USER MODELS~ DISCOURSE MODELS, AND SOME OTHERS 
Karen Sparck Jones 
Computer Laboratory 
University of Cambridge 
Cambridge, England 
My context for this note is natural language interfaces to 
complex systems, considered, as the aim is system 
building, from the system's point of view. I hold that the 
discourse model relates four other models that are all 
required in dialog, but which need to be distinguished. 
These other models, as I define them in this note, are 
the world model, the user model, the system model, and 
the text model. It is a mistake, in my view, to try to 
establish the relation between discourse models and 
user models without reference to the other models that 
figure in the interaction between man and machine. 
The case for all the models mentioned can be made 
through a hypothetical system example. Suppose we 
have an automated reference librarian in an industrial 
company, i.e., a highly sophisticated expert system. 
Suppose also that the system can sustain the kind of 
dialog we can, and so imagine a dialog between the 
system and a user as follows: 
U1. I'm looking for a book on logic programming. 
S1. You're in Accounts, aren't you? 
U2. Yes. 
$2. So you'd like something with financial examples. 
U3. Well that would be very convenient, but I'm 
really interested in logic programming itself. 
$3. Just as well: we don't have anything on Prolog 
for accountants. Do you want something really 
noddy? We've got one pretty basic book. 
U4. How good do you think it is? 
$4. It's very popular. It doesn't assume you know 
anything about programming. 
U5. Oh, but I've done a bit of Cobol. 
$5. There isn't anything on Prolog from a Cobol 
point of view, I'm afraid. 
U6. I wouldn't have thought that mattered: Cobol's 
a good general lead in, isn't it? 
$6. Not especially. But we have another book that 
would probably suit you better than the first: it's very 
clear and has a lot of different examples like planning 
the week's menus and building an office meeting 
diary system, but it does assume you have some 
experience of programming. 
I assume that the system knows about the material in 
the library, about the company, and about the library's 
role in the company's activities. It also knows about the 
company's employees (which include the illustrative 
user, known to be a woman in Accounts who thinks her 
boss underestimates her). The system's knowledge of 
the library and the company are part of its permanent 
world knowledge, and its knowledge of the company's 
employees is in principle no different in kind, though the 
system's knowledge of people's intensional states may 
be less certain than some of its other knowledge. 
There is nevertheless a good reason for distinguish- 
ing the system's knowledge of its users, and hence of 
any specific user, from the rest of its world knowledge. 
This is a functional reason: the system exists to solve 
problems presented by its users. The individual user is 
a necessary participant in the system's problem solving, 
and the system's actions are driven by its view of the 
user's specific needs. In the dialog, the system's re- 
sponse $6 is motivated by the perceived fact that the 
user, as a user, is a person with particular characteris- 
tics who is separated out from the rest of the world 
because her needs have to be met. Thus the response is 
an appropriate one in relation not to individual utter- 
ances or facts about the user, but to the entity consti- 
tuting the user model as a whole, namely that she's a 
non-novice female, seeking to impress her boss by 
improving her computing skills. (This functional view, 
of course, implies that whether the user is human or not 
is an independent, contingent matter.) 
But for the same functional reason, the system has to 
have a model of itself embodying, for example, its plan 
to extract more detail about the user's book request. 
Though the system is in principle, like the user, part of 
the world, it has to be functionally distinguished for it to 
carry out its task: thus it is the system's aim, not the 
general state of the world, which leads it to choose $2, 
asking about financial examples. 
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 
98 Computational Linguistics, Volume 14, Number 3, September 1988 
Karen Sparck Jones User Models, Discourse Models, and Some Others 
However, there are no good grounds for assuming 
that the system's immediately accessible world model 
for a given interactive session contains all its world 
knowledge, its immediate user model everything it 
knows about the user, or its immediately accessible 
system model everything it knows about itself: for 
example, that its models for the illustrative dialog cover 
its knowledge of the company's buildings, the user's 
pay, or its own plan to revise its book descriptions (it's 
a powerful system). The system's entire stock of knowl- 
edge is not needed for effective interaction in a partic- 
ular dialog: and indeed, given the large amounts of 
permanent knowledge presupposed, having all this to 
hand would simply clog everything up. 
For a given dialog, therefore, the world model, user 
model, and system model will be selections, which are 
functionally motivated and hence distinguished, from 
the system's complete and previously undifferentiated 
stock of knowledge. This, of course, requires an invo- 
cation mechanism of the kind sought by Sperber and 
Wilson (1986), yet one which not merely selects infor- 
mation as relevant but assigns it a functional role. For 
example, being in Accounts becomes part of the user 
model, but knowledge about the relation between Pro- 
log and logic programming becomes part of the world 
model, and proffering specific books as a strategy for 
clarifying user needs becomes part of the system model. 
The invocation is continuous, triggered by the progress 
of the dialog, so the system's knowledge of the user's 
desire to impress her boss, for example, which was not 
necessarily initially selected as relevant, is invoked to 
motivate $6. 
