DISTINGUISHING USER MODELS FROM DISCOURSE MODELS 
Wolfgang Wahlster 
Department of Computer Science 
University of Saarbriicken 
West Germany 
1 INTRODUCTION 
In the discussion about the relationship between user 
models (UMs) and discourse models (DMs) so far, at 
least three positions have been stated explicitly: 
P1. the DM is a part of the UM (e.g., Schuster) 
P2. the DM intersects the UM (e.g., Chin) 
P3. the DM and the UM are distinct (e.g., Wahlster 
1986, Cohen) 
Of course, the interpretation of these positions depends 
on the definition of the terms involved and the under- 
lying notion of the "part-of", "intersect", and "dis- 
tinct" relations. The relationships cannot simply be 
interpreted in a set-theoretic sense, since all definitions 
for UMs and DMs proposed so far depend not only on 
representation structures, but also on processes used 
for the construction, maintenance, and exploitation of 
these structures. 
Since this is a terminological, and not an empirical, 
discussion, as I pointed out in Wahlster (1986), P1-P3 
are primarily normative statements. So, P3, for in- 
stance, must be interpreted as "The terms UM and DM 
should be defined in such a way, that they do not 
overlap". 
This view seems not to be shared by all participants 
in the discussion. Schuster, for example, tries to prove 
her position (PI) in a set-theoretic sense. First, she 
argues that "the user model contains information that 
does not appear in the discourse model" and then she 
"proves" that "any information in the discourse model 
is also in the user model". 
I disagree not only with the form, but also with the 
content of Schuster's argumentation. She writes "only 
if the discourse model is part of the user model can the 
system take it into account in its responses and its 
reasoning about the users". By considering an isomor- 
phic argumentation like "only if a tomato is part of 
cheese, can one use it to prepare pizza" it becomes 
clear that this proof is flawed. 
Also, Morik points out correctly that if one follows 
Schuster's argumentation one should "view the gram- 
mar as part of the user model, because the grammar is 
necessary for understanding and producing utter- 
ances". 
Today, it is a standard hypothesis in AI and compu- 
tational linguistics that models for the language under- 
standing and generation process must exploit various 
knowledge sources, including in many cases a DM and 
a UM. For example, in Jameson and Wahlster (1982) we 
described the NP generator of the HAM-ANS system, 
in which the generation of a definite or indefinite des- 
cription was influenced both by the UM and the DM. 
But this in no way means that one must be included in 
the other. 
As long as there is no definitive evidence (e.g., from 
psychology or the neurosciences) for a particular struc- 
ture, content, and use (or even existence) of UMs and 
DMs in the human information processing system, in AI 
the notions of UM and DM are concepts that help on the 
one hand to construct a theory of natural language 
dialog behavior, and on the other hand to structure the 
software systems that realize natural language systems. 
From the second point of view, which is the engi- 
neering perspective, the question of whether P1, P2, or 
P3 holds, is easy to decide so far. In most of the 
implemented systems the data structures and proce- 
dures labeled UM and DM are completely distinct. 
Even the recent GUMS package (Finin 1988), a general 
user modeling component, contains no specific repre- 
sentation structures or processes for discourse model- 
ing. 
Since the discussion above suggests that we view the 
relation between the UM and the DM mainly as a 
terminological problem, in the next section we focus on 
possible definitions for UMs and DMs. Although often 
terminological discussions become quite tedious, at this 
point it seems to be important to define these concepts 
as precisely as possible, since many researchers are 
discovering interesting relationships between discourse 
and user models. 
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 101 
Wolfgang Wahlster Distinguishing User Models from Discourse Models 
2 DEFINING USER MODELS AND DISCOURSE MODPLS 
Some authors define user models simply as information 
that the system has about its users (e.g., Schuster). I 
think this definition is too broad. Consider an NL 
interface to a data base, which contains the following 
relation: 
26 Jones 32 40 
When Mr. Jones happens to be the user of this system 
and asks, "What is my bonus?", the system should 
respond "40". In this case, the system has information 
about the user, but one would not like to say that its 
response was based on a user model. 
