Domain-Dependent and Domain-Independent 
Rhetorical Relations 
Jong-Gyun Lim 
Department of Computer Science 
450 Computer Science Building 
Columbia University 
New York, N.Y. 10027 
lim@cs.columbia.edu 
1. Rhetorical Relations as Text Planning Operators 
Rhetorical relations have been used for text planning in many text generation sys- 
tems (\[McK82\] \[Hov88\] \[Moo89\], among others), hut how they are used vary rather 
significantly from one text planner to another. While studying them in detail side by 
side (\[Lim92\]), I have observed the following: 
1. A plan operator in a typical AI planner is to carry out an action whereas that in 
a text planner is to inform rhetorical relations among the actions and objects. 
2. A goal (or intention) in an AI planner is to change the state of the world whereas 
that in a text planner is to change the mental state of the hearer. 
More precisely, I was led to believe in the analogy between an text planning and 
a typical AI planning task as shown in Table 1. 
Intentions are goals of the text planner that can be realized by planning a text 
in terms of the rhetorical relations. Thus, the role of the rhetorical relations is to 
manipulate the mental objects in the mind of the hearer by creating or altering the 
relations among them. 
With this analogy, I view rhetorical relations as realization of intentions. Further- 
more, this analogy leads me to believe that rhetorical relations should be as numerous 
and varied as there are relations among actions and objects in the domain. This view 
seems to be in conflict with Dale's view (in this proceedings) which argues that 
rhetorical relation should only include textual relations rather than mirroring domain 
relations. Assuming many-to-many relation between intentional and informational 
56 
AI Planning \[ Text Planning 
operators (actions) ! rhetorical relations 
state of the world i mental states 
goals intentions 
plan (network of actions) texts (network of rhetorical relations) 
Table 1: Comparison Between Text Planning and A Typical AI Planning 
levels (see \[MP92\] and Korelsky and Kittredge in this proceeding), however, I think 
the number of rhetorical relations should depend on the number of relations in the 
domain. For example, if there is a domain relation R, then depending on who his 
hearer is and what his intentions are, a speaker may choose different method M to 
inform R to the hearer. The method M may be planned in terms of several rhetorical 
relations or a single rhetorical relation may be used for more than one method like 
M. More on this is discussed later. 
This analogy is in line with Traum's position of viewing rhetorical relation as 
speech acts or actions in general (in this proceedings). In particular, this view agrees 
with his point that a relation can be planned, performed, and recognized. Since 
actions are unbounded so should be the relations, which concurs with his criticism 
that it is meaningless to find the boundary for the right set of rhetorical relations. 
Traum, however, tends to emphasize the importance of intentions so much as to 
indicate that the role of rhetorical relations is only secondary and may not even be 
necessary in communicating intentions. Very often (especially in casual conversation), 
rhetorical relations are implied and thus hidden from the surface form (e.g. (2a) and 
(2b) in Traum's) . However, it is clear that a coherent text is structured with some 
meaningful rhetorical relations among its segments. Therefore, without identifying 
those relations (hidden or otherwise) planned by the speaker, we cannot say that the 
hearer understood the speaker. Thus, I take the position that in both text planning 
and recognition, rhetorical relations play a primary role in communicating intentions. 
2. Need for Domain-Dependent Rhetorical Relations 
Typical AI planner represents actions and objects in a hierarchical knowledge base 
where both domain-dependent and domain-independent concepts are represented. 
Hence, it is possible to plan at abstract level and generate plans in terms of the 
abstract plan operators. Similarly, rhetorical relations and intentions in text planning 
may include domain-independent types as well as domain-dependent ones. 
Without the domain-dependent Counterpart, however, abstract rhetorical relations 
by themselves are not very useful except for the limited use of meta conversation 
(talking about abstract relations among abstract concepts.) For example, rhetorical 
relations in TEXT \[McK82\], Hovy's RST plan operators \[Hov88\], and Moore's RST 
57 
plan operators \[Moo89\] are all domain-independent rhetorical relations. Thus, they 
all apply some facilities to the domain-independent relations to generate information 
about domain-dependent relations. 
In TEXT, a domain-independent rhetorical predicate of a schema is implemented 
with a detailed predicate function that contains directions to search through the 
knowledge base to retrieve the domain-dependent relations. Unfortunately, this makes 
it hard to write new schemas. Hovy's RST plan operators, on the other hand, are 
easier to write because no domain-dependent plan operators need to be implemented. 
