TAILORING OBJECT DESCRIPTIONS TO A USER'S LEVEL OF EXPERTISE 
C~cile L. Paris 
Department of Computer Science 
Columbia University 
New York, NY 10027 
A question answering program providing access to a large amount of data will be most useful if it can 
tailor its answers to each individual user. In particular, a user's level of knowledge about the domain of 
discourse is an important factor in this tailoring if the answer provided is to be both informative and 
understandable to the user. In this research, we address the issue of how the user's domain knowledge 
can affect an answer. By studying texts, we found that the user's level of domain knowledge affected the 
kind of information provided and not just the amount of information, as was previously assumed. 
Depending on the user's assumed domain knowledge, a description can be either parts-oriented or 
process-oriented. Thus the user's level of expertise in a domain can guide a system in choosing the 
appropriate facts from the knowledge base to include in an answer. We propose two distinct descriptive 
strategies that can be used in a question answering program, and show how they can be mixed to include 
the appropriate information from the knowledge base, given the user's domain knowledge. We have 
implemented these strategies in TAILOR, a computer system that generates descriptions of devices. 
TAILOR uses one of the two discourse strategies identified in texts to construct a description for either 
a novice or an expert. It can merge the strategies automatically to produce a wide range of different 
descriptions to users who fall between the extremes of novice or expert, without requiring an a priori set 
of user stereotypes. 
1 INTRODUCTION 
The tailoring of answers according to a person' s domain 
knowledge frequently occurs in human communication. 
For example, an explanation aimed at a child of how a 
car engine works will be different than one aimed at an 
adult, and an explanation adequate for a music student 
is probably too superficial for one in mechanical engi- 
neering. We have found further evidence of this phe- 
nomenon in naturally occurring texts (Collier 1962, 
Britannica 1984, Britannica-Junior 1963, New Book of 
Knowledge 1967, Encyclopedia of Science 1982, Chev- 
rolet 1978, Weissler 1973), where the information pre- 
sented to readers varies with their assumed level of 
general knowledge. 
Likewise, a question answering program that pro- 
vides access to a large amount of data to many different 
users could be more effective if it could customize its 
answers to each user, retrieving from its knowledge 
base the facts that are most appropriate and useful for a 
given user. Much research to date has focused on 
tailoring an answer depending on a user's goals, but 
customizing an answer depending on what the user 
knows about the domain of the question has been 
overlooked. This is an important factor in tailoring an 
answer if the answer provided is to be both informative 
and understandable to the user. The answer should not 
provide information that is obvious to the user (Grice 
1975). However, if the answer assumes knowledge that 
the user does not have, it may be very hard (if not 
impossible) for the user to understand the answer (Wilson 
and Anderson 1986). In this paper we show the feasibility 
of incorporating the user's domain knowledge, or level of 
expertise, into a generation system and address the issue 
of how this factor might affect an answer. 
Through an analysis of texts, we found that two 
distinct discourse strategies were used in describing 
texts. We postulate that the writers' choice of strategy 
might be based on the assumed domain knowledge of 
the expected readers. If so, then the reader's level of 
knowledge about the domain affects the kind of infor- 
mation provided as opposed to just the amount of 
information. In previous generation systems, the user's 
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 
64 Computational Linguistics, Volume 14, Number 3, September 1988 
C~ile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
domain knowledge affected only the amount of detail 
provided in the text (Wallis and Shortliffe 1982). 
In this work, we show how the two discourse strat- 
egies found in texts can be used to provide answers to 
users whose domain knowledge falls anywhere along a 
knowledge spectrum, from naive to expert. We have 
implemented them in TAILOR, a computer system that 
generates descriptions of physical objects. 
TAILOR uses one of the two discourse strategies 
identified in texts to construct a description for either a 
novice or an expert. It can merge the strategies auto- 
matically in a systematic way to produce a wide variety 
of different descriptions for users who fall between the 
extremes of novice and expert. This means TAILOR is 
able to generate descriptions to a whole range of users, 
rather that just for an a priori set of user stereotypes. 
1.1 PREVIOUS WORK ON USER MODELING IN QUESTION 
ANSWERING PROGRAMS 
In studying the factors involved in tailoring the content of 
an answer to a user, research to date has focused mainly 
on the problems of inferring and using user goals, plans, 
and beliefs (Appelt 1982, 1985; Carberry 1983, and this 
issue; McKeown et al. 1985), recognizing and dealing with 
misconceptions (Kaplan 1982; McCoy 1983; McCoy 1986, 
and this issue; Quilici et al., this issue), and superposing 
various stereotypes (Rich 1979). The issue addressed here 
differs from these because we are not concerned with the 
users' goals in asking the question, nor with correcting 
their view of the domain, but rather with providing an 
answer that is optimally informative (without being over- 
whelming) given how much the user knows about the 
domain. We are not interested in building a user model 
using stereotypes (as was Rich), but in determining an 
answer based on a user model involving user types. As in 
McCoy (1986, and this issue), we are more concerned 
about using information from the user model to generate 
an answer than building the model itself. 
While the need for a model of the user's domain 
knowledge in question answering systems has been 
noted by various researchers (Lehnert 1977; McKeown 
1985), few programs have actually had one. The HAM- 
ANS system (Hoeppner et al. 1984) has a model of the 
user's knowledge, but this knowledge is mainly used for 
anaphora resolution and production. In our work, we 
are more interested in studying how a user's knowledge 
affects the content of an answer as opposed to its 
phrasing. Wallis and Shortliffe (1982), who have used 
the naive/expert distinction in providing an answer (or 
explanation), did so mainly by giving more or less 
detail, without addressing the issue of whether the level 
of detail was the only important factor to vary. The 
issue we confront in this work is identifying the role 
played by a user's level of knowledge in determining the 
content of an answer. The UNIX Consultant (UC) (Chin 
1986) uses a user's knowledge level about the UNIX 
system to provide help to its users. UC, however, uses 
stereotypes for both the user and the knowledge base 
(set of UNIX commands). Stereotypes for the knowl- 
edge base include "simple", "mundane", and 
"complex". UC matches the user type against the 
command type to decide on the answer. In this work, 
we are looking at a different kind of domain, the domain 
of complex physical objects, in which this categoriza- 
tion of the knowledge base is not possible. Further- 
more, we would like to be able to tailor answers to users 
whose domain knowledge level falls anywhere along a 
knowledge spectrum without necessarily having to clas- 
sify users in several different stereotypes. 
1.2 THE DOMAIN 
In our work, we are mainly concerned with describing 
complex devices such as telescopes, telephones, and 
disk drives to users with different levels of expertise. 
Our choice of domain has been motivated by RE- 
SEARCHER, a program being developed at Columbia 
University. RESEARCHER reads, remembers, and 
generalizes from patent abstracts written in English 
(Lebowitz 1983, 1985, 1986). The resulting knowledge 
base is organized in a generalization hierarchy. The 
abstracts describe complex physical objects in which 
spatial and functional relations are important. In this 
domain, the amount of information contained in the 
knowledge base is very large and the information can be 
very detailed. Moreover, the knowledge base contains 
several different kinds of information: spatial, func- 
tional, and attributive (properties attached to objects). 
A program can choose from among facts representing 
different kinds of information about an object, and facts 
at different levels of detail in the knowledge base, 
rendering the decision process a complicated one. 
A request for the description of an object cannot be 
translated into a simple database query and thus cannot 
be answered by a straightforward retrieval from the 
knowledge base. This type of question has been termed 
high level questions. (Tennant 1978, McKeown 1985). 
There are no clear constraints on what information 
should be included in the answer. Since the amount of 
information contained in the knowledge base is very 
large and the information very detailed, a program 
cannot just state all the facts contained in the knowledge 
base about the object as there will typically be too 
many. Rather, it needs to select a subset of facts to 
present to the user. As the answer will be composed of 
several facts, a generation program needs to organize 
these facts in order to construct from them a coherent 
text (McKeown 1985). When a generation system can 
choose among many facts, a user model representing 
what the user presumably knows about the domain can 
guide the system in choosing information that the user 
understands and does not already know (and cannot 
easily infer), thereby improving the resulting answer. 
