ESTABLISHING THE RELATIONSHIP BETWEEN DISCOURSE MODELS 
AND USER MODELS 
Ethel Schuster 
Department of Computer and Information Science 
The Moore Schoon 
University of Pennsylvania 
Philadelphia, Pennsylvania 19104 
1 INTRODUCTION 
Many current research efforts have focused on building 
cooperative systems that interact with the{r users in a 
natural language such as English. To be effective, these 
systems must be robust, their dialog must be coherent, 
and their responses must be helpful to the user. A user 
model (UM), which can be modified during the interac- 
tion to represent updated beliefs about the current user, 
is one mechanism that can contribute to a robust, 
coherent, and cooperative dialog. 
In general, when we as speakers describe certain 
situations, we try to communicate these situations to 
our listeners. As proposed by some researchers 
(Webber 1978, Kamp 1984), speakers do so by attempt- 
ing to get their listeners to construct an appropriate 
model: a discourse model. A discourse model (DM) is 
viewed as containing representations of entities, along 
with their properties and relations they participate in. 
The key, then, in successful communication is for the 
speaker to transmit as much information about those 
entities, their properties and relations to the listener so 
as to achieve the goals of the current interaction. From 
the point of view of a system, a computational discourse 
model is used by the system to generate and/or interpret 
a discourse. 
This paper focuses on the relationship between DMs 
and UMs. It starts by describing what a DM is, and the 
role it plays in a coherent di~dog. It then describes what 
a UM is, and the role it plays in a cooperative dialog. I 
argue that the DM should be viewed as one part of the 
UM--that is, as one part of the system's model of the 
user. The examples of the natural language interactions 
are presented in the context of a natural language 
interface to an expert system that provides advice on 
cooking with chilies. (Part of the data was taken from 
the section "Cooking with Chilies" that appeared in 
Bon Appetit magazine, December 1986. The expert 
system can provide information about the different 
varieties of chili peppers as well as descriptions of how 
to "turn down the heat" of the chilies (make them less 
spicy), and how to cook with them without getting any 
kind of skin or eye irritations.) I justify this by showing 
how DMs can be viewed as part of UMs and how both 
models can affect each other. In other words, part of the 
UMs that systems have correspond to the DM, that is, 
a representation of what is talked about in a specific 
interaction. This piece, which changes with each dis- 
course, affects the UM and varies from interaction to 
interaction. 
2 DESCRIPTION OF DISCOURSE MODELS 
A piece of discourse is a collection of utterances that are 
spoken by one or more speakers. Usually the sentences 
in a discourse are connected in a way that makes them 
comprehensible and coherent. One way in which sen- 
tences in a piece of discourse are connected is via the 
use of anaphoric expressions. In general, anaphoric 
expressions refer to things that have been mentioned 
previously in clauses. (Note that there may be cases in 
which the anaphoric expression may refer to an entity 
that will be mentioned afterwards (i.e., cataphora) 
rather than before it, as in the following: 
i. After he finished the race, John went drinking to 
celebrate his victory. 
In this paper, I am concerned only with anaphoric 
expressions that refer to entities previously mentioned, 
such as: 
ii. After John finished the race, he went drinking to 
celebrate his victory). 
In English, anaphoric pronouns contribute to coherence 
in the discourse by avoiding repetitions of entities 
already mentioned. Consider the following: 
1. John went to the store and bought a pepper. He 
then went home to cook with it, 
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-0503.00 
82 Computational Linguistics, Volume 14, Number 3, September 1988 
Ethel Schuster Establishing the Relationship Between Discourse Models and User Models 
where instead of repeating "John" and "pepper" we 
have used the pronouns "he" and "it", respectively. 
We use sentences in the discourse to describe certain 
situations to our listeners. This we do by attempting to 
get our listeners to construct an appropriate model: a 
discourse model. A speaker's DM enables him to gen- 
erate what he believes will be coherent utterances. 
Similarly, the listener's DM enables him to comprehend 
discourse in an organized manner. Several researchers 
(Webber 1978, Kamp 1984, Heim 1982, Sag and 
Hankamer 1984) have been concerned with how DMs 
can be used to identify the referent of an anaphoric 
expression. (Not all these authors use the term "dis- 
course model". For instance, Kamp (1981) describes 
the utterances as being represented in a discourse 
representation structure (DRS). The entities mentioned 
in the sentence are represented in the DRS and they are 
called discourse referents (DRs). Heim's (1982) frame- 
work is the File Change Semantics.) They have sug- 
gested that speaker and listener each build a model of 
the discourse from the incoming sentences, including 
representations of the entities introduced by the dis- 
course, their properties, and the relations they partici- 
pate in. When an entity is later referred to via an 
anaphoric expression, the discourse participants can 
use their DM to make the appropriate link to an entity 
and hence interpret that anaphoric expression correctly. 
