From Planning to Actions: 
Realizing Intentions by Rhetorical Means 
Helmut Horacek 
University of Bielefeld, LILI-Faculty 
Univ.str. 25, \¥-4800 Bielefeld 1, GERMANY 
Fax: +49-521-106-2996 (or 5844) 
horacek©t echf ak. un±-bielefeld, de 
1. Introduction 
This essay describes experiences made in connection with tile consultation system WISBER \[llo- 
racek et al. 1988\] and with the expla.nation device DIAMOD \[Horacek 1993a\], featuring tile role 
of intentions in the former and the role of rhetoric in the latter. Assessing the results obtained, 
we clarify our standpoint concerning the workshop topic, emphasizing on future prospects for text. 
planning. 
2. Tile role of intentions in discourse planning 
In sinmlating a consultant's role, a central capahility of WISBER is reasoning about the eonversants' 
intentions to determine a communicative action that is likely to achieve progress towards a long- 
term consultation goal \[Gerlach, Horacek 1989\]. The mental model of tile agents involved consists 
of BEI, IEFs, KNOWs, aud WANTs, which are expressed as I)rimitive, time-dependent states. An 
agent's INTENTION is defined as a WANT to achieve a state, including a BELIEF that performing 
a certain action will lead to that state. A rule-hased mechanism is invoked to mai,tain consistency 
by updating tile temporal validity of these mental states, to infer a communicative system goal, and 
to associate appropriate speech acts with that goal (this association is made more profound by a 
rule- based mechanism in Cohen and Levesque's style \[Horacek 1991b\]). 
Evidence about relations between several WANTs derived from subsequent utterances is kept \[tIo- 
racek, Gerlach 1990\]. The contextual interpretation of these relations is done on the basis of CA- 
PABII, ITIES and COMPETENC, ES ascribed to the agents involved, augmented by the actual POS- 
SIBII,ITY to carry out the actions referred. This way, intentions are related via their contents 
(one is a PAR'F of another, or one is the REASON for another). In a more elahorate model of 
consultatio. \[lIoracek 1990@ a. agent's evaluations are expressed in terms of ESTIMA'FES and 
(derived) PREFERENCES. These assessments may refer to general and vague concepts introduced 
in the dialog context, which the consultant has to relate meaningfully to the concrete properties of 
the candidate domain objects. 
31 
The system typically produces one- or two-clause dialog contributions composed in a two-stage 
lexicalization process \[Horacek 1990b\] which respect the addressee's command of terminological 
knowledge and reflect simple stylistic preferences. Only when producing a recommendation, the 
surface realization structurally differs from the intentional representation, so that rhetorical as- 
pects are really a concern. In this case, the consultant's intention comprises the identification of a 
recommendable object, and the description of a property considered most interesting for the advice- 
seeking person. The surface form consists of two separate sentences serving these purposes in the 
indicated order. Because of that order, interpretations that the property is meant to be restrictive 
or a motivation or justification for the choice made are excluded, and the fact that only one prop- 
erty is mentioned carries the message about its importance - the environment is cooperative, not 
mentioning other properties must be on purpose. 
3. Applying rhetorical means to communicate more complex messages adequately 
In producing explanations about an expert system's problem solving behavior, we concentrate our 
effort on generating logical explanations. For this issue, det.erntil~ation of semantic content proves 
to be a complex task, guided by finding and exploiting dominances and redundancies among po- 
tentially useful argulnents \[Horacek 1992b\]. The resulting argumentative structure is converted into 
a text plan in a structure-preserving way \[Horacek 1992a\]. For instance, hierarchical relations are 
maintained so that an argumentative pattern consisting of a justification breaking down into sub- 
arguments is carried over to the text structure. For increasing readability, the text is augmented 
by a.n itd, roductory statement about (.he structure of the message and by a summary, llowcver, the 
motivation for tho a.i)plication of these tneasurements still has to be cal)tured in a formal way. 
Moreover, we have designed a mechanism that aims a.t exploiting the addressee's inference capability 
for avoiding (boring) redundancies, by anticipating his/her coruprehension process in the generation 
task. This procedure is justified by the observation that (the relevance of) generic regularities can 
be inferred from referential facts and vice-versa, due to the expected relevance attributed to infor- 
mation conveyed in a cooperative, task-oriented enviromnent. Assumptions about the addressee's 
acquaintance with domain knowledge have significant influence on anticipating his/her inferences. 
In addition, psychologically motivated principles (like maximal coverage and minimal complexity) 
guide preferences among possible interpretations \[Iloracek 1992c\]. tlence, only some ccntral pieccs 
of information need to be said explicitly for getting the entire message across. 
4. Future prospects for text planning 
We think that these considerations provide sufficient evidence for the roles of intentions and rhetoric 
in discourse. Intentions drive internal goal-oriented reasoning about mental models of agents and 
manifest themselves in terms of l)lans adopted and actions carried out, independently of whether 
communication is involved or not. Intentions address envisioning physical or mental states related 
among each other in a way described in section 2. Rhetoric manifests itself in the differences between 
(rei)resentations of) mental states and the variety of surface forms used to satisfy the underlying 
goals by adequate l)resentations. These measurenaents include structuring and ordering pieces of 
text, applying rhetorical figures, and taking consequences of the Gricean maxims into account (for 
instance, conveying a piece of information or not is on purpose). 
in our view, relations between rhetoric and intentions are described well in (classical) literary rhetoric 
32 
which, therefore, should also guide adequate formalizations. In that approach, intentions comprise 
a pure semantic content (the 'proprium') and an attitude holding towards it. The art, of rhetoric 
is then to select forms that convey the intended attitude towards the 'proprium'. In computational 
approaches, on the contrary, semantic, intentional, and textual relations are frequently treated in 
a uniform way, subsumed under tile term 'rhetorical relations', which makes organizating the text 
planning process less clear. 