Descriptions are important because they can be used 
to answer other types of high level questions. For 
example, to compare two objects, it may be necessary 
to describe each of them. Furthermore, with a knowl- 
Computational Linguistics, Volume 14, Number 3, September 1988 65 
C~cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
edge base of physical objects, users are likely to ask 
such questions. In order to focus on how the level of 
expertise affects a description, we have not considered 
how the goal of the questioner could affect the descrip- 
tion. It is clear that in a sophisticated question answer- 
ing program the user's goal should also play an impor- 
tant part. An answer for a user whose goal is to buy an 
object should include different kinds of information than 
an answer for a user who wants to repair this object. 
Detecting and using the user's goal to provide an 
appropriate response has been the focus of extensive 
research (Appelt 1985, Carberry 1983, McKeown et al. 
1985). In this work, being more concerned with the role 
played by the user's domain knowledge, we simply as- 
sume that users want to find some information about an 
object. A description should provide meaningful informa- 
tion about an object and allow the user to build a mental 
functional model of the object. Therefore, we assume that 
the goal of a description is to help the user construct a 
mental functional model of the object under consideration. 
2 IDENTIFYING WHAT NEEDS TO BE IN THE USER 
MODEL 
Our goal is to provide a characterization of the role of 
the user's domain knowledge in generating descriptions 
that is computationally usable by a generation system. 
Even though we will not be addressing the problem of 
how to determine how much the user knows about the 
domain, we still have to ask what kinds of knowledge a 
user can possess about a domain that can affect gener- 
ation and that can be explicitly represented in a user 
model. Instances of these kinds of knowledge Will be the 
information contained in our user model. Having iden- 
tified what needs to be in the user model, we will take 
the user model as given, and study how a system can 
use the information about a user's domain knowledge 
contained in the user model to tailor the answer. 
Analysis of natural language texts suggests the exist- 
ence of at least two kinds of domain knowledge that 
affect the type of descriptions that can be provided in 
our domain: 
• knowledge about specific items in the knowledge 
base. We define "knowing" about an object to mean 
knowing about the existence of the object, its purpose 
and how this purpose is achieved (that is, how the 
various subparts of the object work together to 
achieve it). Knowing about an object thus means 
understanding the functionality of the object and the 
mechanical processes associated with it. 
• knowledge about various basic underlying concepts. 
In a domain of complex physical objects, such con- 
cepts might include electricity and voltages. 
We define an expert user as one whose knowledge about 
the domain includes functionality of objects and mechan- 
ical processes. An expert user knows all the underlying 
basic concepts and the majority of the generalized objects 
contained in the knowledge base for a particular domain. 
Given an object that is new but similar to a known one, 
such a user has enough domain knowledge to infer how 
the palls of this new object work together to perform a 
function. A naive user is one who does not know about 
specific objects in the knowledge base, and does not 
necessarily understand the underlying basic concepts. 
A user is not necessarily naive or expert, however. For 
example, a user may. know about several objects in the 
knowledge base. In this work, instead of rating the user as 
having some intermediate level of domain knowledge, we 
use explicit parameters to indicate the user's knowledge. 
These parameters are a list of items in the knowledge base 
which are known to the user and information about 
whether the user understands the underlying basic con- 
cepts. So, for example, a user model in our system may 
contain a parameter indicating that the user only has local 
expertise with respect to disk drives. We retain the 
terminology naive and expert only for users at the two 
ends of the knowledge spectrum. Thus our emphasis is on 
studying how object descriptions can be varied when 
these parameters vary. We will not try to categorize a 
user's domain knowledge, or attempt to determine what 
levels exist between the two extremes. 
3 Two DESCRIPTIONS STRATEGIES FOUND IN TEXTS: 
CONSTITUENCY SCHEMA AND PROCESS TRACE 
3.1 THE TEXTUAL ANALYSIS 
To develop effective strategies for tailoring a description 
to a particular level of expertise, we began by studying 
descriptions in a variety of texts: adult encyclopedias 
(Britannica 1984, Collier 1962) and junior encyclopedias 
(Britannica-Junior 1963, New Book of Knowledge 1967, 
Encyclopedia of Science 1982), manuals (Chevrolet 1978, 
Weissler 1973), and high school textbooks. This range of 
texts was chosen because it provided a good source of 
descriptions ~ and because these texts seem to address 
audiences at the two ends of the knowledge spectrum: 
naive and expert. Texts from adult encyclopedias are 
directed at an audience much more knowledgeable in 
general than the audience addressed by high school text- 
books and junior encyclopedias. Likewise, the Chevrolet 
manual is aimed at knowledgeable users (i.e., professional 
mechanics), while the other manual claims to be directed 
towards novices. We studied descriptions of devices, 
taking the description of the same object in all sources 
whenever possible. The descriptions we have studied are 
generally several paragraphs in length. 
Besides providing us with examples of descriptions, 
encyclopedias have the added advantage (for our study) 
that people read them for a variety of reasons. Yet, they 
all obtain the same texts (and therefore the same 
information). An encyclopedia is thus providing its 
readers with information about an object without taking 
the reader's goals into account. 
We analyzed the different texts using methods devel- 
66 Computational Linguistics, Volume 14, Number 3, September 1988 
C~eile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
{Identification (description of an object in terms of its 
superordinate)} 2 
Attributive* (associating properties with an entity)/ 
Cause-effect* 
Constituency (description of subparts or subtypes) 
{Depth-identification / Depth-attributive 
{Particular Illustration / Evidence} 
{Comparison ; Analogy} }+ 
{Attributive / Explanation/Analogy} 
Figure 1. The Constituency Schema. 
oped by other researchers (Hobbs 1978a, 1980; Mann 
1984, McKeown 1985), decomposing paragraphs in terms 
of their primitive rhetorical structure in an attempt to find 
consistent structures in the texts. We found that the texts 
fell into two groups: most of the descriptions in the adult 
encyclopedia entries and in the car manual for mechanics 
were organized around object subparts and their proper- 
ties, while the descriptions in the junior entries and in the 
car manual for novices traced process information. 
3.2 TEXTS ORGANIZED AROUND SUBPARTS 
The texts from the adult encyclopedias and the Chev- 
rolet manual can be characterized in terms of the 
discourse strategy the constituency schema, one of the 
textual structures posited in McKeown (1985). In her 
work on natural language generation, McKeown studied 
the problems of what to say when there are many facts 
to choose from and how to organize a text coherently. 
To this end, she examined texts and transcripts in order 
to determine whether there were standard patterns of 
discourse structure used in naturally occurring texts. 
Where patterns of discourse structure could be identi- 
fied for various discourse goals, these patterns could be 
used to guide a generation system in choosing and 
organizing facts to construct a text. 
McKeown analyzed the texts by classifying each 
sentence as one of a set of rhetorical predicates. Rhe- 
torical predicates characterize the structural purpose of 
sentences and have been discussed by a variety of 
linguists (Shepherd 1926, Grimes 1975, Hobbs 1978b). 
Some examples are constituency (description of sub- 
parts or subtypes), attributive (providing detail about an 
entity or event), and analogy. McKeown found that in 
the texts she studied, some combinations of predicates 
were more likely to occur than others, and that for each 
discourse situation (such as providing a definition), 
some combination was the most frequent. McKeown 
encoded these standard combinations as schemas that 
are associated with a particular discourse situation. One 
of these schemas is the constituency schema, which is 
used to describe an object (or concept) in terms of its 
subparts and their properties. The constituency schema 
is shown in Figure 1. 
The descriptions from the adult encyclopedias and 
the car manual for mechanics followed the pattern 
The following text illustrates the decomposition of a 
description using the Constituency Schema. 
1) The hand-sets introduced in 1947 2) consist of a 
receiver and a transmitter in a single housing avail- 
able in black or colored plastic. 
3) The transmitter diaphragm is clamped rigidly at its 
edges 4) to improve the high frequency response. 5) 
The diaphragm is coupled to a doubly resonant 
system 6) -a cavity and an air chamber- 7) which 
broadens the response. 8) The carbon chamber con- 
tains carbon granules, 9) the contact resistance of 
which is varied by the diaphragm's vibration. 
10) The receiver includes a ring-shaped magnet sys- 
tem around a coil and a ring shaped armature of 
anadium Permendur. 11) Current in the coil makes 
the armature vibrate in the air gap. 12) An attached 
phenolic-impregnated fabric diaphragm, shaped like 
a dome, 13) vibrates and sets the air in the canal of 
the ear in motion. 