Some of the work on anaphora has concentrated in 
describing what characterizes the entities in the DM. 
For instance, Webber (1978) looked at the problem of 
definite noun phrases (where the references are to 
individuals and sets) and Schuster (1986, 1988) is look- 
ing at references to events and actions. The description 
of how things, sets, events, actions, facts, and so on, 
are represented in a discourse model and how one can 
refer to them gives us a clue to what characterizes a 
discourse model. Because the representations in the 
discourse model are of specific objects or events which 
are talked about during the interaction, the discourse 
model can be viewed as a temporary knowledge base. 
Since a discourse has relatively short duration, the 
discourse model that supports the interaction contains 
short term or temporary information. 
It is important to note that the representations of 
entities, as they appear in the discourse have a structure 
as proposed by Grosz and Sidner (1986). While Grosz 
and Sidner do not specifically deal with discourse 
models, their view on discourse is applicable to dis- 
course models. The discourse model reflects the struc- 
ture of the dialog. In the same way that items are 
highlighted in the actual discourse, they appear as being 
more salient in the discourse model. Because some 
items are more salient than others, the representation is 
not just a flat representation, but has a hierarchical 
structure in which the more salient entities are repre- 
sented in the same way as they appear in the discourse. 
The structure is needed because the ordering of the 
representations does not necessarily correspond to the 
order in which the entities are mentioned in the dis- 
course. A focusing mechanism plays a very important 
role in understanding discourse. This mechanism is 
needed to process sentences at any point in the dis- 
course by indicating which objects, things, events, or 
facts are more salient at any point in the discourse. 
When processing a part of discourse, only those entities 
that are salient come into play. 
3 A VIEW OF USER MODELS 
In this paper, the UM is viewed as "the system's beliefs 
about its users". Many views have been proposed to 
describe what UMs are. The various UMs proposed so 
far fall under the general category described here. For 
example, this definition of UMs includes McCoy's 
(1985) concept of a UM: the system's beliefs about how 
the user views objects in the domain. It also includes 
Paris's definition of a UM: the system's beliefs about 
the user's levels of expertise as well as the definition of 
UMs as viewed by researchers concerned with plan 
recognition: the system's beliefs about what the user is 
trying to do. 
Many distinctions have been made when character- 
izing user models. Kobsa (1985) and Kass and Finin 
(this issue) distinguish between user models and agent 
models. (Kobsa actually uses the term Akteurmodell 
(actor model) since, according to him, the primary 
meaning of the German Agent is "secret (foreign) 
agent".) For them, the agent model is the model of the 
person that the system can model and there can be 
many agent models. The user model is the model of the 
specific agent that interacts with the system. Often the 
agent model and the user model coincide. In this paper, 
I will assume that this is the case. Also, Rich (1979) 
distinguishes between models of individual users and 
models for classes of users, as well as between long- 
term as compared to short-term UMs. This notion of 
short- and long-term UMs provides a spectrum of parts 
of the UM, some of which are temporary and some of 
which remain after the discourse ends. I will show more 
on this issue in the next section. 
In this paper, I assume that the system has represen- 
tations for three possible stereotypes of users: a begin- 
ner, an intermediate, and an expert. The system can 
modify its own user model as the interaction occurs, as 
a result of the information that flows out from the 
discourse model into the user model. Thus the user 
model is dynamic. In general, information that is rele- 
vant to the user and which is represented in the DM 
becomes part of the UM. 
Consider a simulated expert system HOT, which 
provides information about cooking with chilies. The 
system provides advice to aficionados (amateurs) about 
buying, cutting, peeling, storing, and cooking with 
chilies. The system also has a general UM from which it 
can identify three possible users: beginner, intermedi- 
ate, and expert. These are canonical UMs, and they are 
Computational Linguistics, Volume 14, Number 3, September 1988 83 
Ethel Schuster Establishing 'the Relationship Between Discourse Models and User Models 
representations of three potential classes of users of the 
system. The beginner stereotype contains information 
about simple and well-known varieties of chili peppers. 
It also contains information about storing chilies. The 
stereotype for intermediate users assumes that the user 
knows more than a beginner, while an expert is assumed 
to know about unusual varieties of peppers and to be 
interested in more sophisticated information concerning 
chili peppers and detailed information about using dif- 
ferent kinds of them. 
