. 
Diverse Accounts of Coherence 
Various accounts of discourse coherence have been offered in the literature, each with their 
own merits and limitations. For example, Grosz & Sidner \[1986\] defined a coherent discourse to be 
one in which a single discourse purpose is "shared bffy all the participants and when each utterance 
of the discourse contributes to achieving this purpose." Grosz and Sidner provided a clear account 
of certain aspects of the hierarchical and prerequisite structure of discourse. However they did 
not address the variety of communicative roles that sibling text segments can play with respect 
to each other. 
Intersegmental coherence is the concern of rhetorical structure theory (KST), whose propo- 
nents have stated (in various forms, and most directly by Hovy \[1988\]) that a coherent text is 
one in which each contiguous segment of text is related by a rhetorical relation. This claim is 
concisely stated, easy to understand and apply, and intuitively appealing. However it is deficient 
in being too weak in explanatory power and too strong in its conditions for coherence. It is 
too weak because the rhetorical relations themselves were developed by identifying relationships 
between contiguous segments of text in a corpus of texts considered to be coherent. Because 
of this circularity, the rhetorical characterization of coherence does not account for the sources 
of coherence or say why the text is coherent: the rhetorical relations are merely descriptive of 
coherent text. The claim is too strong because coherence does not necessarily require rhetorical 
relations between every contiguous segment of a text. The ordering of a text can be arbitrary at 
points. There is unlikely to be a well motivated rhetorical relation between those segments whose 
contiguity is arbitrary. This problem has shown up in text generation: Moore ~ Paris \[1991\] 
encountered certain difficulties in the design of their explanation planner that they attributed to 
the requirement that a rhetorical relation be present between each pair of contiguous segments 
in a plan. It.ST has also been criticized for conflating different kinds of relations that should be 
distinguished \[Grosz & Sidner 1986, Moore & Pollack 1993, Suthers 1990\]. 
Finally, Hobbs' \[1979\] characterization of coherence as the ease of inferring coherence rela- 
tions linking the parts of the text has also been criticized on the basis that the coherence relations 
serve no purpose once they have been inferred. The reader is more concerned with discerning 
128 
the communicative intent of a text than with verifying its coherence, and the theoretician should 
be as well. The relations by which coherence is defined should be independently motivated for 
reasons of communicative functionality. This view is not inconsistent with the view that co- 
herence relations must be recognized to succeed in making a text coherent. The former simply 
requires that coherence relations serve some communicative function other than being recognized 
as contributing to coherence. Hobbs' characterization can be restated as follows: a discourse is 
coherent to the extent that it makes its communicative functionality plain. However this leaves 
us with the question of precisely which functionally motivated relations are needed to account for 
coherence, and how these relations can provide a structural "test" for coherence that replaces the 
initial approximation provided by Hovy's interpretation of RST. 
Some Aspects of Discourse Structure 
My own work \[Suthers 1993a\] has been concerned with explanation as a question-answering 
activity in an instructional setting. I have found it necessary to distinguish the following structural 
aspects because each requires examination of a different source of information and/or application 
of a different kind of planning knowledge: 
1. Exchange Structure: The appropriateness of the dominating rhetorical acts I with respect 
to expectations created by prior exchanges. (Is the rhetorical act appropriate for the question 
just asked? Is the question being deferred?) 
2. Informative Structure (a special kind of goal satisfaction): The relevance of expressed 
information for performing an intended rhetorical act. (E.g., When does a causal, functional, 
or structural description constitute an answer to a question?) 
3. Propositional Structure (a.k.a. "informational" or "ideational"): The propositional 
inter-relatedness of references to different domain entities in a given communicative act. 2 
(What propositions are being expressed? How is each entity that is mentioned related to 
the topic?) 
4. "Contextual structure" in \[Suthers 1993a\] (but better understood in terms of tradeoffs 
between context-sensitive constraints on content choice rather than as structural): The 
appropriateness of the contents of a communicative act in the context of the knowledge 
available to the interlocutors and the knowledge expressed in prior dialogue. (E.g., does the 
utterance provide new information? Does it elaborate on an explanation or argument under 
construction?) 
5. Supplemental Structure: The ways in which one segment plays a supporting role with 
respect to another segment, such as facilitating comprehension or retention \[Zukerman 1990\]. 
Supplemental relations correspond roughly to those rhetorical relations where nuclearity is 
determined by intention rather than information. As Moore & Moser note in their workshop 
abstract, these provide different kinds of dominance. 
6. Sequential Structure: the consistency of the sequential presentation of the units of an ex- 
planation at all granularities with their intended communicative functionality. A partial 
ordering in terms of constraints derived from the above functional relationships \[Suthers 
1993b\]. 
tOne of 14 rhetorical iLlocutions (account, compare, define, describe, illustrate, ...) applied to a topic (entity 
reference or proposition). 
=A communicative i\]locution (assent, inform, dissent, request) applied to one or more propositions. 
129 
Based on this structural taxonomy, two concise (though oversimplified) conditions for a co- 
herent explanation can be stated: 
• The graph of informative, supplemental, and propositional relations between segments is a 
connected graph meeting the definitions of these relations (i.e., there are no isolated units of 
analysis, weaker than requiring a relation between each contiguous pair of units). 