Using rhetorical predicates, we can classify the sen- 
tences of the above description in the following way: 
I. Attributive 
2. Constituency 
Depth-attributive for the Depth-attributive for the 
transmitter (Description of the receiver (Description of the 
transmitter) receiver) 
3. Depth-Attributive 10. Depth-Attributive 
4. Cause-effect 11. Cause-effect 
5. Depth-Attributive 12. Attributive 
6. Depth-identification 13. Cause-effect 
7. Cause-effect 
8. Depth-Attributive 
9. Cause-effect 
Figure 2. Constituency Schema Example. 
shown in this schema: the descriptions are organized 
around the parts of the object. An example of such a 
description is shown in Figure 23 . This entry, describing 
the telephone, is taken from Collier (1962). In the first 
paragraph, the parts (constituents) of the telephone are 
given. Then, each main part is described in turn: first 
the transmitter, then the receiver. In the descriptions 
from this set of texts, the parts are also described with 
their subparts and their properties (depth-attributive). 
3.3 TEXTS FROM JUNIOR ENCYCLOPEDIAS AND FROM THE 
CAR MANUAL FOR NOVICES 
The texts from junior encyclopedias, high school text- 
books, and the car manual for novices are organized in 
a significantly different manner. No known schema or 
other organizing structure consistently accounted for 
the descriptions in the junior encyclopedia texts. In 
looking for other types of organizing strategies, we 
discovered that the main strategy used in these descrip- 
tions is to trace through the process that allows the 
object to perform its function, that is, to mainly describe 
Computational Linguistics, Volume 14, Number 3, September 1988 67 
C~cile L. Paris Tniloring Object Descriptions to a User's Level of Expertise 
I. 1) When one speaks into the transmitter of a modern 
telephone, these sound waves strike against an alumi- 
nium disk or diaphragm and cause it to vibrate back and 
forth in just the same way the molecules of air are 
vibrating. 
II. 2) The center of this diaphragm is connected with 
the carbon button originally invented by Thomas A. 
Edison. 3) This is a little brass box filled with granules 
of carbon composed of especially selected and treated 
coal. 4) The front and back of the button are 
insulated. 
III. 5) The talking current is passed through this box so 
that the electricity must find its way from granule to 
granule inside the box. 6) When the diaphragm moves 
inward under the pressure from the sound waves the 
carbon grains are pushed together and the electricity 
finds an easier path. 7) Thus a strong current flows 
through the line. 8) When a thin portion of the sound 
wave comes along, the diaphragm springs back, allow- 
ing the carbon particles to be more loosely packed, and 
consequently less current can find its way through. 9) 
So a varying or undulating current is sent over the line 
whose vibrations exactly correspond to the vibrations 
caused by the speaker's voice. 10) This current then 
flows through the line to the coils of an electromagnet in 
the receiver. 
IV. 11) Very near to the poles of this magnet is a thin 
iron disc. 
V. 12) When the current becomes stronger it pulls the 
disc toward it. 13) As a weaker current flows through 
the magnet, it is not strong enough to attract the disk 
and it springs back. 14) Thus the diaphragm in the 
receiver is made to vibrate in and out... 
Figure 3. Description from a junior entry. 
processes associated with the operation of the object. 
We characterize these descriptions as process descrip- 
tions. An example of such a description, from Britanni- 
ca-Junior (1963), is presented in Figure 34. 
We see that the organizing principle of this text is the 
mechanical process description. The process descrip- 
tion gets interrupted when descriptive information can 
be included concerning a subpart that was just men- 
tioned as part of the process description. (Such infor- 
mation is shown indented in the example). Further- 
more, in this text, not only is the description made 
mainly through a process trace, but this process trace is 
given in great detail and substeps are explained if there 
are any. 
The information contained in this group of descrip- 
tions corresponds to the causal links that connect the 
various processes contained in the knowledge base. To 
generate such a description, it is then necessary to 
follow these links, giving rise to a process trace that 
(For each object, given a chain of causal links) 
(1) Follow the next causal link 
(2) {Mention an important side link} 
(3) {Give attributive information about a part just 
introduced} 
(4) {Follow the substeps if there are any. (These sub- 
steps can be omitted for brevity.)} 
(5) Go back to (1). 
(This process can be repeated for each subpart of the 
object.) 
Figure 4. The Process Trace. 
describes how the object functions. The algorithm used 
for the process trace, summarized in Figure 4, is as 
follows: given the chain of causal links that constitutes 
the functional information of the object, the first link is 
taken. If there is an important side effect at this point, it 
is mentioned. If a new part was introduced when the 
causal link was mentioned, attributive information 
about the part may be included. If the step just ex- 
plained can be subdivided into substeps, the trace may 
continue at the substeps level. Finally, the next causal 
link is taken and the algorithm repeats. (This strategy is 
described in detail in Paris and McKeown (1987) and 
Paris (1987).) 
Substeps happen, for example, in the following case: 
"... causes the diaphragm to vibrate". This step can 
be divided into the two substeps: "the diaphragm 
moves inward" and "the diaphragm moves outward". 
Substeps can also arise when a complex object is made 
of several other complex parts. The strategy first de- 
scribes how the parts work together to achieve the 
object's function. If a long description is desired, it is 
possible to step through each of the parts, describing 
how it achieves its own function. This is similar to the 
schema recursion for the constituency schema men- 
tioned by McKeown (1985). 
3.4 TAILORING DESCRIPTIONS 
Given that the two types of descriptions occur in texts, 
a system that generates device descriptions should also 
be able to provide the two kinds of descriptions. We 
have thus implemented both strategies in our generation 
system. 
We use the constituency schema when a user has 
expertise about the domain of discourse, giving rise to a 
parts-oriented description. Recall that we assumed that 
the goal of a description is to allow the user to form a 
mental model of the functionality of the object. Re- 
search in psychology indicates that expert users have 
more knowledge not only about individual components, 
but also about the causal models involved and the 
interconnections among parts (Lancaster and Kolodner 
1987; Chi et al. 1981). Expert users, then, are likely to 
have functional knowledge about the domain and to 
know how parts might interact with each other. As they 
68 Computational Linguistics, Volume 14, Number 3, September 1988 
C6cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
can use this knowledge when reading the description 5, 
they should be able to pull all the parts provided in the 
description together in order to "understand" the de- 
scription as a whole. That is, they should be able to 
figure out how the parts fit together to form an object 
capable of performing a function. Since the reader is 
able to construct a mental model, it is unnecessary to 
include the process information in the description. 
Actually, assuming that the reader will be able to infer 
the processes involved, providing such useless informa- 
tion would contradict the principles of cooperative 
behavior (Grice 1975). 
On the other hand, a user who does not have enough 
knowledge to infer the processes linking the parts would 
be unlikely to understand a mostly structural descrip- 
tion of an object and to be able to construct a functional 
mental model of the object from such a description. For 
this sort of naive user, if the description is to be 
informative and understandable, it must describe how 
the parts perform the function of the object. The de- 
scription must therefore include process information. 
Previous research in reading comprehension strength- 
ens our belief that a user who does not have knowledge 
about the functions of the various parts will not be able 
to make sense of a description centered around parts. 
Wilson and Anderson (1986) demonstrate the impor- 
tance of prior knowledge in comprehending new texts, 
and show how readers can fail to understand a text 
mainly because the text contains implicit knowledge 
that the readers do not have. We thus propose to use the 
process trace when providing a description to a naive 
user. 
To summarize, we suggest that the user's domain 
knowledge affects the content of a description with 
respect to the kind of information included, and not just 
to the level of detail, and postulate that the choice of 
strategy might be based on the assumed level of exper- 
tise of reader/user. Namely, the process trace can be 
used when the expected readers are relatively naive 
about the domain of discourse, while the constituency 
schema can be used when the expected readers have 
expertise about the domain. We will show how a user's 
level of expertise can be incorporated in a generation 
system and how it can guide the system in choosing a 
discourse strategy. 
4 MLXlNG THE STRATEGIES 
The two strategies presented account for the main 
differences found between the adult and junior encyclo- 
pedia entries and we proposed to use them to describe 
objects to naive or expert users. Users are not neces- 
sarily either naive or expert in a domain however. They 
may have local expertise, knowing about some objects 
in the domain and not others (Paris 1984). Such users 
would not be considered naive users, but, as there are 
many objects they do not know in the domain, they 
would not be considered expert users either. The user 
models for such users would indicate for which objects 
of the knowledge base they have local expertise. We 
believe that, to describe objects to users with interme- 
diate levels of expertise, a combination of the two 
strategies presented for naive and expert users is appro- 
priate. Based on the user model, a generation program 
can decide which strategy to use for which object. 
