ON THE RELATIONSHIP BETWEEN USER MODELS AND 
DISCOURSE MODELS 
Robin Cohen 
Department of Computer Science 
University of Waterloo 
Waterloo, Canada N2L 3G1 
The best way to summarize my view of the relationship 
between user models and discourse models is that they 
are separate, but related to each other. This paper will 
show why the two terms have completely distinct 
elements, and where the common ground between the 
two lies. It is also important to acknowledge at the 
outset that the two terms have not been well defined in 
the literature. 
For the discourse model, I am including everything 
that should be derived from an analysis of discourse, to 
present a representation for the structure of the dis- 
course, useful in subsequent responses. In this sense, I 
focus on the interpretation of a discourse from the point 
of view of one of the conversants. I essentially include 
in the discourse all the components covered by the 
model of Grosz and Sidner (1986). For the definition of 
the user model, I also ground the discussion in the point 
of view of one conversant. The model is thus an analysis 
of the other conversant (subsequently referred to as the 
speaker). The term user model is especially obscure, 
because in the context of this journal it is confined to a 
derivation of background knowledge and goals of a user 
which influence the language used in the discourse. 
(One can use a similar term in the design of graphical 
interfaces, for instance). 
The discourse model must thus contain the following 
key elements: an indication of the structure of the 
discourse and an organization of the objects of the real 
world mentioned in the discourse (to help anaphora 
resolution, for example). As soon as this kind of history 
of objects is included (covered in the model of Grosz 
and Sidner (1986) by tracking attentional state and the 
objects currently in focus), there are elements that are 
not specifically attached to the user himself. 
The structure of the discourse is essentially provided 
in two different ways. Which of the actual utterances of 
the discourse group together into logical segments is 
covered by the "linguistic structure" of Grosz and 
Sidner (1986). Often clue words (such as "but anyway") 
will indicate how to segment the utterances into logical 
segments, without concern for how individual utter- 
ances within that segment relate. In addition, there is an 
indication of the intentional structure. Here, I would 
reinterpret slightly the term as used in Grosz and Sidner 
(1986) (see Cohen 1986). Intentional structure should 
indicate the intentional relations between, again, actual 
utterances. For instance, it is important to determine 
the cases where the goal underlying an utterance "con- 
tributes to the satisfaction of" the goal underlying 
another utterance--e.g., getting the hearer to believe 
some proposition p contributes to the satisfaction of 
getting the hearer to believe some proposition q (deter- 
mined as dominance relations in Grosz and Sidner 
(1986). In this sense, my interpretation of the derivation 
of intentional structure agrees well with Wahlster's 
appeal for an incremental derivation of the discourse 
model). 
I believe that the intentional structure is related to, 
but not identical with, the plan of the speaker underly- 
ing discourse. For one, the plan of the speaker can be 
reconstructed at a different level of detail than what is 
actually uttered. For an example, see Appendix 1. 
This leads me to where I feel the discourse model and 
user model relate. The plan of the speaker underlying 
discourse is one part of the user model and is related to 
the intentional structure of the discourse. But there is 
more to the user model as well. There has been a good 
deal of work on co-operative responses (e.g., Joshi et al. 
(1984)). Van Beek (van Beek and Cohen 1986, van Beek 
1986) shows that goals and systemwide preferences of a 
user can influence appropriate responses (e.g., prefer- 
ring to take numerical analysis courses in a course 
advisor domain). It is thus important to include a model 
of the user's goals (beyond an understanding of the goal 
underlying each individual utterance, useful for the 
reconstruction of the intentional structure). This kind of 
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. 
0362-613 X/88/0100e-e$03.00 
88 Computational Linguistics, Volume 14, Number 3, September 1988 
Robin Cohen On the Relationship Between User Models and Discourse Models 
goal is not part of the discourse per se. Other important 
components of the user model are background knowl- 
edge of the user (e.g., Cohen and Jones (1988) show that 
it is important to vary the response to a parent vs. a 
teacher in a domain of educational diagnosis; Paris 
(1985) shows how the form of response can vary accord- 
ing to the level of expertise; Chin (1986) also uses the 
level of expertise of the user, together with a labeling of 
difficulty of the system's knowledge, to produce good 
responses). Especially if the user is modeled over a 
period of time, these "values" can change and must be 
monitored. 
The bottom line, I feel, is that what is included in the 
user model or the discourse model is dependent on what 
the system employing these models is being designed 
for. (Again, I agree with Wahlster's perception of the 
problem--the discussion of what is in the UM or DM 
can be guided by how systems should be designed). 
I prefer the point of view of a NLUS, taking the role 
of a conversant, analyzing the discourse of the speaker 
(using a model of the user as well as an incrementally 
built model of the discourse), to be used to eventually 
respond. If a representation to facilitate response is 
what is needed, it seems clear that both a picture of the 
discourse as it proceeded and an understanding of the 
person producing the discourse will be important dis- 
tinct factors. Understanding the structure enables the 
hearer to comprehend the points made by the speaker, 
to then evaluate and address a response. Deeper know- 
ledge of the speaker will then facilitate constructing a 
response that can be well understood (for which the goal 
of the hearer in producing the response will succeed). 
For an example, see Appendix 2. 
I will close with some comments about the terminol- 
ogy used by Schuster. I feel that the definition of 
discourse model here is too narrow--there is more to a 
model of discourse than an indication of the underlying 
entities (objects, events). Schuster seems to suggest 
that some of the structuring provided in Grosz and 
Sidner (1986) is there only to highlight the entities, In 
my view, the actual utterances themselves are worth 
examining as participating in some structure. 
