DISCOURSE MODELS~ DIALOG MEMORIES~ AND 
USER MODELS 
Katharina Morik 
Technical University 
Berlin, West Germany 
1 INTRODUCTION 
In this paper, we discuss some terminological issues 
related to the notions of discourse models, dialog mem- 
ories, and user models. It is not our goal to show how 
discourse modeling and user modeling should actually 
interact in a cooperative system, but to show how the 
notions of discourse model, dialog memory, and user 
model can be defined and related in order to prevent 
misunderstandings and confusion. We argue that dialog 
memory may be subsumed under user model, as well as 
under discourse model, but that the three concepts 
should not be identified. Several separating criteria are 
discussed. We conclude that discourse modeling and 
user modeling are two lines of research that are orthog- 
onal to each other. 
2 DIALOG MEMORY AS PART OF USER MODEL 
A dialog memory can be viewed as part of a user model, 
namely the part that represents the dialog-dependent 
knowledge of the user (Morik 1984). Entries out of the 
dialog memory may cause entries in the user model, and 
entries of the user model may support the interpretation 
of an utterance, the interpretation then being stored in 
the dialog memory. However, in order to keep technical 
terms precise, user modeling on the one hand, and 
building and exploiting a dialog memory on the other 
hand should not be identified. This would lead to a 
reduction of what user modeling is about by disregard- 
ing all aspects other than dialog-dependent knowledge 
of the user as known to the system, while in fact there 
is some information that is to be covered by a user 
model and that may not be covered by a dialog memory. 
Let us think, for example, of a visit to the dentist's. 
The dentist will have some expectations concerning the 
client before the client said a word---even before he 
opened his mouth. This is due to the conversational 
setting, the roles of dentist and client. The same two 
persons meeting in another environment (e.g., at a 
political event, a horse race, or the opera) would not 
rely on the dentist-client expectations but on the infor- 
mation that then belongs to their roles. 
A user model contains explicit assumptions on the 
role of the user and the way a particular user plays it. 
The system exploits the user model systematically for 
playing its role more cooperatively by adopting to 
diverse users. To that end it uses rules which are 
parametrized according to the facets of the user. A user 
model is built up based on a "naive psychology", which 
forms a consistent image of the user. 
Schuster also states that the user model covers 
entities that do not belong into the dialog memory. In 
addition to the argument mentioned above (the dialog 
memory being the part of the user model that represents 
the dialog-dependent knowledge of the user), she points 
out that the dialog memory is used for building up parts 
of the user model and that both, user model and dialog 
memory, are used for generating an adequate answer. 
But if this were a valid argument for establishing a 
subsumption relation, we should also view the grammar 
as part of the user model, because the grammar is 
necessary for understanding and producing utterances. 
All the knowledge sources of a natural language system 
(hopefully) work together. We separate them conceptu- 
ally not because of their independence, but because 
they contain different kinds of knowledge that contrib- 
ute to the overall task in different ways. 
A dialog memory contains all beliefs that can be 
inferred with certainty from utterances, so that they 
belong to the mutual belief space. For example, the 
objects and their properties introduced in a dialog are 
typical entries in a dialog memory. Also, presupposi- 
tions that can be inferred from articles or question 
particles belong into the dialog memory. The linguistic 
rules that determine the inferences are valid and binding 
for all conversational settings. General rules establish 
mutual beliefs on the basis of utterances. The dialog 
memory is then used for, e.g., determining the appro- 
Copyright 1988 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided 
that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To 
copy otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/88/0100o-o$03.00 
Computational Linguistics, Volume 14, Number 3, September 1988 95 
Katharine Morik Discourse Models, Dialog Memories, and User Models 
priate description (definite/indefinite), anaphoric expres- 
sion, or characterization. 