As an example, suppose we are providing a descrip- 
tion of an elevator to a user who knows the function of 
a motor and how this function is achieved, but not how 
the elevator itself works. In describing the elevator to 
this user it is necessary to first describe how the parts of 
the elevator work together, using the process trace 
strategy, since the user model would indicate local 
expertise about the "motor" only. This local expertise 
is too narrow to allow the user to understand how all the 
parts of the elevator work together to perform their 
required function. A process explanation is thus neces- 
sary. In describing the individual parts in turn, it should 
not be necessary to fully explain what the motor does, 
as the user already knows about it. The constituency 
schema strategy can thus be used to describe the motor. 
For the other parts, however, the process strategy is 
still appropriate in order to explain their mechanisms. 
Such combinations were actually also found in natu- 
rally occurring texts. Figure 5 presents an example of a 
text that uses a combination of the constituency schema 
and the process trace to generate a description aimed at 
users with intermediate levels of domain knowledge. 
This text is taken from the Encyclopedia of Chemical 
Technology (Chemical 1978). The description starts 
with the constituency schema strategy but ends with a 
process trace: the "IR (Infra-Red) spectrometer" is first 
described in terms of its parts; each part is then de- 
scribed in turn (depth-attributive); finally, the authors 
revert to a process trace to describe the "thermocouple 
detector", assuming it is unknown to the reader. To 
fully understand this text, the reader must already know 
(or be able to infer information) about the IR spectrom- 
eter's purpose, the "IR radiation" and the "monochro- 
mator". In Figure 5, the text corresponding to the 
process trace is shown in italics. 
4,1 DECISION POINTS WITHIN THE STRATEGIES 
Since we want to be able to combine the two strategies 
to generate a description aimed at users with interme- 
diate levels of expertise, we must specify under which 
conditions one strategy would be preferable to the other 
and how to switch from one strategy to the other. The 
decision of which strategy to use at any point is based 
on information about the user's domain knowledge 
contained in the user model. Notice that it would be 
hard without a thorough psychological study to specify 
the exact conditions necessary to choose one strategy 
over the other and to switch from one strategy to the 
other. However, we have identified some heuristics that 
determine when to mix the strategies. (Should data from 
psychological experiments later become available, we 
Computational Linguistics, Volume 14, Number 3, September 1988 69 
C~,cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
(I) The IR spectrometer consists of three essential 
features: a source of IR radiation, a monochromator and 
a detector. (2) The primary sources of IR radiation are 
the Globar and Nernst glower. (3) The Globar is a 
silicon carbide rod heated to 1200 degrees C. (4)The 
Nernst glower is a rod containing a mixture of yttrium, 
zirconium, and erbium oxides that is heated electrically 
to 1500 degrees C. (5) Earlier IR spectrometers con- 
tained prism monochromators but today gratings are 
used almost exclusively. (6) Most detectors in modern 
spectrometers operate on the thermocouple principle. 
(7) Two dissimilar metal wires are connected to form a 
junction. (8) Incident radiation causes a temperature 
rise at the junction and the difference in the tempera- 
ture between head and tail causes a flow of current in 
the wires which is proportional to the intensity of the 
radiation. 
Text Decomposition 
1. Constituency 
2-4. Depth identification for the IR radiation 
5,6. Depth identification for the monochromators 
7. Depth identification for the thermocouple 
spectrometer 
8. Process trace for the thermocouple principle. 
Figure 5. Text from the Encyclopedia of Chemical 
Technology. 
Constituency Schema (with decision points) 
Identification (introduction of the superordinate) 
If there is no local expertise for the superordinate 
do a Process Trace (for the superordinate) before 
proceeding. 
Constituency (description of the subparts) 
For each part, do: 
If there is local expertise on this part (or its 
superordinate), do Depth-identification 
Else do a Process Trace (for the part) 
Attributive 
Figure 6. The Constituency Schema strategy and its decision 
points. 
after the identification predicate: once the superor- 
dinate of an object has been introduced, we could 
provide a process trace for this superordinate. 
After the constituency predicate: after mentioning 
the parts of an object, the constituency schema 
dictates to fill the depth identification predicate for 
each subpart. Instead, we could provide a func- 
tional description of one or more of the parts. This 
has been done in the Encyclopedia of Chemical 
Technology text presented in Figure 5, for 
example. 
could use the results as our heuristics to generate 
appropriate texts given the user's domain knowledge.) 
To decide on the strategy to use, the program looks in 
the user model to check whether the user is an "expert" 
or has a local expertise about the object to be described 
or about its superordinate in the generalization hierar- 
chy. In either case, the constituency schema is used; 
otherwise, the process trace is chosen. This process is 
repeated when the program has to decide on which 
strategy to employ to describe the subparts. 
If the user knows about most of the parts that play an 
important role in the mechanical process of the object 6, 
it is possible to describe the object with the constitu- 
ency schema instead of explicitly describing the process 
information that Connects these parts. Note however 
that, in this case, providing a process trace for the 
top-level description, thus indicating how the parts 
work together, might still provide an adequate descrip- 
tion. Currently, "most" is set to be at least half of the 
functionally important parts 7. 
To mix the strategies, we must specify when it might 
be possible to switch. Whenever an object is introduced 
and needs to be described, the system must decide 
whether to provide chiefly structural information (with 
the constituency schema) or functional information 
(with the process trace). This gives us some clear 
decision points in the strategies: 
• Within the constituency schema: 
• Within the process trace: 
• When a part is introduced while traversing the 
causal links, the process strategy dictates to include 
attributes of this part (to describe it). Here, we 
could also choose to describe the part more fully 
with the constituency schema. 
• When the subparts have to be described, the con- 
stituency schema can be used to provide structural 
information about them instead of including func- 
tional information. 
Figure 6 and 7 summarizes the two strategies in their 
simplest form with the decision points. 
Process Trace (with decision points) 
Next causal link 
Properties of a part mentioned during the process trace 
If a fuller description of the part is desired, do 
Constituency Schema (for the part) 
Substeps 
Back to next causal link 
Repeat for each of the subparts: 
If there is local expertise on this part (or its super- 
ordinate), do Constituency Schema 
Else do a Process Trace 
Figure 7. The Process Trace strategy and its decision 
points. 
70 Computational Linguistics, Volume 14, Number 3, September 1988 
C6cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
The decision to switch strategy is based on the user 
model: the program looks into the user model and 
checks the local expertise of the user. If there is no local 
expertise about the object, the program decides to 
follow the process trace; otherwise, the constituency 
schema is used. The user model is also examined to 
decide whether substeps should be traced: if the sub- 
steps involve any basic underlying concepts and the 
user model indicates that the user knows these con- 
cepts, the substeps are traced. Otherwise, the program 
does not trace through the substeps, as they might 
confuse the user. In that case, after the description has 
been generated, the user is asked whether he or she 
would like to see the substeps, although they might 
involve unknown concepts. 
Figure 8 shows an example of a text generated by 
TAILOR by combining the two strategies. More exam- 
ples can be found in Paris (1987).) Based on the user 
model that indicates local expertise about loudspeaker, 
TAILOR chooses the constituency schema. It first 
identifies the telephone by providing its purpose and 
then introduces its parts. Structural information is then 
provided about each of the parts, except for the trans- 
mitter, because the transmitter plays an important role 
in the function of a telephone and the user model shows 
no local expertise about it, or about the microphone, its 
superordinate in the generalization hierarchy. Thus 
TAILOR chooses to provide process information for 
the transmitter, switching momentarily to the process 
trace strategy. The process trace is shown underlined in 
the figure. 
5 TAILOR 
5.1 OVERVIEW OF THE SYSTEM 
Description of the telephone, given a user model 
User Model: Local Expertise: Loudspeaker (a 
pointer to the concept of a loudspeaker in the knowl- 
edge base); Basic concepts: electricity 
The telephone has two main functional parts: the trans- 
mitter (an instance of a microphone) and the receiver 
(an instance of a loudspeaker). Because the user knows 
one of the two parts of the telephone, the system 
decides on the constituency schema strategy at first. 
