USER MODELS AND DISCOURSE MODELS: 
UNITED THEY STAND . . . 
Alfred Kobsa 
SFB 314: AI-Knowledge-Based Systems 
Department of Computer Science 
University of Saarbriicken, W. Germany 
Opinions on the relationship between discourse models 
(DMs) and user models (UMs) are obviously influenced 
by preassumptions about their respective contents. As 
far as DMs are concerned, two divergent views have 
been expressed in the discussion published here: 
I. The DM contains only representations of the 
objects mentioned so far in the discourse (i.e., a 
mentioned-object memory--see Schuster, this 
issue). The term "object" will be used here in the 
broad sense of Schuster, thus also denoting 
events, properties, etc. 
2. The DM contains in addition 
a. a representation of the purpose underlying 
the segments of the dialog (i.e. a dialog 
purpose--see Grosz Sidner 1986, Chin, this 
issue). 
b. an attentional structure, which is a subset of 
the representations mentioned in (1) contain- 
ing the currently focused objects which are 
ordered in a focus stack (Cohen, this issue; 
Chin, this issue, who requires only that the 
user must be familiar with these objects). 
Less disagreement seems to exist about the components 
of a UM. Generally, it is regarded as containing explicit 
representations of the system's assumptions about all 
relevant aspects of the user, i.e., assumptions about his/ 
her "objective situation" (e.g., marital status, number 
of children), as well as about his/her prior knowledge, 
goals, plans and false beliefs with respect to the domain 
of discourse. In order to meet Wahlster's personnel- 
database counterexample, it must be further required 
that the user model be separable by the system from the 
rest of the system's knowledge. 
To discuss the relationship between DMs and UMs, 
a general belief, goal, and plan maintenance system 
(BGP-MS) will be presented here, the purpose of which 
is to store and update the beliefs, goals, and plans of 
both the system and an arbitrary number of other 
agents, including the system's current user. Specific 
subcomponents and subfunctions of this system hope- 
fully capture the general consensus on what constitutes 
a discourse model and a user model, respectively. 
However, we will see that these subcomponents are 
strongly interwoven and that--apart from a few rarely 
occurring exceptions--the DM is part of the UM at least 
at the level of content. The question arises then, of 
course, whether it makes sense to separate these no- 
tions conceptually. 
The belief, goal, and plan maintenance system out- 
lined here is being implemented (in a somewhat simpli- 
fied form) in XTRA, a natural language access system to 
expert systems (Allgayer et al. 1988). A previous imple- 
mentation was VIE-DPM (Kobsa 1985a,b). In the 
knowledge base of BGP-MS, the representation of the 
various types of (nested) beliefs and goals (Kobsa 1988) 
is separated into a number of hierarchically ordered 
partitions (see Figure 1). If it is shared knowledge 
between S and U that U possesses certain beliefs 
(knowledge), then this knowledge or these beliefs are 
represented in MB(UB). 1 MB(UW) contains those goals 
and plans of the user, MB(SB) those beliefs of the 
system, and MB(SW) those goals of the system for 
which the same holds true. "Private" beliefs of the 
system about the domain of discourse / about the user's 
beliefs / about the user's beliefs about the system's 
goals are represented in SB, SBUB, and SBUBSW, 
respectively. MB contains the mutual beliefs (knowl- 
edge) with respect to the domain, and MW the mutual 
goals and plans of S and U. The arrows between the 
partitions denote inheritance relationships. 
In the partitions of BGP-MS, the content of the individ- 
ual beliefs, goals, and plans can be expressed through 
arbitrary representational structures (e.g., a KL-ONE- 
like representation as used in XTRA). Various markers 
for non-belief and uncertainty can be added: For in- 
stance, in SBUB it can be expressed, among other 
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/88/0100e-e$ 03.00 
Computational Linguistics, Volume 14, Number 3, September 1988 91 
Affred Kohsa User Models and Discourse Models: United they stand... 
MW 
t 
SB 
..................... i .............................. i \[ i 
i 
t 
\[ J sso B I 
IqlB 
Figure 1. Hierarchical belief, goal and plan representation 
in BGP-MS. 
things, that S is uncertain whether (or does not believe 
that) U knows some fact; and in MB(UB), that S is 
uncertain (or does not believe) that a belief of the user 
is mutually known. 