• The sequencing of the segments is consistent with the intended communicative functionality 
of the segments, the functionality being discerned by the aforementioned relations \[Suthers 
1993b\]. 
(Given a more adequate treatment of exchange structure, the conditions could be extended to 
include the appropriateness of the dominating rhetorical acts for the exchange context.) This 
characterization of coherence addresses the weakness of the previous accounts by identifying the 
sources of coherence in independently motivated constraints on the content and organization of 
an explanation without unnecessarily strong conditions or assumptions about hearer inferences. 
Additionally, the proliferation of coherence relations is reduced by separating supplemental roles 
from domain-specific relations. 
My argument for the above taxonomy depends on my theoretical assumptions about the text 
planning process rather than on direct linguistic evidence. This deficiency can be addressed by 
empirical work. However, regardless of whether this particular taxonomy proves to be the most 
useful, I expect that it is plausible enough to make the point that a complete description of the 
coherence and communicative functionality of discourse requires acknowledgment of functionally 
distinct categories of relationships between segments, not just different relationships within a 
category. This is especially the case in text genres that mix informational and intentional goals 
(such as instructional texts). As Owen P~ambow pointed out in his workshop abstract, such genres 
require the development of new text planning architectures - the primary topic of Suthers \[1993a\]. 
Some Remaining Questions 
I believe that the need to consider the above aspects of discourse structure during text 
planning is fairly well estabhshed. (I take the above list to be an approximation to a lower bound, 
being the items I needed to process in my own planner.) The question remains: which aspects 
of discourse structure are processed by an agent in understanding a text. 7 This is an empirical 
question which I won't speculate on further here. However it should be noted that the above 
characterization of coherence is oriented towards the needs of generation rather than the process 
of comprehension. "Coherence" from the point of view of the understander need not be the same 
as "coherence" from the point of view of the generator, as their processing requirements may 
differ. 
A related question is whether coherence is the combined effect of its aspects identified by the 
analyst, or whether it suffices to attend to only one source of coherence when many are possible. 
Gisela Kedeker 3 argues that a hearer need only recognize the presence of one coherence relation; 
therefore when multiple relations hold between two segments (ideational, rhetorical, or segmental 
in her terminology), the speaker can choose which to mark. (She provides evidence that one 
could have marked both.) Yet marker choice does not rule out the need to process other aspects 
of coherence, for example the planner must decide whether and how to use an ideational (e.g. 
causal) relation in constructing a rhetorical (e.g., evidential) one. 
aTalk at LB.DC, Pittsburgh, May 1993. 
130 
Conclusion 
Both the analysis and generation of coherent and functional text depend on the coordina- 
tion of multiple types of structural relations at different granularities. It is not productive to 
view discourse structure as consisting of a single class of relations: one risks omitting or con- 
flaring relevant yet distinct functionalities and sources of coherence. The author has proposed 
distinctions between exchange, informative, propositional, contextual, supplemental, and sequen- 
tial structure \[Suthers 1993a\]: these differentiate types of intentional and rhetorical structure. 
Improved taxonomies are possible and may depend on whether one is primarily interested in 
modeling generation or understanding. Regardless, we need to deal with multiple descriptions 
when analyzing a text. The community should focus on factoring functionally distinct aspects of 
discourse structure according to their functionality in generation and understanding processes. 

References 
Grosz, B. & Sidner, C. \[1986\]. Attention, intention, and the structure of discourse. Com- 
putational Linguistics. Vol. 12, no. 3. pp. 175-204. 
Hobbs, J. \[1979\]. Coherence and coreference. Cognitive Science. Vol. 3, no. 1, pp. 67-90. 
Hovy, E. \[1988\]. Approaches to the planning of coherent text. Presented at the 4th Interna- 
tional Workshop on Text Generation. Catalina Island, California, July 1988. 
Moore, J. &: Paris, C. \[1991\]. The EES explanation facility: its tasks and its architecture. 
Workshop on Comparative Analysis of Explanation Planning Architectures, 9th National 
Conference on Artificial Intelligence. July 1991, Anaheim, CA. 
Moore, J. & Pollack, M. \[1993\]. A problem for KST: The need for multi-level discourse 
analysis. Computational Linguistics. Vol. 18, no. 4, pp. 537-544. 
Suthers, D. \[1990\]. Reassessing rhetorical abstractions and planning mechanisms. 5th Inter- 
national Workshop on Natural Language Generation. June 1990, Dawson, Pennsylvania. 
Suthers, D. \[1993a\]. An analysis of explanation and its implications for the design of expla- 
nation planners. Ph.D. Dissertation, University of Massachusetts. 
Suthers, D. \[1993b\]. Sequencing explanations to enhance communicative functionality. To 
appear in Proc. Cognitive Science. June 1993, Boulder, Colorado. 
Zukerman, I. \[1990\]. Generating peripheral rhetorical devices by consulting a user model. 5th 
International Workshop on Natural Language Generation. June 1990, Dawson, Pennsyl- 
vania, pp. 156-163. 