However, before providing structural information about 
each subpart, the system consults the user model and 
decides to switch strategy to describe the transmitter 
since the user has no local expertise about it. 
TAILOR output: 
The telephone is a device that transmits soundwaves. 
The telephone has a housing that has various shapes and 
various colors, a transmitter that changes soundwaves 
into current, a curly-shaped cord, a line, a receiver to 
change current into soundwaves and a dialing_me- 
chanism. The transmitter is a microphone with a small 
diaphragm. A person speaking into the microphone 
causes the soundwaves to hit the diaphragm of the 
microphone. The soundwaves hitting the diaphragm 
causes the diaphragm to vibrate. The vibration of the 
diaphragm causes the current to vary. The current 
varies, like the intensity varies. The receiver is a 
loudspeaker with a small aluminium diaphragm. The 
housing contains the transmitter and it contains the 
receiver. The housing is connected to the dialing__ 
mechanism by the cord. The line connects the dialing_. 
mechanism to the wall. 
Figure 8. A description generated by TAILOR combining 
the two strategies. 
We have implemented the discourse strategies pre- 
sented above in TAILOR, a program that generates 
descriptions tailored to a user's level of expertise. The 
discourse strategies guide the program to choose the 
appropriate information from the knowledge base, un- 
der the constraint of the user model. TAILOR generates 
descriptions aimed at users anywhere along the knowl- 
edge spectrum. TAILOR uses the knowledge base built 
by RESEARCHER, as depicted in Figure 9, and looks 
at the information contained in the user model to decide 
on the strategy to employ. After generating a descrip- 
tion, TAILOR updates the user model based on the 
objects that were included in the description provided to 
the user. At this point, no parsing of the questions is 
done (the input consists of a request for a description). 
In the ideal system, RESEARCHER would parse the 
question using the same parser as that used in reading 
patent abstracts, produce the request, and hand it to 
TAILOR. The question (along with other factors) could 
also be used to determine to the level of expertise. The 
user model includes the parameters that TAILOR uses 
to constrain its decision process during generation. The 
user model is not determined by the program but is 
given as input. 
RESEARCHER's knowledge base contains detailed 
descriptions of complex devices, including both struc- 
tural and functional information about the objects. We 
use a frame-based knowledge representation (Was- 
serman and Lebowitz 1983, Wasserman 1985) in which 
the basic frames represent objects. The knowledge base 
contains about 120 object frames and 150 frames of 
other types. Objects are organized in a generalization 
hierarchy. In addition to the generalization links, or 
instance-of links, there exist two additional kinds of 
links joining entities: part-of links, which indicate that 
an entity is a part in a larger structure, and relations, 
which convey information about spatial or functional 
relationships. Functional relations corresponds to the 
various events (or processes) that occur. Finally, there 
are links between relations, that is links between 
events. These links include cause-effect relations, tem- 
poral relations (such as "X happens at the same time as 
Computational Linguistics, Volume 14, Number 3, September 1988 71 
C6cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
(:=: ,,..+,,,,,,, ,,+.,°., o,,,,,. ) ( ,.,.+,..,. °-,*°".,°, o..°,,,**°.,:,') 
(organ£zod in a/ 
generalization | 
h£arazch¥).,, . J 
RIBSIhM~WI1 
(permel and 
generalizes) 
~ntern&~ £o~1 
USgR NODZL 
(Parameters r 4escr£bLng the l 
level of expertiselJ 
+l 
TAILOR 
Descr ~pttons 
Figure 9. RESEARCHER and the TAILOR System. 
Y"), and analogical relations (such as "X corresponds 
to Y"). 
A top-level diagram of TAILOR is shown in Figure 
I0. Input to TAILOR includes: I. a request for the 
definition of an object, and 2. the level of expertise of 
the user, that is, either one of the two stereotypes naive 
or expert, or, for users with intermediate levels of 
expertise, the set of parameters that describe the user's 
level of knowledge about the domain, including: 
• A list of basic underlying concepts that are important 
in the domain of the knowledge base and that the user 
understands. 
• The specific objects the user knows in the domain, 
that is the user's local expertise (a list of pointers into 
the knowledge base). 
The textual component of the system decides what to 
include in the description. It looks into the knowledge 
base and, based on the user model and the discourse 
strategies, chooses appropriate facts. The output of the 
textual component is a conceptual representation of the 
content of the description. This representation is passed 
through an interface, which makes lexical choices for 
the various concepts included in the description. The 
interface uses the focus of a proposition and the past 
discourse to guide its decision process. (However, as 
our emphasis in this work is on the content of a 
description as opposed to its phrasing, we have not 
studied in depth the complexity and subtleties of lexical 
choice.) Finally, a surface generator constructs English 
sentences. The surface generator is based on the one 
used by McKeown (1985) in the TEXT system. This 
generator unifies the input with a functional grammar 
(Kay 1979) to produce English sentences. We have 
extended and improved the performance of this pro- 
gram, and augmented the functional grammar it uses 
(Paris and Kwee 1985; McKeown and Paris 1987). 
Request for an object desczipt~on 
OSm NO~I~ I r ~-~ ~ ~r 
- - t- - +Az~_ .. 
I NI~¢ concepts?l I , -- ,, l ,."J ,..,o., L I 
I 1o e l content of 
I LLst of ob~ectsl I the_description 1 
L - ~2*~-;-~j ! 
i 
I i Dict£ona~ Interface I (whirl lexical cho£¢e ~l ude) 
I l 
I content of|the delc:~pt~on 
I vi~ lex£cal choice ude 
! 
ob\]ee~ ~n Enql£sh 
Figure 10. The TAILOR System. 
5.2 IMPLEMENTATION OF THE STRATEGIES 
The constituency schema and the process trace strate- 
gies are implemented using an augmented transition 
network (ATN) (Woods 1973). The arcs joining the 
various nodes in the network specify what information 
is to be retrieved from the knowledge base, under what 
conditions (the arcs contain a test), and which node to 
go to next. Figures 11 and 13 present the nets used for 
the strategies. 
I"13mUTIVlS a 
Jl.rklP Ctlt,~iit ~I~ICY 
C4\[/,!~t w~ 
AT'ntmU1W! 
~111x4 
AI"nuBuTIYE 
Figure 11. The Constituency Schema. 
In the constituency schema, shown in Figure 11, the 
arcs correspond to the predicates from the schema. (See 
McKeown (1985) for details of a similar system.) These 
72 Computational Linguistics, Volume 14, Number 3, September 1988 
C~cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise / 
predicates define the type of information to be taken 
from the data base. They are: 
• identification, presenting the more general concept of 
which the present object is an instance; 
• constituency, giving the components of an entity, if 
there are any; 
• attributive, providing different attributes of an object 
(such as its shape or material); and 
• cause-effect, providing some causal relations be- 
tween entities or relations. 
The process of filling the ATN for the constituency 
schema strategy to describe a microphone is shown in 
Figure 12. The constituency schema would be chosen 
by TAILOR if the user model exhibits local expertise 
about the microphone (or if the user is classified as an 
expert). In this example, the identification predicate is 
first applied to the microphone. The identification pred- 
icate provides the superordinate of the object together 
with the function of the object. The constituency pred- 
icate is then applied, providing the subparts of the 
microphone, together with their properties or purposes. 
Finally, the depth-identification predicate is applied to 
each subpart, the doubly-resonant system and the dia- 
;Stepping through the Constituency Schema to describe 
; a MICROPHONE. (English noun-phrases are used 
here for clarity sake.) 
Applying the predicates to MICROPHONE: 
Identification 
predicate: 
Constituency 
predicate: 
DEVICE; (used-for: 
change soundwaves into current) 
DIAPHRAGM 
(shape disc, material aluminium), 
DOUBLY-RESONANT-SYSTEM 
(used-for: broaden response) 
Depth-constituency for DOUBLY-RESONANT-SYS- 
TEM: 
CARBON-CHAMBER, AIR-CHAMBER 
Depth-attributive for DIAPHRAGM: 
(edges clamped) 
TAILOR output: 
The microphone is a device that changes soundwaves 
into current. It has a disc-shaped aluminium diaphragm 
and a doubly-resonant system to broaden the response. 
The system has a carbon chamber and an air chamber. 
The diaphragm is clamped at its edges. 