Even if one restricts the definition above to "infor- 
mation about the user put to use" (see Sparck Jones), it 
is not strong enough. If a deductive data base in addition 
to the relation above includes a rule like "If AGE(X) > 
30 and BONUS(X) > 35 then STATUS(X) = 10" and 
Mr. Jones asks, '°What is my status?" the system 
should respond "10". Even though the deductive 
DBMS uses information about the user to instantiate the 
inference rule, such a system should not be construed as 
having a user model. 
I propose the following joint definitions of user model 
and user modeling component (see Wahlster and Kobsa 
1988) as well as discourse model and discourse model- 
ing component in the context of NL dialog systems: 
A user model is a knowledge source that contains 
explicit assumptions on all aspects of the user that may 
be relevant for the dialog behavior of the system. A user 
modeling component is that part of a dialog system 
whose function is to 
• incrementally build up a user model; 
• store, update, and delete entries in it; 
• maintain the consistency of the model; and 
• supply other components of the system with assump- 
tions about the user. 
A discourse model is a knowledge source that contains 
the system's description of the syntax, semantics, and 
pragmatics of a dialog as it proceeds. A discourse 
modeling component is that part of a dialog system 
whose function is to 
• incrementally build up a discourse model; 
• store and update entries in it; and 
• supply other components of the system with informa- 
tion about the structure and content of the previous 
segments of the dialog. 
While it seems commonly agreed upon that a DM 
should contain a syntactic and semantic description of 
discourse segments, a record of the discourse entities 
mentioned, the attentional structure of the dialog in- 
cluding a focus space stack, anaphoric links, and de- 
scriptions of individual utterances on the speech act 
level, there seem to be many other ingredients needed 
for a good discourse representation which are not yet 
worked out in current computational discourse theory. 
Therefore, I prefer to refer only to the abstract levels of 
necessary discourse representation in the definition 
above. 
3 SOME DIFFERENCES AND SIMILARITIES BETWEEN USER 
MODELS AND DISCOURSE MODELS 
An important difference between a discourse model and 
a user model is that entries in the user model often must 
be explicitly deleted or updated, whereas in the dis- 
course model entries describing the structure and con- 
tent of utterances of the ongoing dialog are never 
deleted (except for forgetting phenomena, which are 
beyond the scope of the current discussion). Thus, 
according to our definition above, a belief revision 
component is an important part of a user modeling 
component. 
Consider the following dialog with a hypothetical 
tutoring system in the SCHOLAR tradition. 
System: (1) Tell me about California. 
User: (2) San Francisco is the capital of 
California. 
System: (3) No, that's wrong. 
User: (4) I see. So, that's not the capital. 
(5) Then, what is its capital? 
System: (6) Sacramento. 
(7) Now, tell me why you mentioned San 
Francisco first, when you began to talk 
about California. 
A simple consequence of the user's response (2) is an 
entry in the system's user model, which represents the 
fact, that the system believes that the user believes 
(B1). After (3), and certainly after (4) the user model 
should contain (BI'). 
(B1) capital(California, San-Francisco) 
(B 1') not(capital(California, San-Francisco)) 
(B2) capital(California, Sacramento) 
This means that the user modeling component has to 
remove (B1) from the user model (in a reason mainte- 
nance system this causes (B1) to be added to the set of 
beliefs, which are currently "out"). After (6) the user's 
belief (B2) should be added to the system's user model. 
If the apriori user model contains "For each state there 
exists one and only one capital" as a mutual believed 
fact, then the user modeling component can also re- 
move (BI') after adding (B2). 
In the discourse model, of course, the fact that the 
user uttered sentence (2) should not be deleted. For 
example, the system could go on and ask the user a 
question like (7), which explicitly refers to the fact that 
(2) was the first reaction to (I). What this simply means 
is that the fact that the user made a particular assertion 
102 Computational Linguistics, Volume 14, Number 3, September 1988 
Wolfgang Wahlster Distinguishing User Models from Discourse Models 
remains true even if the user's belief changes and he 
withdraws his previous assertion. 
Even a metacommunicative act like (9) should not 
delete entries in the discourse model, as the successful 
anaphoric reference in (10) to a discourse entity intro- 
duced in (8) suggests. But it is obvious that in the user 
model the corresponding representation of the user's 
wants has to be changed. 
User:(8) I don't want to travel with my kids. 
(9) Forget what I just said. 
(10) I want to travel with them. 