Where are the user model and the discourse model 
located in this architecture? The UM part of BGP-MS 
consists of all partitions except SB and SW, plus all 
representations in SB (and probably SW) in which an 
individual constant occurs denoting the user (the rest of 
SB corresponds to Sparck Jones's world model). The 
DM cannot be so easily identified. In the following 
sections I will discuss how the different functions of a 
discourse model as outlined above can be fulfilled by 
the proposed architecture. 
1. THE MENTIONED OBJECT MEMORY 
When an object is mentioned during the discourse, then 
mutual knowledge about its existence is usually estab- 
lished. Thus a representation of the object can be 
entered into MB (implying, for instance, that S can now 
use definite NPs to refer to these objects). Finer dis- 
tinctions in what is now known can also be expressed: 
to take Chin's example, if S mentions a name unknown 
to U (i.e., the relationship between the name and its 
bearer is contained in SB only), then the existence of a 
person with this name can be entered into MB. In 
MB(UB) it can be represented that U does not know, 
and in bAB_(SB) that S does know the bearer of this 
name. 
All of the mentioned partitions are part of the user 
model. One might argue that the above architecture may 
completely cover the mentioned object memory func- 
tion in an NL dialog system. However, three kinds of 
information are lost thereby (this defect also seems to 
apply partly to Schuster's model). 
a. The information that objects have been explicitly 
mentioned in the discourse: 
MB contains not only those objects which were 
explicitly mentioned in the discourse (and which 
are therefore mutually known by all dialog partic- 
ipants), but also representations of those objects 
whose existence is mutually known due to stereo- 
types (Rich 1988) or to inferences from the dis- 
course. Sometimes, however, the system should 
possess information about whether or not some 
object had been explicitly mentioned, for example 
in order to increase coherency in its own dialog 
contributions ("As I/you said before... ") or to 
point out inconsistencies in dialog contributions of 
the user ("But previously you said..."). 
b. Information about the sequence (and thus recency) 
of objects' mention: 
This information is very important in NL systems, 
since the choice of various forms of anaphora 
depends on the degree of recency. 
c. hlformation about the linguistic structure of dialog 
contributions: 
Sometimes the system should also possess infor- 
mation about the wording or the syntactic struc- 
ture of the user's and the system's previous dialog 
contributions (i.e., information on how objects 
have been mentioned). This information can be 
exploited by the system for reiterating a descrip- 
tion in its own dialog contributions or for avoiding 
reiteration, for instance. 
In XTRA, two additional knowledge bases have been 
introduced which serve the above-mentioned functions, 
among others: the FSS knowledge base and the so- 
called Linguistic Dialog Memory. The FSS (Allgayer 
and Reddig 1986) represents the functional semantic 
structure of both the user's and the system's dialog 
contributions. FSS contents can also be linked to the 
linguistic surface forms (NPs, PPs, etc.) which caused 
their creation (in the case of user input) or became their 
linguistic realizations (in the case of system dialog 
contributions). The dialog memory, among other things, 
records the objects that have been mentioned during the 
discourse, and, in its dialog sequence part, the sequence 
of the objects' mention. 
In general, all system knowledge about what objects 
have been mentioned in the on-going discourse, in what 
92 Computational Linguistics, Volume 14, Number 3, September 1988 
Alfred Kobsa User Models and Discourse Models: United they stand... 
order they were mentioned, and how they were de- 
scribed (i.e. all parts of the dialog memory) are regarded 
as being part of MB, and hence part of the user model. 
This is necessarily so, since, by definition, MB contains 
all knowledge that is shared by system and user. And 
only if knowledge about the previous discourse is 
shared between both participants can it be safely em- 
ployed in the generation of dialog contributions. (For 
example, an anaphor generated by the system will 
probably fail to fulfill its referential function if only the 
system, but not the user, believes that its intended 
referent has been mentioned just recently. Hence the 
system should check whether its records in the dialog 
sequence memory are shared by the user.) 