Figure 12. Stepping through the Constituency Schema. 
NEXT-MAIN-LI~ SIDE LINK? 
JUMP 
~~/~~ ATTRIBUTIVE OR 
(NO SUBSTEPS) I SWITCH? 
SUBSTEPS/ 
HEXT-HAXN-LINK SUBSTEPSI SIDE LINK? 
~ATTRIBUTIVZ OR 
SWITCH? 
Figure 13. The Process Trace. 
phragm. The English output shown in the figure (and in 
the following figures) is the actual output from TAI- 
LOR. 
The network for the process trace is shown in Figure 
13. An example of following the process trace strategy 
is presented in Figure 14. In this network, the arcs 
dictate how to trace the knowledge base to form an 
answer, but they are not linguistic predicates as in the 
network corresponding to the constituency schema. 
They mainly dictate how to follow the causal links in the 
knowledge base. (Details about the process trace can be 
found in Paris and McKeown (1987) and Paris (1984)). 
The arcs are: 
Next-main-link: this arc dictates to follow the next 
link on the main path. The main path is the sequence 
of events that is performed in order for an object to 
achieve its function. 
Side-link?: A side-link is a link that is not part of the 
main path, but that comes off an event on the main 
path. This arc tests to see whether there is a side link 
caused by an event at this point. The decision to 
mention the side link is based on the importance of 
that link. s 
Attributive: This arc is similar to the attributive 
predicate in the constituency schema. If information 
about a part just introduced is available in the knowl- 
edge base, this arc will be taken. 
Substeps?: If an event consists of several substeps, 
the substeps are traced first. To traverse the substeps, 
the subroutine substep is called for each substep. This 
subroutine is very similar to the main graph, but does 
not allow for a further decomposition of events. 
In Figure 14, the link "the diaphragm vibrates" can 
be divided into substeps, namely "the diaphragm 
goes forward" and "the diaphragm goes backward". 
Computational Linguistics, Volume 14, Number 3, September 1988 73 
C~cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
;The program traces the process information for a naive user 
Causal link(l): {M-CAUSES} relates the two relations: 
\[ONE\] P-SPEAKS.-INTO \[MICROPHONE\] 
\[SOUNDWA VES\] P-HITS \[DIAPHRAGM1 
Attributive Information about DIAPHRAGM: \[material: aluminium\] 
\[shape: d:isc\] 
Causal link(2): {M-CAUSES} relates the two relations: 
\[SOUNDWAVES\] P-HITS \[DIAPHRAGM\] 
\[DIAPHRAGM\] P..VIBRATES 
;Substeps: "Soundwaves increasing" and "soundwaves decreasing" 
Substepl {M-CAUSES} relates the two relations: 
\[SOUNDWA VES\] P-INCREASES \[DIAPHRAGM\] 
\[DIAPHRAGM\] P-SPRING \[Direction; forward\] 
{M-CAUSES} relates the two relations: 
\[DIAPHRAGM\] P-SPRING \[Direction: forward\] 
\[GRANULES\] P-COMPRESS 
{M-CAUSES} relates the two relations: 
\[GRANULES\] P-COMPRESSES 
\[RESISTANCE\] P-DECREASES 
{M-CAUSES} relates the two relations: 
\[RESISTANCE\] P-DECREASES 
\[CURRENT\] P-INCREASE 
Substep2 {M-CAUSES} relates the two relations: 
\[SOUNDWA VES\] P-DECREASES \[DIAPHRAGM\] 
\[DIAPHRAGM\] P-SPRING \[Direction: backward\] 
; \[The remainder of the trace is omitted here for brevity.\] 
TAILOR output: 
A person speaking into the microphone causes the soundwaves to hit the diaphragm of the microphone. The 
diaphragm is aluminium and disc-shaped. The soundwaves hitting the diaphragm causes the diaphragm to vibrate. 
When the intensity of the soundwaves increases, the diaphragm springs forward. This causes the granules of the 
button to be compressed. The compression of the granules causes the resistance of the granules to decrease. This 
causes the current to increase. Then, when the intensity decreases, the diaphragm springs backward. This causes 
the granules to be decompressed• The decompression of the granules causes the resistance to increase .This causes 
the current to decrease• The vibration of the diaphragm causes the current to vary. The current varies like the 
intensity varies• 
Figure 14. Tracing the process information for a naive user. 
The process trace can be continued at the substep 
level• Then, once the substeps have been traced 
through, the trace returns to the top-level description• 
(Only one substep is fully shown in Figure 14 for 
brevity. See (Paris 1985) for details.) We can also 
choose to not follow the substeps in order to generate 
a shorter description. This factor is incorporated into 
the arc test. 
By representing the two strategies in this formalism, we 
immediately obtain the control structure necessary to 
switch strategies, since it is possible to jump from a 
node in one part of the network to a node in a different 
part. The decision points are thus marked as special 
tests on the arcs joining nodes. 
As a general test for deciding which strategy to use to 
describe a part, TAILOR looks into the user model to 
check if a superordinate of the part (or the part itself) is 
known to the user. If the part is known, the constitu- 
ency schema is used. Otherwise, the process trace is 
chosen. This test is invoked before beginning generation 
and at any point in the schema where it is possible to 
switch strategies. The test also checks on the length of 
the discourse planned so far and the number of parts to 
avoid generating overwhelmingly long texts. The proc- 
ess of stepping through the ATN and switching strategy 
is shown in Figure 15. (The corresponding generated 
text was shown in Figure 8.) Note that to avoid gener- 
ating very long texts, substeps are omitted when the 
process trace is chosen to describe a subpart. After the 
74 Computational Linguistics, Volume 14, Number 3, September 1988 
C6cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
User Model: Local Expertise: Loudspeaker 
; Stepping through the Constituency Schema to describe a TELEPHONE. 
; Switching to the Process Trace to describe the TRANSMITTER. 
Applying the predicates to TELEPHONE: 
Identification predicate: DEVICE; (used-for: 
change soundwaves into soundwaves) 
Constituency predicate: DIALING MECHANISM 
TRANSMITTER 
(used-for: change soundwaves into current) 
LINE 
CORD 
RECEIVER 
(used-for: change current into soundwaves) 
HOUSING 
(properties: color: various, shape: various) 
Need to switch to process trace for the TRANSMITTER 
Introduction: 
Identification predicate: MICROPHONE 
Causal link (1):{M-CAUSES} 
relates the two relations: 
\[ONE\] P-SPEAKS-INTO \[TRANSMITTER\] 
\[SOUNDWAVES\] P-HITS \[DIAPHRAGM\] 
Causal link (2): {M-CAUSES} 
relates the two relations: 
\[SOUNDWAVES\] P-HITS \[DIAPHRAGM\] 
\[DIAPHRAGM\] P-VIBRATES 
\[Substeps omitted\] 
Causal link (3): {M-CAUSES} 
relates the two relations: 
\[DIAPHRAGM\] P-VIBRATES 
\[CURRENT\] P-VARIES 
Side Link (4): {M-EQUIVALENT-TO} 
relates the two relations: 
\[CURRENT\] P-VARIES 
\[SOUNDWAVE-INTENSITY\] P-VARIES 
Returning to the Constituency Schema: 
Applying the predicates to RECEIVER: 
Identification: RECEIVER is a LOUDSPEAKER 
(difference: small aluminium diaphragm) 
Applying the predicates to HOUSING: 
Attributive: HOUSING r-contains TRANSMITTER 
HOUSING r-contains RECEIVER 
HOUSING r-connected-to DIALING MECHANISM 
by CORD 
Applying the predicates to LINE: 
Attributive: DIALING MECHANISM r-connected-to WALL 
by LINE 
Figure 15. Switching strategy. 
initial description is generated, a mechanism allows the 
user to request the information that was omitted for 
brevity sake (see Paris 1987). 
For each object of the knowledge base, TAILOR can 
generate descriptions aimed at expert and naive users, 
and texts that combine the two strategies for users who 
are along the knowledge spectrum and only know a few 
objects. The ability to combine strategies allows TAI- 
LOR to generate a greater variety of texts than would 
otherwise be possible. For example, starting with the 
constituency schema as the initial strategy for describ- 
ing the pulse-telephone, which has one superordinate 
Computational Linguistics, Volume 14, Number 3, September 1988 75 
C~cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
and three functionally important parts, TAILOR can 
generate eight different descriptions depending on. var- 
ious user models. This ability also allows TAILOR to 
generate a description tailored to any individual user 
model. 