This does not imply that the discourse model is static 
and the user model is dynamic. The discourse model is 
also highly dynamic (consider, e.g., focus shifting), but 
it lacks the notion of logical consistency, which is 
important for belief revision and default reasoning in a 
user modeling component. In my view, the discourse 
model is like an annotated trace of the various levels of 
the system's processing involved in understanding the 
user's utterances and generating its own dialog contri- 
butions. 
Let's consider another example to emphasize the 
differences between a UM and a DM. Suppose that the 
system plays the role of a travel agent, who wants to sell 
trips to the well-known holiday places A and B, for 
which it has some reasonably priced offers. When the 
user asks, "What are your cheapest trips?" the system 
lists A and B first, followed by a hastily presented list of 
eight other places with names, which it assumes are 
totally unfamiliar to the user. In the system's DM all ten 
places appear, but the user modeling component of the 
system explicitly assumes that the user only believes 
"cheap-trip-to(A)", "cheap-trip-to(B)" together with 
the belief that there are some other cheap trips avail- 
able. This is exactly the aim of the uncooperative 
behavior of the travel agent: Now, it is likely that the 
user wants to know more about the offers A and B, 
which the agent wants to sell. But if the user later finds 
out that a trip to one of the other places is much cheaper 
and better, and complains to the travel agent, "Why 
didn't you suggest this trip right at the beginning?", the 
travel agent can refer back to his DM and say, "I 
mentioned this place among my first suggestions". 
Some authors claim that the discourse model expires 
at the end of a dialog, while parts of the user model may 
be saved for further use (e.g., Chin). I think that is 
wrong. Often a dialog participant is able to paraphrase a 
segment of a previous dialog without remembering who 
the dialog partner was and at what time and location the 
dialog took place. While he may not be able to recon- 
struct the exact phrasing, he has access to a represen- 
tation of the semantics and pragmatics of the interac- 
tion. Furthermore, I think that often conversational 
rules and tactics are learned by induction over a large 
set of interaction patterns extracted from discourse 
models, which were partially saved in episodic memory, 
where they are not necessarily associated with a long- 
term user model. In order to learn how to use a language 
it seems to be important to not always discard the 
complete discourse model after the end of a conversa- 
tion. 
On the other hand, one often has many assumptions 
about the beliefs, plans, and goals of a dialog partner 
before a new dialog begins (cf. Wahlster and Kobsa 
1988), without having a clear idea from which actual 
dialogs the assumptions in this user model were de- 
rived. Thus I agree with Morik that the short-term/ 
long-term criterion cannot be used to distinguish user 
models and discourse models. If one prefers to restrict 
the term discourse model to an ongoing conversation 
and to define the saved portions of it as part of the world 
knowledge, then one should do the same for the term 
user model, so that again the criterion does not discrim- 
inate. 
While in many cases the UM component and the DM 
component process the same input (e.g., a meaning 
representation of the last utterance), and their output is 
used by the same processes, I would suggest that both 
components be kept separate. Even if there is some 
information, which should be present in both models, it 
will be represented in another form, since, as I pointed 
out above, the functionality and the type of processing 
of the UM and DM components are so different. In this 
case we have a multiple knowledge representation in the 
UM and the DM, which is quite common in complex A1 
systems. 
As I remarked in the beginning, in asking who is right 
in this discussion, one must carefully evaluate the 
corresponding definitions for UM and DM proposed by 
the respective authors. In this paper, I introduced and 
motivated definitions, under which a UM and a DM are 
separate, but related to each other. 

REFERENCES 
Finin, T. W. 1988 GUMS: A General User Modeling Shell. In Kobsa, 
A. and Wahlster, W. (eds.), User Models in Dialog Systems. 
Springer-Verlag, Berlin--New York. 
Jameson, A. and Wahlster, W. 1982 User Modeling in Anaphora 
Generation: Ellipsis and Definite Description. In Proceedings of 
the 1982 European Conference on Artificial Intelligence, Orsay, 
France; 222-227. 
Wahlster, W. 1986 Some Terminological Remarks on User Modeling. 
Paper presented at the International Workshop on User Modeling, 
Maria Laach, W. Germany. 
Wahlster, W. and Kobsa, A. 1988 User Models in Dialog Systems. In 
Kobsa, A. and Wahlster, W. (eds.), User Models in Dialog 
Systems. Springer-Verlag, Berlin--New York. 