There are several prerequisites for orienting text planning on classical rhetoric. A basic and general 
vocabulary of semantic relations is needed (CAPABILITY, POSSIBILITY, COMPETENCE, and 
the associated state-transition model constitute a first step towards that goal). On tile basis of these 
semantic relations, relations between intentions can be defined, including formalizations of aspects 
of cooperativity (a primary sketch can be found in \[tloracek 1991b\]). In our view, this approach 
also seems to be promising to develop formalizations for complex relations like SOLUTIONHOOD 
and CONCESSION, on tile basis of the underlying facts and attitudes holding towards them. More- 
over, tile associated reasoning process has to reflect degrees of graduality which characterize several 
aspects in communication: an agent's assessments (e.g., ESTIMATE \[Horacek 1990a\]), his/her ac- 
quaintance with domain terms and concepts, and progress towards a long-term communicative goal. 
'Fhe quantitative aspect is missing in compul.ational apl)roaches. 
When the employment of rhetorical figures is invoked to compose a text that is supposed to serve 
a comnmnicative intention, the choices to make have to be oriented ou anticipating the addressee's 
comprehensioll process in several respects: 
• Making use of his/her inference capabilities, only parts of tile inforlnation to convey need to 
be expressed explicitly (see \[noracek 1991a1, \[noracek 1992c\]). 
• There is a trade-off between fluent expressions, which may entail anlbiguities, and precise ones, 
which may occasionally sound awkward 
• \[Block, Horacek 1990\] argue in fervor of naturalness, if the resulting ambiguity is tolerable. 
• Envisioning a reasonable degree of comprehension (see \[Horacek 1993b\]), tile amount of infor- 
mation communicated should not exceed a certain quantity, among others; this goal would also 
yield motivations for tile incorporation of purely textual relations like summaries and hints 
about the structure of a text.. 
We, think that these argmnenl,s have ident, ificd a couple of problems in tile area of intentionality and 
rhetoric, but also some promising ways towards solutions. 

References 
\[Horacek et al. 1988\] H. tloracek, It. Bergmann, M. Fliegner, II. Marburger, M. Poesio, M. Sprenger, 
From Meaning to Meaning - a Walk Through WISBEIUs Semantic-Pragmatic Processing. In Proc. 
of Germau Workshop ou A1-88, W. Iloeppuer (ed.), Geseke, 1988. 
\[(.~erlach, lloracek 1989\] M. Gerla.ch, II. lloracek, 1)ialogue C, ontrol in a Natural l,anguage System. 
In Proc. of EACL-89, 11. Somers, M. McGee (eds.), Manchester, 1989. 
\[lloracek 1990a\] 11. ttoracek, A Framework for Consultation. In Cybernetics and Systems'90, R. 
rFral)pl (ed.), Wien, 1990. 
\[lloracek, Gerlach 1990\] H. Horacek, M. Gerlach, Tile Role of Goals ill Dialog Control. Ill Proc. of 
ECA1-90, L. Aiello (ed.), Stockholm, 1990. 
\[Block, Iloracek 1990\] R. Block, H. lloracek, Generating Referring Expressions Making Use of Mul- 
tiple Kllowledge Sources. Ill Proc. of COLING-90, F. Karlsson (ed.), Helsinki, 1990. 
\[Horacek 1990b\] H. Horacek, The Architecture of a Generation Component in a Complete Natural 
Language Dialog System. In Current Issues in Natural Language Generation, R. Dale, C. Mellish, 
M. Zock (eds.), Academic Press, 1990. 
\[lloracek 1991a\] H. ltoracek, Exploiting Conversational hnplicature for Generating Concise Expla- 
nations. In Proc. of EACL-91, J. Kunze, D. Reimann (eds.), Berlin, 1991. 
\[llora.cek 1991b\] I1. lloracek, A Model of "l'ask-()ri('ld.ed Communication Based on Principles of 
Ita.tioJlal Action a lid interactioll. In Proc. of Ausl.riall Conferellc(~ on AI-91, H. Kaindl (ed.), Wien, 
19!) 1. 
\[lloracek 1992a\] 11. Horacek, An Integrated View of Text Plamling. Ill Aspects of Automated Natural 
Language Generation, ft. Dale et al. (eds.), Springer, Berlin, 1992. 
\[lloracek 1992b\] H. Horacek, Explanations for Constraint Systems. In Proc. of ECAI-92, B. Neu- 
mann (ed.), Wien, 1992. 
\[lloracek 1992c\] H. tloracek, How to Avoid Explaining Obvious Things (Without Omitting Central 
hfformation). DIAMOD-Report 20, University of Bielefeld, 199:2. Also submitted to IJCAI-93. 
\[lloracek 1993a\] 1t. lloracek, Generating Explana.tiolls for Constraint-Based Expert Systems. Sub- 
rail.ted to IEEE-EXpert. 
\[lloracek 1993b\] 11. ltoracek, Viewing Natural Language Generation as an Optimization Process. 
Submitted to European Workshop on Natural Language Generation, 1993. 