6 FURTHER WORK AND RELATED ISSUES 
We are extending our work by examining the recursive 
use of both strategies. Currently, TAILOR can call the 
strategies recursively for each subpart of the top level 
object being described, but not for the components of 
these subparts. Similarly, the process is only traced one 
level down, that is only the top-level relations are 
broken into substeps. This kind of recursion could 
occur as deeply as the knowledge base allows for. We 
think that such a level of detail is appropriate only if the 
user asks for a longer description. We plan to implement 
a mechanism that would allow the user to request such 
a description, either initially or after a description has 
already been provided by the system and the user wants 
additional information. As we already mentioned, we 
think that the user's domain knowledge, in particular 
knowledge of basic concepts, plays a role in determin- 
ing the depth needed. The depth of the knowledge base 
itself may also affect the depth of the description, as 
parts or processes may be considered too minor to be 
mentioned beyond a certain threshold. 
6.1 DETERMINING THE LEVEL OF EXPERTISE 
In this work, we have not addressed the issue of 
determining the level of expertise of the user. This is 
obviously an important question that needs to be stud- 
ied. We believe that it is possible to infer the user's level 
of expertise. Some relevant factors are: 
• The user type. Some classes of users may be likely to 
be naive while others may be likely to be expert. 
Identifying a user as part of these classes can give 
insight into how much that user might know, and 
provide a starting point. This is similar to using 
stereotypes to model the user (Rich 1979; Chin 1986). 
The question type. The kind of question asked 
("What is a tape recoder?" as opposed to "Does this 
disk drive have three bearings?") can give further 
information about how much the user knows. 
• The depth in the knowledge base (both in the compo- 
nents tree and the generalization hierarchy) of the 
object of the question ("Describe a disc drive" as 
opposed to "Describe the track assembly in a disk 
drive")(Paris 1984). 
• The previous discourse. If a user asks the same 
question twice, it is indicative that the answer was not 
understood, perhaps because it assumed knowledge 
the user did not have. 
It would be necessary to study how the user's domain 
knowledge can be inferred from these factors and how 
these factors affect each other. This seems to be a hard 
problem. 
Finally, our user model at this point is coarse 
grained, in that it contains a list of objects and concepts 
that the user knows. A more detailed model might 
include exactly which facts the user knows about ob- 
jects, and specifically which basic concepts are under- 
stood. We have shown that a system benefits from a 
user model that indicates a user's knowledge about the 
domain, even when the model is not a detailed one. 
While we feel that a detailed user model would be much 
harder to obtain, it would be interesting to see whether 
such a model would allow a system to provide more 
appropriate answers. 
6.2 INCORPORATING THE USER'S GOAL 
It is clear that the goal of the user also influences the 
content of an answer, or of a description. It would be 
interesting to examine how the goal and the level of 
expertise combined can help a program in choosing 
appropriate facts from the knowledge base. In this 
research, we found that the level of expertise affects the 
kind of information to include in a description. This is 
also clearly true for a user's goal: the information 
provided in the description should help the user in 
pursuing his/her goal. Thus, depending on a user's goal, 
the information included would vary. The interaction of 
these two factors can become quite complex. However, 
until both factors are fully understood individually, we 
feel that it will be very hard to determine exactly how 
they interact. 
6.3 FEASIBILITY AND EXTENSIBILITY OF THIS APPROACH 
In this work, we make the assumption that a system that 
tailors its answer to its users will be most useful. This is 
true only provided that this tailoring does not hinder the 
system's performance or increase its complexity signif- 
icantly. We argue that tailoring a description to a user's 
level of expertise using the method described above will 
not add to the cost of generation and yet might provide 
better answers. Whether a generation system tailors its 
answers to users or not, it needs to employ a discourse 
strategy to guide its decision process lest the resulting 
text be incoherent. The two discourse strategies used by 
TAILOR are of comparable complexities, and there is 
not much cost added to combine them. 
While this work was done only with respect to 
generating descriptions of complex devices, we think 
this approach will be useful in any information seeking 
environment to which users with different background 
and knowledge levels have access. Such environment 
could be a large knowledge base of facts (of the kind of 
an encyclopedia), or a help system. Providing different 
information in an answer might also be done in explain- 
ing the behavior of an expert system that is used both as 
a teaching tool and as a problem solving engine. 
76 Computational Linguistics, Volume 14, Number 3, September 1988 
C~cile L. Paris Tailoring Object Descriptions to a User's Level of Expertise 
7 CONCLUSIONS 
In this paper we have demonstrated that the user's 
domain knowledge can be used as a factor in tailoring an 
answer. In particular, we have shown how the descrip- 
tion of a complex physical object might be tailored to a 
user's level of expertise. We presented different kinds 
of knowledge users can have, explaining how a system 
can take them into consideration in order to generate a 
description. From our studies of texts, we have found 
two distinct discourse strategies that are used in de- 
scribing complex devices. We postulated that the level 
of expertise of the user affects the kind of information 
given as opposed to just the amount of detail provided. 
Even though we conducted this study in the domain of 
complex physical objects, we believe this result can 
extend to other domains. We thus propose that a user 
model containing information about the user's domain 
knowledge can be used in a question answering system 
to guide the decision process. We presented the two 
distinct descriptive strategies that can be used in a 
question answering program and showed how they can 
be mixed to include the appropriate information from 
the knowledge base, based on the information contained 
in user model. 
Finally, we presented TAILOR, a program that gen- 
erates descriptions tailored to users with various levels 
of expertise. TAILOR employs one of the two discourse 
strategies described to generate a text for a novice or an 
expert. TAILOR is also able to automatically mix the 
strategies to provide device descriptions tailored to 
users whose domain knowledge fall anywhere along the 
knowledge spectrum. By representing explicitly the 
user's domain knowledge in terms of parameters, TAI- 
LOR does not require an a priori set of stereotypes but 
can provide wide variety of descriptions for a whole 
range of users. 
ACKNOWLEDGMENTS 
Many thanks go to Kathleen McKeown and Michael Lebowitz for 
their help in both the research and the writing of this paper, and to 
TjoeLiong Kwee for his work on the functional grammar. We also 
thank the anonymous reviewer for his/her useful comments. 
This research was supported in part by the Defense Advanced 
Research Projects Agency under contract N00039-84-C-0165, and the 
National Science Foundation grant ISI-84-51438. 
NOTES 
1. We hope that by choosing several sources, stylistic differences on 
our results are minimized. We studied about 15 examples from 
each encyclopedia and textbook and a few from the manuals. 
2. We are using McKeown's notation: "{}" indicate optionality, "/" 
alternatives, "+" that the item may appear 1 or more times, and 
"*" that the item may appear 0 or more times. Finally, ";" is 
used to indicate that the propositions could not be clearly 
classified as corresponding to one predicate. We changed 
McKeown's schema slightly, by adding the identification predi- 
cate as an option for the first predicate of the schema. 
3. The original entry was in one paragraph only. We divided it into 
three paragraphs for clarity. More details about this analysis are 
given in Paris (1985). 
4. The original entry contained two paragraphs. The second one has 
been divided for clarity. 
5. Research in reading comprehension indicates that readers indeed 
use their previous knowledge in order to understand new texts 
(Anderson et al. 1977; Wilson and Anderson 1986). 
6. Some parts do not play an important part on the mechanical 
process associated with the object. For example, the housing of 
the telephone does not have an important role in the functionality 
of the telephone. Such a part would not be involved in one of the 
causal links contained in the knowledge base. 
7. We do not claim that this is the optimal value for the threshold 
indicating at which point the constituency schema should be used. 
This threshold cannot be set with certitude without further 
experimentation. Also note that, while these heuristics allow the 
system to generate reasonable descriptions given the user's 
domain knowledge, there is no clear "best" descriptions. 
8. We have mentioned that there are different kinds of links between 
events. The types of links are ranked in order of importance, and 
the most important ones are mentioned. There are actually 
different types of side-links. These are not indicated in the figure 
for simplicity. The reader is referred to Pads (1987) for details. 

REFERENCES 
Anderson, R.C.; Spiro, R.J.; and Anderson, M.C. 1977 Schemata as 
scaffolding for the representation of information in connected 
discourse. Technical Report 24, Center for the Study of Reading, 
Urbana, IL. 