The users interact with the system by asking ques- 
tions. From these, HOT can decide how to fit each user 
into any of the particular UMs that it has available. 
Also, the sample responses from HOT are used as a way 
of demonstrating how the UM participates in the dis- 
course. In other words, the responses show evidence of 
interaction between the UM and the DM. The following 
example illustrates the interaction between HOT and 
one of its users. 
2. U: Hi! I love to eat spicy food and I love to cook 
with chilies. I just found some fresh peppers in the 
health food store called banana-peppers and I was 
told they are very hot. How can I peel them? 
From this introduction the system can deduce that the 
user is an intermediate user in cooking with chilies, and 
invokes the stereotype for intermediate users. How 
does the system decide that this user is an intermediate 
and not a beginner? Firstly, the user explicitly mentions 
that he likes to eat and cook spicy food. Also, the 
system can realize that a more experienced person in 
spicy food knows about the need to peel hot chilies 
(sometimes), while a novice may not realize that some 
kinds of peppers need to be peeled. And an expert 
would know how to peel hot peppers. These facts 
trigger the intermediate stereotype in the user modeling 
system. Notice that the user mentions specific entities 
(e.g., himself, peppers, health food stores, and so on) as 
well as events: "user loves to eat spicy food", "user 
cooks", and so on. All these entities and event descrip- 
tions are represented in the DM and they are used to 
infer the correct level of the user in the UM. This fact is 
evidence that the DM is part of the UM. Once the 
system has decided that the user is an intermediate, it 
can respond not only in terms of what the user wants to 
know, but also what will be most helpful to the user. 
4 RELATIONSHIP BETWEEN DM AND UM 
In the previous sections I have shown the role of the 
DM in a user-system interaction. I have also described 
the role that a UM plays in a user-system interaction. 
The system uses the information in the UM to decide 
what kind of user it is interacting with, as well as how to 
respond to the particular user. 
Given the definition of UMs in the previous section, 
the DM seems to clearly be part of the UM, that is, it is 
the system's beliefs about what the user believes about 
the discourse. The question then is whether the DM is 
the system's beliefs about the discourse or is it the 
system's beliefs about the user's beliefs about the 
discourse. I would argue that it is the latter. It has been 
claimed lhat both dialog participants must be focused on 
the same subset of knowledge for communication to be 
successfal. If the system has a DM that allows it to 
comprehend utterances one way and the user has a DM 
that causes it to interpret an utterance differently, the 
interaction is going to fail. So if the system is going to 
use its DM to generate utterances that it believes the 
user can understand as the system intended, then it 
must believe that its DM reflects the user's beliefs about 
what has been talked about. (One might argue that we 
have to go all the way to mutual beliefs--namely, that 
the DM is the system's beliefs about what is mutually 
believed about the discourse.) 
Furthermore, if the DM were separate from the UM, 
then an entity introduced by the discourse could always 
be referred to. But that may not be possible unless the 
system believes the user knows about this particular 
entity. On the other hand, if the DM is part of the UM, 
then only those entities that the system believes the user 
knows about can be represented implicitly in the DM, 
since in this case the DM must represent the system's 
beliefs about the user's beliefs about the discourse. 
Then the system can only coherently refer to entities 
that it believes the user knows about, since these are the 
only ones represented in its DM. 
In the previous section, I described a view of UMs 
with three stereotypes. Pictorially, this can be seen as a 
kernel of information with several possible levels: 
I Expert 
\[ ....................... 
i i Inter I 
II ............... I 
Ill Novice I t 
II ............... I 
\[ ...................... 
........................... 
The INITIAL-UM is the representation of the UM that 
the system has initially (before any interaction). During 
its interaction with the user, the system builds the DM. 
In turn, information taken from this DM is used to 
update the INITIAL-UM into an UPDATED-UM. The 
UPDATED-UM becomes a FINAL-UM when the in- 
teraction ends (possibly after several updates). Note 
that only parts of the FINAL-UM persists for future use 
after the current interaction ends. 
All the information that the user provides is repre- 
sented in the DM. Consider the following: 
3. U: I want to know how to peel banana-peppers. 
Imagine, my mother was in Mexico and I asked 
her to buy some for me. She decided to try one of 
them and she burnt her throat. She had to be 
rushed to the hospital, blah, blah, blah. 