In cases of communicative failure, however, there 
exists system discourse knowledge that is not shared by 
the user, and thus not part of MB and the user model. In 
such cases, entries are made into SB instead, and 
thereby form part of the system's knowledge only. For 
instance, when S assumes that U does not remember 
what has been said (see Wahlster's hastily-presented- 
names example), the FSS descriptions and the repre- 
sentations of their referents in the dialog memory can be 
entered in SB instead of MB, and thus do not form part 
of the user model. In addition, however, all sorts of 
uncertain assumptions about what the user has or has 
not, in fact, kept in mind can be expressed in SBUB or 
MB(UB), i.e., in the user model. (For example, the 
system can note in the user model that the user probably 
remembered the first two but not the subsequent 
names.) For simplicity, however, neither of the de- 
scribed cases of communicative failure is dealt with in 
XTRA, and for implementational reasons the FSS part 
of MB forms a separate partition. 
2A. THE DIALOG PURPOSE 
Here the problem arises in research on NL dialog 
systems as to whether or not one should regard the 
intentional structure of individual utterances (Cohen, 
this issue) or dialog segments (Grosz Sidner 1986) as 
being independent of the dialog setting and the dialog 
participants. To put it in a more provocative way, do 
dialog constituents or dialog participants have an inten- 
tional structure? In my view, the essence of problem- 
solving dialog lies in the recognition of the dialog 
participants' goals and plans, and in the construction of 
mutually known goals and plans. Dialog contributions 
of the dialog partner serve as a more-or-less helpful aid 
in this process (Pollack et al. (1982) present transcripts 
in which dialog contributions of clients with misconcep- 
tions even impair the recognition of their actual goals). 
Apart from that, no intentional character pertains to 
dialog constituents that is independent of the dialog and 
situational context and of the current (beliefs about) 
goals and plans. 
In the BGP-MS philosophy, (beliefs about) goals and 
plans are contained in those partitions whose labels 
include a W. A user's dialog contribution is first repre- 
sented by FSS structures in MB. If user plans or goals 
can be inferred by S, they are represented in MB(UW), 
or in MW if there is mutual knowledge that they have 
been accepted by S. Conversely, when S gradually 
communicates its plans or goals to U, the corresponding 
representation structures are transferred from SW to 
MB(SW), and finally hopefully to MW. Thus any sort of 
intentional structure is part of the user model. 
2B. THE ATTENTIONAL STRUCTURE 
I agree with Chin (this issue) that the attentional struc- 
ture is also not a context-independent characteristic of 
discourse (although Chin's notion of attentional struc- 
ture seems to be broader than mine). Only mutually 
known objects can be in focus. In XTRA, focus is 
expressed by focus values in the dialog memory which, 
logically, can only be applied to representation struc- 
tures in MB and MW and therefore form part of the user 
model. 
SUMMARY 
The above discussion demonstrates that the function of 
a UM and of all mentioned DM components can be 
completely fulfilled by the outlined belief~ goal, and plan 
maintenance system. I cannot deal here in detail with 
the question of whether this is also the case for other 
components that have been proposed for a DM, for 
example, the structuring of the dialog into dialog seg- 
ments (Grosz Sidner 1986), a context space grammar 
(Reichman 1981), or rhetorical predicates (schemata; 
McKeown 1982). With respect to the analyzed compo- 
nents, we have seen that the discourse model almost 
completely overlaps with the user model at the level of 
content. Only if the user does not fully catch the 
system's dialog contributions are entries in the DM 
created which do not form part of the UM (see, for 
instance, Wahlster's example, this issue). But at a 
procedural level as well, only a few processes can be 
found which operate exclusively on that part of the user 
model that is identical with the discourse model, or 
upon the remaining parts of the user model. 
This large degree to which the DM is included in the 
UM, however, is not surprising: Discourse models are 
ultimately based on linguistic conventions. In order for 
the linguistic, intentional, attentional, etc., structure of 
the previous discourse to be exploited for future dialog 
contributions, conventions about what the structure of a 
particular ongoing dialog actually is must exist. Knowl- 
edge about convention is mutual knowledge, however, 
(Lewis 1969, Schiffer 1972), and thus part of MB. The 
same holds true for the above-mentioned additional 
components of the DM that could not be dealt with in 
this paper. And, by the way, it also holds true for the 
grammar the system employs (but see the opposing 
views of Morik and Wahlster, this issue). If the system 
did not assume that its assumptions about the syntactic 
structure of language (as expressed in its grammar) be 
Computational Linguistics, Volume 14, Number 3, September 1988 93 
Alfred Kobsa "Jser Models and Discourse Models: United they stand... 
shared by the user, it could not justifiably use it in the 
analysis and generation of dialog contributions without 
risking miscommunication. And there definitely exists 
work in user modeling (e.g., Schuster I985, Kilbury 
1986, Lehman CarboneU 1988), which is concerned with 
the recognition of those parts of a user's idiosyncratic 
grammar that deviate from the mutually shared kernel 
grammar. Of course, an entry in MB never means that 
the system assumes that the user "has the same struc- 
ture in his/her mind" (e.g., ATNs, KL-ONE, or LISP), 
but only that these structures are functionally equiva- 
lent reconstructions of the user's competence. 