Appelt, D.E. 1982 Planning Natural Language Utterances. In Pro- 
ceedings of the Second National Conference on Artificial Intelli- 
gence, Pittsburgh, PA: 59-62. 
Appelt, D.E. 1985 Planning English Sentences. Cambridge Univer- 
sity Press, Cambridge, England. 
The New Encyclopedia Britannica 1984 Encyclopedia Britannica 
Incorporation. 
Britannica Junior Encyclopedia 1963 Encyclopedia Britannica Incor- 
poration, William Benton Publisher. 
Carberry, S. 1983 Tracking User Goals in an Information-Seeking 
Environment. In Proceedings of the American Association of 
Artificial Intelligence. 
Carberry, S. (this issue) Plan Recognition and User Modeling. Com- 
putational Linguistics. Computation on User Modeling, 1988. 
Kirk and Othmer, (eds.) 1978 Encyclopedia of Chemical Techology. 
John Wiley & Sons, New York, NY. 
Chevrolet Service Manual 1978 General Motors Corporation. 
Chi, M.T.H.; Glaser, R.; and Rees, E. 1981 Expertise in Problem 
Solving. Advances in the Psychology of Human Intelligence. 
Lawrence Erlbaum, Hillsdale, NJ. 
Chin, D.N. 1986 User Modeling in UC, the UNIX Consultant. In 
Proceedings of Computer Human Interactions. 
William Halsey (ed.) 1962 Collier's Encyclopedia The Crowell-Collier 
Publishing Company. 
The New Encyclopedia of Science 1982 Raintree Publishers. 
Grice, H.P. 1975 Logic and Conversation. In Cole P. and Morgan J.L. 
(eds.), Syntax and Semantics. Academic Press, New York, NY. 
Grimes, J.E. 1975 The Thread of Discourse. Mouton, The Hague, 
Paris, France. 
Hobbs, J. 1978 Why is a Discourse Coherent? Technical Report 176, 
SRI International, Menlo Park, CA. 
Hobbs, J. 1978 Coherence and Coreference. Technical Note 168, SRI 
International, Menlo Park, CA. 
Hobbs, J. and Evans, D. 1980 Conversation as Planned Behavior. 
Cognitive Science, 4(4): 34%377. 
Hoeppner, W.; Morik, K.; and Marburger, H. 1984 Talking it Over: 
The Natural Dialog System HAM-ANS. Technical Report ANS- 
26, Research Unit for Information Science and Artificial Intelli- 
gence, University of Hamburg, W. Germany. 
Kaplan, S.J. 1982 Cooperative Responses from a Portable Natural 
Language Query System. Artificial Intelligence 2(19). 
Kay, Martin 1979 Functional Grammar. In Proceedings of the 5th 
meeting of the Berkeley Linguistics Society, Berkeley, CA. 
Lancaster, J.S., and Kolodner, J.L. 1987 Problem Solving in a 
Natural Task as a Function of Experience. Technical Report, 
School of Information and Computer Science, Georgia Institute of 
Technology, Atlanta, GA. 
Lebowitz, M. 1983 RESEARCHER: An Overview. In Proceedings of 
the Third National Conference on Artificial Intelligence. Ameri- 
can Association of Artificial Intelligence, Washington, DC. 
Lebowitz, M. 1985 RESEARCHER: An experimental intelligent 
information system. In Proceedings of the Ninth International 
Joint Conference on Artificial Intelligence, Los Angeles, CA: 858- 
862. 
Lebowitz, M. 1986 An experiment in intelligent information systems: 
RESEARCHER. Intelligent Library and Information Systems, 
Ellis Horwood, London, England. 
Lehnert, W. 1977 A Conceptual Theory of Question Answering. In 
Proceedings of the International Joint Conferences of Artificial 
Intelligence, Cambridge, MA. 
Mann, W.C. 1984 Discourse Structure for Text Generation. Technical 
Report ISI/RR-84-127, Information Sciences Institute, Marina del 
Rey, CA. 
McCoy, K.F. 1983 Correcting Misconceptions: What to say When the 
User is Mistaken. In Proceedings of the Conference on Human 
Factors in Computing Systems. Human Factors, Boston, MA. 
McCoy, K.F. 1986 The ROMPER System: Responding to Object- 
Related Misconceptions Using Perspective. In Proceedings of the 
24th Annual Meeting of the Association of Computational Linguis- 
tics, New York, NY. 
McCoy, K.F. (this issue) Reasoning on a Highlighted User Model to 
Respond to Misconceptions. Computational Linguistics 1988. 
McKeown, K.R. 1985 Text Generation: Using Discourse Strategies 
and Focus Constraints to Generate Natural Language Text. 
Cambridge University Press, Cambridge, England. 
McKeown, K.R.; Wish, M.; and Matthews, K. 1985 Tailoring Expla- 
nations for the User. In Proceedings of the Ninth International 
Joint Conferences on Artificial Intelligence, Los Angeles, CA. 
McKeown, K.R. and pads, C.L. 1987 Functional Unification Gram- 
mar Revisited. In Proceedings of the 25th Annual Meeting of the 
Association of Computational Linguistics, Palo Alto, CA. 
The New Book of Knowledge-The Children's Encyclopedia 1967 
Grolier Incorporation, New York, NY. 
Pads, C.L. 1984 Determining the Level of Expertise. In Proceedings 
of the First Annual Workshop on Theoretical Issues in Conceptual 
Information Processing. Atlanta, GA. 
Pads, 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. 
Pads, C.L. and Kwee, T.L. 1985 Guide to the Unification Process 
and its Implementation; Progress Report on Extending the Gram- 
mar. Technical Report, Computer Science Department, Columbia 
University, New York, NY. 
Pads, C.L. and McKeown, K.R. 1986 Discourse Strategies for 
Describing Complex Physical Objects. Paper presented at the 
Third International Workshop on Natural Language Generation, 
Nijmegen, The Netherlands. In G. Kemper (ed.), Natural Lan- 
guage Generation: Recent Advances in Artificial Intelligence, 
Psychology, and Linguistics. Kluwer Academic Publishers, Bos- 
ton, MA/Dordrecht, Holland. 
Pads, C.L. 1987 The Use of Explicit User Models in Text Generation: 
Tailoring to a User's Level of Expertise. Ph.D. thesis. Columbia 
University, Department of Computer Science, New York. 
Quilici, A.; Dyer, M.; and Flowers M. (this issue) Providing Explan- 
atory Responses to Plan-Oriented Misconceptions. Computational 
Linguistics 1988. 
Rich, E.A. 1979 User Modeling Via Stereotypes. Cognitive Science 3: 
329-354. 
Shepherd, H.R. 1926 The Fine Art of Writing. Macmillian Co., New 
York, NY. 
Tennant, H. 1978 The Evaluation of Natural Language Question 
Answerers. Technical Report, University of Illinois at Urbana/ 
Champaign Ph.D. proposal, Department of Computer Science, 
Advanced Automation Group, Coordinated Science Laboratory. 
Wallis, J.W. and Shortliffe, E.H. 1982 Explanatory Power for Medical 
Expert Systems: Studies in the Representation of Causal Relation- 
ships for Clinical Consultation. Technical Report STAN-CS-82- 
923, Heuristics programming Project, Department of Medicine and 
Computer Science, Stanford University, Stanford, CA. 
Wasserman, K. 1985 Unifying Representation and Generalization: 
Understanding Hierarchically Structured Objects. Department of 
Computer Science, Ph.D. thesis, Columbia University, New 
York, NY. 
Wassernmn, K. and Lebowitz, M. 1983 Representing Complex Phys- 
ical Objects. Cognition and Brain Theory 6(3): 333-352. 
Weissler, A. and Weissler, P. 1973 A woman's guide to fixing the car. 
Walker ~md Company, New York, NY. 
Wilson, P.q. and Anderson, R.C. 1986 What They Don't Know Will 
Hurt Them: The Role of Prior Knowledge in Comprehension. 
Reading Comprehension: From Research to Practice. Erlbaum, 
Hillsd~de, NJ. 
Woods, W. 1973 An Experimental Parsing System for Transition 
Network Grammars. In Rustin, R. (ed.), Natural Language Proc- 
essing. Algorithmics Press, New York, NY. 