84 Computational Linguistics, Volume 14, Number 3, September 1988 
Ethel Schuster Establishing the Relationship Between Discourse Models and User Models 
This information is also part of the UM. Given the 
definition of the UM as the system's beliefs about the 
user, then this information provided by the user is the 
system's beliefs about the user's beliefs about what has 
occurred. For instance, now the system believes that 
the user believes that you can buy banana-peppers in 
Mexico. 
In replying to its users, the system not only decides 
what information to include in the reply, but can also 
use anaphoric expressions (i.e. pronouns) in its re- 
sponses. The only way the system could have used 
those pronouns was by having a representation of the 
discourse in which the mentioned entities were repre- 
sented and available for reference. Also, since the 
system responded in terms of its model of the users, 
only if the DM is part of the UM, can the system take it 
into account in its responses and its reasoning about the 
users. Both the UM and the DM were needed in 
creating the response, not only because of the specific 
information used in the response, but also in the way in 
which that information was actually presented to the 
users. In other words, the DM is part of what the system 
needs to consult when responding to its users. 
One of the ways to identify how the UM contains the 
DM is by looking for what information might be in the 
UM but not in the DM. In the earlier examples, the 
responses generated by HOT made use of information 
taken from the stereotype invoked for the individual 
user. This information was not present in (or implied by) 
the previous discourse. Hence the UM contains infor- 
mation that does not appear in the DM. 
Note also that the DM can affect the rest of the UM. 
Suppose a user comes often in contact with the system, 
and keeps referring to the same things. After several 
interactions, these things the user mentions should 
eventually become part of the long-term UM. The 
question that is left is whether it is indeed worthwhile to 
perform this transfer from the DM into the long-term 
UM. For instance, if a user talks about the same things 
over a course of several interactions and the informa- 
tion is moved to the UM, what happens if the user stops 
talking about those specific things? Do we then delete 
the information from the UM and allow for new infor- 
mation to come in? Also, with respect to the short- and 
long-term UMs, we could consider the short-term parts 
to be the DM, which is removed once it is no longer 
relevant. The intermediate parts could correspond to 
the beliefs that the system has about what the user is 
trying to do. And the long term would be the beliefs 
about the user's level of expertise, his likes, or dislikes. 
These are among the many issues that remain to be 
solved. 
ACKNOWLEDGEMENTS 
Many thanks go to Kathy McCoy, Sandy Carberry, Bob Kass, Julia 
Hirschberg, Alfred Kobsa, and Tim Finin for their discussions that 
helped me in understanding this controversial issue as well as for their 
comments on earlier drafts of this paper. I would also like to thank the 
participants of the 1986 User Model workshop at Maria Laach for 
their comments during the heated discussion that this subject gener- 
ated. 

REFERENCES 
Grosz, B. and Sidner, C. 1986 Attention, Intention, and the Structure 
of Discourse, Computational Linguistics 12: 175-204. 
Heim, I. 1982 The Semantics of Definite and lndefinite Noun Phrases. 
Ph.D. thesis, University of Massachusetts, Amherst, MA. 
Kamp, H. 1984 A Theory of Truth and Semantic Representation. In 
Groenendijk, J.; Janssen, T.M.V.; and Stokoff, M. (eds.), Truth, 
Interpretation and Information, Foris Publications: 1--41. 
Kass, R. and Finin, T. 1987 Modelling the User in Natural Language 
Systems. Computational Linguistics forthcoming, 1987. 
Kobsa, Alfred 1985 Benutzermodellierung in Dialogsystemen. Ph.D. 
thesis. Springer-Verlag, Berlin--New York. 
McCoy, Kathleen 1985 Responding to Users Misconceptions. Ph.D. 
thesis, University of Pennsylvania, Philadelphia, PA. 
Rich, Elaine 1979 User Modeling Via Stereotypes. Cognitive Science 
3: 329-345. 
Sag, I. and Hankamer, J. 1984 Toward a Theory of Anaphoric 
Processing. Linguistic and Philosophy: 325-345. 
Schuster, Ethel 1986 Towards a Computational Model of Anaphora in 
Discourse: Reference to Events and Actions. Ph.D. dissertation 
proposal. Technical Report MS-CIS-86-34, University of Pennsyl- 
vania, Philadelphia, PA. 
Schuster, Ethel 1988 Ph.D. thesis, University of Pennsylvania, Phil- 
adelphia, PA. Forthcoming. 
Webber, B.L. 1978 A Formal Approach to Discourse Anaphora. 
Ph.D. thesis, Harvard University, Cambridge, MA. \