3 DIALOG MEMORY AS PART OF DISCOURSE MODEL 
Another notion that is very close to user model as well 
as to dialog memory is discourse model. Sometimes 
dialog memory and discourse model are treated as 
synonyms (e.g., Wahlster 1986). Given the above defi- 
nition of dialog memories, however, there is a differ- 
ence between the two notions. As opposed to Schuster, 
who defines a discourse model as "containing represen- 
tations of entities, along with their properties and rela- 
tions they participate in", which corresponds exactly to 
our dialog memory, I use discourse model according to 
the framework of Grosz and Sidner (1986), where a 
discourse model is the syntactic structure of a dialog. 
One part of it, though, could be identified with the 
dialog memory, namely the focus space stack. The 
overall discourse model additionally represents the 
structure of the dialog with the segments and their 
relations, which is not part of the user model. Decom- 
posing a dialog into segments and establishing relations 
between them does not depend on a particular conver- 
sational setting. As is the case with dialog memories the 
overall discourse model, too, is built up by general 
linguistic rules that need not be parametrized according 
to a certain user. 
4 SEPARATING CRITERIA 
Previous attempts to separate user models from dis- 
course models have used the short-time/long-time crite- 
rion, arguing that entries in the dialog memory can be 
forgotten after the end of the dialog, whereas entries in 
the user model are to be remembered. The same argu- 
ment applies to dialog memories as part of the discourse 
model. The rationale of this argument is that anaphors 
are not applicable from one dialog to another and that 
the structure of a dialog is unlikely to be recalled as the 
syntactic structure of uttered sentences--just to men- 
tion these two phenomena. 
But does that mean that the entities with all their 
properties and relations as communicated in the dialog 
are forgotten? What would be the reason to talk to each 
other then? How could we learn from each other? 
Knowledge is to a great extent transferred via dialogs. 
Second, how could speech acts have social obligations 
as a consequence that may well hold for a long time? 
(Think of promises, for example!) Although the speaker 
may have forgotten the dialog, the hearer has--by very 
general conventions of language use--the right to insist 
on the speaker's commitments (Lewis 1975, Searle 
1969, Wunderlich 1972). 
The synthesis of those seemingly conflicting obser- 
vations is that the content of the dialog memory is 
integrated into the world knowledge. In other words, 
the conteilt of the focus space stack is partially incor- 
porated into the world knowledge when it gets popped 
off the stack. So, the structure is lost, but the content is 
at least partly saved. 
Turniing things the other way around, why couldn't 
properties or character traits of a user be forgotten? 
What makes entries of a user model more stable? Think, 
for instance, of a post office clerk. Although he may 
adapt his behavior to the particular customer during the 
dialog, he normally forgets the information about her or 
him immediately after the dialog. As Rich (1979) pointed 
out, user models may be short term or long term. Thus 
short-time/long-time or forgettable/unforgettable is no 
criterion for dividing user models from dialog memories 
or discourse models. 
Another criterion could be, whether the knowledge is 
used for generating the linguistic form (how to say 
something) or for establishing the content of a system's 
utterance (what to say). Clearly, dialog memory and 
overall discourse model deal with the linguistic struc- 
ture of dialogs, e.g., the reference resolution and the 
appropriate verbalization of concepts. The user model, 
on the other hand, also covers information that directs 
the selection of what to utter. The user's level of 
expertise determines the particularity of a system utter- 
ance, the realization of notes of caution, and the word 
choice, for instance. The user's wants establish the 
context of the user utterances and guide the system's 
problem solving, thus keeping the system behavior 
directed towards the user goals. 
This distinction, however, is not clear cut either, for 
two reasons. First, the line between what to say and 
how to say it is rather fuzzy. Referencing a concept, for 
example, also involves choosing the appropriate at- 
tributes for characterization--and this is naturally a 
matter of what to say. Second, this criterion would 
exclude work as presented by Lehman and Carbonell 
(1988) from the area of user modeling. There, linguistic 
rules are specialized for a particular user in a particular 
conversational setting. This is clearly not a matter of the 
dialog memory, but of the user model, although it is 
concerned with the linguistic form. Thus the form/ 
content distinction does not separate user models from 
dialog memories and discourse models, either. 