However, the burden is swifted to the text structuring process where selected pool of 
propositions must be interpreated and matched to some RST plan operators. This 
necessitates the encoding of knowledge that maps a domain specific proposition to 
a domain-independent rhetorical relation which in my opinion is as hard as writing 
domain-dependent RST relations. 
Unlike Hovy's, Moore's plan operators contain intentional goals which make it 
possible to generate structured network of RST relations using hierarchical planning. 
However, an instantiated domain-independent relation typically does not capture 
the specific relations of the instantiated information. For example, the plan oper- 
ator PERSUADE-USER-TO-DO-ACT will be instantiated to (PERSUADED ?user 
(GOAL ?user (DO ?user ?act))) but same strategy will be used to persuade an ?user 
to do an ?act regardless of who the ?user or what the ?act might be. In reality, 
different persuading strategy might be needed for persuading John to replace SETQ 
to SETF and persuading John to jump off from an airplane. 
To solve this problem, Elhadad takes another extreme approach and introduces 
RST-like plan operators called topoi \[Elh92\]. Topoi are functionally equivalent to 
Moore's RST plan operators except that they are completely domain-dependent plan 
operators. The main advantages of using all domain-dependent rhetorical relations 
in text planning is that one does not need to struggle with naming and matching 
a specific relation to a generic level relation (like Evidence and Justifications) while 
still being able to build a text structure that reflects the rhetorical structure among 
the discourse units. The disadvantage is that the recursively embedded structural 
relations that are captured by RST relations among the various levels of discourse 
units can not be captured by topoi. As far as text planning is concerned, however, 
that doesn't seem like a disadvantage because such an information is not necessary for 
generating coherent text nor for responding to follow-up questions. Also, coherence 
of a text seems to depend more on the underlying domain plans than on the abstract 
rhetorical relations. 
3. Where Are the Domain-Dependent Rhetorical Relations? 
These observations lead me to conclude that the types of rhetorical relations and 
intentions that are most useful for text planning are those that are needed by the 
text planner the most, and it seems that the most needed rhetorical relations are 
the domain-dependent relations that capture the specific relations between the two 
discourse units. How then do we collect these domain-dependent rhetorical relations? 
My research is to collect these domain-dependent relations from a multi-agent pro- 
58 
gramming environment \[KP88\] and make them available to different kind of text 
planners that use them to generate explanations to the agents in that environment. 
From this domain, I have learned two things. One is that the domain-dependent 
relations have to be collected from a hybrid sources including rules in expert sys- 
tem, constraints in a scheduler, programs, and variables in the programs. The other 
observation is that one planning method is not sufficient to handle the various text 
planning tasks. Therefore, my system allows both the application programs and the 
users to define domain-dependent rhetorical relations and strategies to generate ex- 
planation for those relations. Depending on the nature of planning task, one planning 
method might be better suited than another. Thus, my system has the flexibility to 
choose among the different planning strategies. So far, domain-dependent rhetorical 
relations have been sufficient for this text planning needs. 

References 
M. Elhadad. Using Argumentation to Control Lexical Choice: A Unification- 
based Approach. PhD thesis, Columbia University, New York, NY, July 
1992. 
E. H. Hovy. Two types of planning in language generation. In Proceedings 
of the $6th ACL Conference, pages 179-186, Buffalo, June 1988. ACL. 
P. H. Kaiser, G. E. Feiler and S. S. Popovich. Intelligent Assistance for 
Software Development and Maintenance. IEEE Software, pages 40-49, May 
1988. 
J. G. Lim. Planning in AI and Text Planning in Natural Language Genera- 
tion. Technical Report CUCS-038-92, Columbia University, New York, NY, 
July 1992. area paper. 
K. R. McKeown. Generating Natural Language Text in Response to Ques- 
tions About Database Structure. PhD thesis, University of Pennsylvania, 
New York, NY, May 1982. Also a Technical report, No MS-CIS-82-05, 
University of Pennsylvania, 1982. 
J. D. Moore. A Reactive Approach to Explanation in Expert and Advice- 
Giving Systems. PhD thesis, University of California, Los Angeles, Los 
Angeles, CA, 1989. 
J. Moore and M. Pollack. A problem for RST: The Need for Multi-level 
Discourse Analysis. Computational Linguistics, 18(4):534-544, 1992. 