Does the large degree of inclusion of discourse mod- 
els in user models at the level of content imply that the 
notion of discourse model is superflous? As was pointed 
out by Morik (this issue), extensionally overlapping 
notions may still prove useful if their intension high- 
lights different aspects of a system. For example, in the 
above architecture, such a concept might characterize 
an orthogonal substructure and denote, for instance, 
entries in different partitions with specific origin or 
function. The above as well as Morik's and partly 
Wahlster's discussions demonstrate, however, that it is 
very hard to find such differential criteria for DMs. I 
therefore suspect that a happy fate of that kind will 
more probably apply to notions such as mentioned 
object memory or discourse sequence memory than to 
the vague notion of discourse model. 
ACKNOWLEDGEMENT 
This research was supported by the German Science Foundation in its 
Special Collaborative Programme on AI and Knowledge-Based Sys- 
tems (SFB 314). I am indebted to Carola Reddig and Norbert 
Reithinger for their comments on an earlier version of this paper. 

REFERENCES 
Allgayer, J. and Reddig, C. 1986 Processing Descriptions Containing 
Words and Gestures: A System Architecture. In Rollinger, C. R. 
and Horn, W. (eds.) GWA1-86 und 2. Osterreichische Artificial- 
lntelligence-Tagung. Springer, Verlag, Berlin--New York. 
Allgayer, J.; Harbusch, K.; Kobsa, A.; Reddig, C.; Reithinger, N.; 
Schm~.uks, D. 1988 XTRA: A Natural-Language Access System to 
Expert Systems. Technical Report, SFB 314: AI-Knowledge- 
Based Systems, Department of Computer Science, University of 
Saarbrficken, W. Germany. 
Kilbury, J. 1986 Language Variation, Parsing, and the Modelling of 
User's Language Variations. In Proceedings of the 7th European 
Conference on Artificial Intelligence, Brighton, England: 29-32. 
Kobsa, A. I985a Benutzermodellierung in Dialogsystemen. Springer- 
Verlag, Berlin--New York. 
Kobsa, A. 1985b Using Situation Descriptions and Russellian Atti- 
tudes for Representing Beliefs and Wants. In Proceedings of the 
International Joint Conference on Artificial Intelligence, Los 
Angeles, CA: 513-515. 
Kobsa, A. 1988 A Taxonomy of Beliefs and Goals for User Models in 
Dialog Systems. In Kobsa, A. and Wahlster, W. (eds.), User 
Models in Dialog Systems. Springer-Verlag, Berlin--New York. 
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. K. 1969 Convention: A Philosophical Study. Harvard 
University Press, Cambridge, MA. 
McKeown, K. R. 1982 Generating Natural Language Responses to 
Questions about Database Structure. TR MS-CIS-82-5, Depart- 
ment of Computer and Information Science, University of Penn- 
sylvania, Philadelphia, PA. 
Pollack, M.E.; Hirschberg, J.; and Webber, B. 1982 User Participa- 
tion in the Reasoning Process of Expert Systems. MS CIS-82-9, 
Department of Computer and Information Science, University of 
Pennsylvania, Philadelphia, PA. 
Reichman, R. 1981 Plain Speaking: A Theory and Grammar of 
Spontaneous Discourse. Report No. 4681, Bolt, Beranek and 
Newman, Cambridge, MA. 
Rich, E. 1988 Stereotypes and User Modeling. In Kobsa, A. and 
Wahlster, W. (eds.), User Models in Dialog Systems. Springer- 
Verlag, Berlin--New York. 
Schiffer, S. R. 1972 Meaning. Clarendon Press, Oxford, England. 
Schuster, E. 1985 Grammars as User Models. In Proceedings of the 
International Joint Conference on Artificial Intelligence, IJCAI- 
85, Los Angeles, CA: 20-22. 