The difficulty to find criteria separating between 
discourse models and user models indicates a case of 
cross-classification. The criteria, namely what is spe- 
cific to a user and what concerns dialog structure 
naturally cross. Dialog memory falls into both catego- 
ries. 
On the one hand, from what the user utters his 
beliefs, his level of expertise in a certain domain, his 
wants, and his language style can be inferred. This 
knowledge can be used by all the system components: 
tuning syntactic analysis, resolving reference, determin- 
ing the input speech act, disambiguating the input, 
selecting relevant information, organizing the text to be 
96 Computational Linguistics, Volume 14, Number 3, September 1988 
Katharine Morik Discourse Models, Dialog Memories, and User Models 
outputted (Paris 1988), choosing the appropriate words, 
referencing, topicalizing, etc. In order to do so, all the 
components must include procedures that are parame- 
trized according to the (particular) user model. Cur- 
rently, the interactive systems do not put all the user 
facets to good use in all their components. This is not 
due to principled limits, however, but rather to a 
shortcoming in the state of the art. 
On the other hand, the user's utterances can also be 
analyzed from another viewpoint, namely incorporating 
them into a coherent discourse model as described by, 
e.g., Grosz and Sidner (1986). Also, this model can be 
used during all processing steps from understanding to 
generating. 
Both user models and discourse models are built up 
(at least partially) from the user utterances. Both con- 
tribute to a cooperative system behavior. But they do so 
from different viewpoints with different aims. Adopting 
to a particular user on the one hand and achieving a 
coherent, well-formed dialog on the other hand are two 
aims for a cooperative system which are orthogonal to 
each other. The terms user model and discourse model 
denote different aspects of a system. Thus although the 
notions are intensionally different, the extension of their 
respective definitions may overlap. 

REFERENCES 
Grosz, B. and Sidner, C. 1986 Attention, Intentions, and the Structure 
of Discourse. In Computational Linguistics 12: 175-204. 
Lehman, J. F. and Carbonell, J. G. 1988 Learning the User's Lan- 
guage: A Step Towards Automated Creation of User Models. In 
Kobsa, A. and Wahlster, W. (eds.), User Models in Dialog 
Systems. Springer-Verlag, Berlin--New York. 
Lewis, D. 1975 Languages and Language. In Gunderson, K. (ed.), 
Language, Mind and Knowledge, Minnesota Studies in the Phi- 
losophy of Science 7. University of Minnesota Press, Minneapolis, 
MN. 
Morik, K. 1984 Partnermodellierung und Interessenprofile bei Dia- 
logsystemen der Kfinstlichen Intelligenz. In Rollinger, C. R. (ed.), 
Probleme des (Text-) Verstehens: Ansdtze der Kiinstlichen lntel- 
ligenz. Niemeyer, Tfibingen, W. Germany. 
Paris, C. L. 1988 Tailoring Object Descriptions to a User's Level of 
Expertise. In Kobsa, A. and Wahlster, W. (eds.), User Models in 
Dialog Systems. Springer-Verlag, Berlin--New York. 
Rich, E. 1979 Building and Exploiting User Models. Ph.D. thesis, 
Department of Computer Science, Carnegie-Mellon University, 
Pittsburgh, PA. 
Searle, J. R. 1969 Speech Acts. Cambridge University Press, Cam- 
bridge, England. 
Wahlster, W. 1986 Some Terminological Remarks on User Modeling. 
Paper presented at the International Workshop on User Modeling, 
Maria Laach, W. Germany. 
Wunderlich, D. 1972 Sprechakte. In Maas, U. and Wunderlich, D. 
(eds.), Pragmatik und sprachliches Handeln, Athen/ium Verlag, 
Frankfurt, W. Germany. 