I also find Schuster's definition for user model--the 
information a system has about the userAsomewhat 
problematic. I think that the user model must concen- 
trate on dynamic information, that is, which has some 
potential for change. In any case, the information 
should be such that different values make for different 
analyses ( of the discourse where the user model is 
derived). Otherwise, why have a model at all? So if all 
the users of a system are male, why record this fact in 
the user model for each one? (My views here thus 
coincide with Sparck Jones's claim that a user should be 
modeled if there are particular characteristics which set 
her apart.) 
Finally, relevant to Schuster's discussion on agent 
models and user models (see also Wahlster and Kobsa 
1988; Kass and Finin, this issue), I reiterate that the 
focus should be on the user as conversant. If the topic 
of conversation is another agent, it is useful to know 
about this person, in the same sense that it is useful to 
know about any topic discussed (e.g., the working of a 
nuclear power plant). (Note that the system's and the 
user's view of the topic may not coincide, and thus this 
view of the world may need to be modeled of the user as 
well). 
1. See that screw 
2. The one with the funny top 
3. Loosen it with the wrench 
4. That black wrench there 
5. OK--now you can slip in the pliers 
6. And the whole pole comes off 
Plan of Speaker: 
The top level goal is get pole off, which succeeds if the following hierarchy of subgoals succeeds: 
get po~ off ~_ 
loosen screw with wrench -~" slip in pliers 
identi~/scr/ew ~ identify wrench 
know chars, of screw know c'~hars, of wrench 
Intentional structure of discourse (as in Grosz and Sidner 1986): 
Primary Intentions: 
I1: intend H (get pole off); 
I2: intend H (loosen screw with wrench) 
I3: intend H (identify screw) 
Computational Linguistics, Volume 14, Number 3, September 1988 89 
Robin Cohen On the Relationship Between User Models and Discourse Models 
Segmentation Structure: 
( ( ( 1 2 (ds3)) 3 4 (ds2)) 5 6 (dsl)) 
There are three segments: ds3 with 13, ds2 with 12, and dsl with I1, where 12 DOM 13 and I1 DOM 12 (i.e. 
13 contributes to the satisfaction of 12, etc.) 
There are two main sources of difference between the 
plan of the speaker and the intentional structure of 
discourse, illustrated by the above example: (i) there 
may be no direct match from the utterances to the units 
(subgoals) of the plan; here, there is no utterance 
corresponding to "identify wrench", on top of utter- 
ance 4, which serves to let the hearer "know charac- 
teristics of the wrench"; (ii) the intentions recorded for 
the intentional structure may be at a higher level of 
detail. 
The examples provided in Grosz and Sidner (1986), 
for instance, only record those attached to segments of 
more than one utterance. There are, indeed, many issues 
regarding the relationship of plans and discourse struc- 
ture; we will not elaborate further here. Our main point is 
that the two terms should be related, but distinct. 
Appendix 1. Plans vs. Intentional Structure. 
Example: 
2. I. Prime Minister Mulroney is wonderful. 
2. He refuses to back off on the free trade plan. 
2b. 1. Prime Minister Mulroney is wonderful. 
2. For example, he refuses to back off on the free 
trade plan. 
In Example 2b, the structure of the discourse, indicated 
by the connecting phrase "for example", suggests an 
intentional connection between (1) and (2). (One way to 
view this discourse is as an argument where the speaker 
utters (2) in order to get the hearer to believe (1)). Clue 
words alone may provide a basis for the determination 
of the segmentation of this small example. 
Now, if we also know of the speaker (or derive, on 
the basis of the likely intended connection above) that 
he is an arch-conservative, we have additional informa- 
tion to facilitate response. One such rejoinder might be: 
2c. Yes, but won't this prevent big private compa- 
nies from making lots of money? 
This rejoinder would not be appropriate for a fiery labor 
supporter (e.g., New Democrat, in Canadian politics). 
Example 2 is also a case where having a model of the 
user's beliefs (e.g., he's an arch-conservative stereo- 
type) may facilitate derivation of the intended structure 
of the discourse, in the absence of clue words. This thus 
also argues for inclusion in the user model information 
additional to the structure of discourse determined so 
far. 
Appendix 2. Using Discourse Structure and User Models for Response. 

REFERENCES 
Chin, D. N. 1986 User Modeling in UC, the UNIX Consultant. In 
Proceedings of the Conference on Human Factors in Computing 
Systems, Boston, MA: 24-28. 
Cohen, R. 1986 An Incremental Model for Discourse Analysis. 
Unpublished draft, Department of Computer Science, University 
of Waterloo, Canada. 
Cohen, R. and Jones, M. 1988 Incorporating User Models into Expert 
Systems for Educational Diagnosis. In Kobsa, A. and Wahlster, 
W. (eds.), User Models in Dialog Systems. Springer-Verlag, 
Berlin--New York. 
Grosz, B. and Sidner, C. 1986 Attention, Intentions, and the Structure 
of Discourse. In Computational Linguistics 12: 175-204. 
Joshi, A.; Webber, B. and Weischedel, R. M. 1984 Living up to 
Expectations: Computing Expert Responses. In Proceedings of 
the National Conference on Artificial Intelligence, Stanford, CA: 
169-175. 
Kass, R, and Finin, T. (this issue): Modeling the User in Natural 
Language Systems. 
Paris, C. L. 1985 Description Strategies for Naive and Expert Users. 
In Proceedings of the 23rd Annual Meeting of the Association for 
Computational Linguistics, Chicago, IL: 238-246. 
van Beek, P. 1986 A Model for User Specific Explanation from Expert 
Systems. M. Math. thesis, Technical Report CS-86-42, Depart- 
ment of Computer Science, University of Waterloo, Canada. 
van Beek, P. and Cohen, R. 1986 Towards User Specific Explanation 
Systems. In Proceedings of the 6th Canadian Conference on 
Artificial Intelligence, Montreal, Canada: 194-198. 
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. 