The fourth factor is the dialog itself. The text of the 
interaction is an inert object: but discourse processing 
implies some model of the text as a linguistic object 
(Sparck Jones 1983). This linguistic model is a text 
model that is functionally required, like the other mod- 
els, to support one subprocess of the whole problem- 
solving interaction. The text model deals with text- 
based entities, which are linguistically characterized 
and need not have real world referents, like the Cobol 
book (cf. Panel VII in Tinlap-3 1987). It is this linguis- 
tically motivated interpretation of the dialog text which 
embodies information, e.g., about the lexical items 
used, topicalisation, and anaphoric links. The text 
model is required to guide further discourse production, 
if not determining at least suggesting the choice of a 
word, constraining sentential structure to maintain co- 
hesion, etc., as in the use of "accountants" rather than 
"financial staff" in $3, for example and of "that" in U6, 
and in the form of $5, where the structure of "Prolog 
from a Cobol point of view" is linguistically related to 
$3 and U5. 
Compared with the other models, the text model may 
not appear to be a subset of a larger body of knowledge 
about the world, including knowledge about individuals. 
But it can apply general knowledge of argument struc- 
tures (in the sense of Reichman (1985)) as well as of 
grammar, just as the permanent world model contains 
general knowledge about humans, and it can in principle 
also depend on prior knowledge of linguistic individu- 
als, e.g., (somewhat trivially) "Accounts". Knowledge 
provided by the text model, like that supplied by the 
other models as they develop through the interactive 
session, can update the permanent stock, in which it has 
no more special status or character than the information 
about people who can function on occasion as users. 
But the text model is not the discourse model: it is 
too shallow for this. It does not, for instance, express 
the relation between Prolog and logic programming, or 
the user's inferred intention to better herself, though 
these are clearly discourse matters. The text model 
includes, on the other hand, information about the order 
of mention of entities that is relevant to the production 
of linguistic responses, which is not obviously a dis- 
course matter if the discourse model is about the 
substantive relations between discourse entities (Co- 
hen, Morik, Schuster, all this issue). 
In fact, none of the four models listed can claim to be 
a discourse model. The discourse model is what relates 
these four models, that is, expresses the relations 
between them. Thus the discourse model relates the 
world model entity, the noddy book, with the user 
model element representing the user's request for a 
book on logic programming; and it relates the system 
model constituent--help a user by suggesting a specific 
book--to these world model and user model entities, 
the noddy book and requested book, respectively. 
Again, the discourse model relates the system's belief 
about the utility of a book on Prolog for a would-be logic 
programmer with the world model link between Prolog 
and logic programming, and it relates this book in the 
world model with the text model entity for "one pretty 
basic book". 
The discourse entities are the entities involved in the 
various models (taking entities as complex structured 
objects as well as simple ones), but the discourse model 
is not the mere aggregate of the other four models: it has 
to explicitly relate their elements in a way that allows 
transitions from one model to another to meet the 
requirements of the system's task. Thus the user's 
belief about the connection between Cobol and logic 
programming is an element of the user model, where it 
is functionally associated (by the system) with user's 
goal of learning about logic programming. But the user's 
belief about the relation between Cobol and logic pro- 
gramming is also related to the world model because it 
is a function of the world model to test for existence, in 
this case for the reality of a connection between Cobol 
and logic programming. The relation between these user 
and world model entities referring to Cobol and logic 
programming, i.e., the relation between a belief in a 
user model functionally concerned with achieving goals 
and a proposition in a world model functionally con- 
cerned with existence testing, is one sort of discourse 
model relation. There is another one in the relation 
Computational Linguistics, Volume 14, Number 3, September 1988 99 
Karen Sparek Jones User Models, Discourse Models, and Some Others 
between the user model belief and the text model entity 
"a good general lead in". But what discourse model 
relations we should recognize, and hence whether' we 
should have an austere or promiscuous ontology for 
these relations, is not so far clear. 
Thus the four different modelslworld, user, system, 
and text--below the discourse model are all needed 
because each serves a distinct function, implying a 
different selection from, and organization of, the total 
information relevant to the interaction between the user 
and the system. The discourse model is then also 
needed to provide the links between the four, which 
support calls from one to another. 
A shifting focus of attention like that represented by 
the point of interaction between participants and dis- 
course in Grosz and Sidner's (1986) account is naturally 
presupposed here. But my argument is that we need to 
separate a participant's (in this case, the system's) view 
of itself from its views of the world and of the user. The 
way these interact with the text model will then be 
reflected in a subset of relations (and hence entities) in 
the discourse model which constitutes the focus of 
attention. Thus, within the type of framework Grosz 
and Sidner propose, I am advocating, on functional 
grounds, a finer allocation of entities to a larger set of 
model,~. This large set of functionally motivated models 
is neede6 for serious systems and interfaces, though in 
simple cases the different models may collapse into one. 
I am similarly arguing, in relation to Wahlster (1987), for 
a finer set of modeling distinctions. These distinctions 
reflect the different status of the various objects and 
perspectives involved in discourse, and have to be 
recognized even if the system's operations are treated 
as if carried out within one overall model. 

REFERENCES 
Grosz, B. and Sidner, C. 1986 Attention, Intentions, and the Structure 
of Discourse. In Computational Linguistics 12: 175-204. 
Reichman, R. 1985 Getting Computers to Talk Like You and Me. MIT 
Press, Cambridge, MA. 
Sparck Jones, K. 1983 Shifting Meaning Representations. In Proceed- 
ings of the 8th International Joint Conference on Artificial Intelli- 
gence: 621-623. 
Sperber, D. and Wilson, D. 1986 Relevance. Basil Blackwell. Oxford, 
England. 
Tinlap-3, Panel VII on Reference 1987 In Wilks, Y. (ed.), Theoretical 
Issues in Natural Language Processing, New Mexico State Uni- 
versity, University Park, NM: 128--154. 
