REASONING ON A HIGHLIGHTED USER MODEL 
TO RESPOND TO MISCONCEPTIONS 
Kathleen F. McCoy 
Department of Computer and Information Sciences 
University of Delaware 
Newark, DE 19716 
Responses to misconceptions given by human conversational partners very often contain information 
refuting possible reasoning which may have led to the misconceptions. Surprisingly there is a great deal 
of regularity in these responses across different domains of discourse. For instance, one reason a user 
might have given an object a property it does not have is that the user confused the object with another 
similar object. In correcting such a misconception, a human conversational partner is likely to point out 
this possible confusion. 
This work describes a method for generating responses like the one just described by reasoning on a 
highlighted model of the user to identify possible sources of the error. Through a transcript study a 
number of response strategies were abstracted. Each strategy was associated with a structural 
configuration of the user model. For example, the above mentioned strategy of pointing out a similar 
confused object is associated with a configuration of the user model that indicates the user believes there 
is an important similar object that has the property involved in the misconception. Upon finding that 
configuration in the highlighted user model, the system can respond with the associated strategy. 
Notice that the reasoning must be done on a highlighted user model since the perception of both an 
object's importance and its similarity with another object change with the perspective being taken on the 
domain. This paper investigates how domain perspective can be modeled to provide the needed 
highlighting and introduces a similarity metric that is sensitive to the highlighting provided by the 
domain perspective. Finally, the paper shows how the highlighting affects misconception responses. 
1 INTRODUCTION 
When people interact with a database or expert system, 
it is reasonable to expect that they might reveal a 
misconception about an object modeled by the system. 
Since a human conversational partner would correct 
such a misconception if it was important to the current 
goals of the conversation, our database and expert 
systems should also be equipped with this ability. 
In order to investigate how the process of correcting 
misconceptions might be automated, a study of tran- 
scripts of both humans .interacting with what they 
thought were expert systems (Malhotra 1975, Malhotra 
and Sheridan 1976, Schuster 1982), and humans inter- 
acting with other humans to achieve some goal (Pollack, 
Hirschberg, and Webber 1982) was undertaken. The 
transcripts, which varied greatly in their domains of 
discourse, were analyzed to determine if there was any 
regularity in the content and rhetorical force of re- 
sponses given to misconceptions. The intention of this 
analysis was not to mimic the actual behavior found in 
the transcripts, but to use them as a source of intuitions 
about the context and textual shape of responses as well 
as the process of generating them. 
The study revealed that a response to a misconcep- 
tion important to the current discourse goals of the 
participants can be viewed as consisting of three parts: 
1. a denial of the incorrect information; 2. a statement of 
the correct information; and 3. justification for the 
denial and correction given. For a particular type of 
misconception (i.e., one involving a particular kind of 
knowledge), variations in responses could be found in 
the form of the justification given. The justification 
often seemed to refute support that might have led to 
the misconception. While the kind of support someone 
might have for a misconception seems unrestricted, the 
form of the justification was limited for misconceptions 
Copyright 1988 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided 
that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To 
copy otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/88/0100e-e$03.00 
52 Computational Linguistics, Volume 14, Number 3, September 1988 
Katlfleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
involving a particular kind of knowledge. A large num- 
ber of responses found could be accounted for by a 
small number of correction strategies based on the kind 
of justification given (and hence the faulty reasoning 
refuted). 
If a principled reason for using one strategy over 
another could be developed, these strategies could be 
used by a natural language generation system for re- 
sponding to a misconception. In this work the faulty 
reasoning refuted by several of the found correction 
strategies is characterized in a domain-independent 
fashion in terms of the user's beliefs about the domain. 
Therefore, given a highlighted model of the user's 
beliefs about the domain, a generation system can look 
for possible support for the misconception. The re- 
sponse strategy that refuted the kind of support found 
could then be instantiated. Notice that the domain- 
independent characterization of the faulty reasoning 
enables the same strategies and same method for choos- 
ing a strategy to be used given a highlighted user model 
for any domain. 
It is crucial that the user model given to the miscon- 
ception corrector be highlighted by previous discourse 
since the kind of response given by a human conversa- 
tion partner is apparently not only dependent on the 
beliefs about the person being corrected, but also on the 
context in which the misconception occurred. For in- 
stance, we could imagine the following dialog where the 
user exhibits the misconception that T-bills have a 
penalty. A reasonable response is shown. 
U: I am interested in investing in some securities to 
use as savings instruments. I want something 
short-term and I don't have a lot of money to 
invest, so the instrument must have small denom- 
inations. I am a bit concerned about the penalties 
for early withdrawal. What is the penalty on a 
T-Bill? 
R: T-Bills don't have a penalty. Were you thinking 
of Money Market Certificates? 
This response might be prompted by R thinking that U 
came to the misconception by confusing T-Bills with 
Money Market Certificates. 
On the other hand, it is reasonable that the response 
might be different given a different preceding dialog. 
For example: 
U: I am interested in investing in some securities. 
Safety is very important to me, so I would 
probably like to get something from the govern- 
ment. I am a bit concerned about the penalties for 
early withdrawal. What is the penalty on a T-Bill? 
R: T-Bills don't have a penalty. Were you thinking 
of T-Bonds? 
This response may have been prompted by R thinking 
that a confusion between T-Bills and T-Bonds was the 
source of the misconception. The two different re- 
sponses to the same misconception suggest that the user 
model should be influenced by previous discourse. I will 
show how part of this influence can be achieved by a 
highlighting due to the perspective being taken on the 
domain. 
Currently, when a user model containing what the 
system takes to be the user's model for the domain is 
accessed, all user knowledge has equal importance. 
When people engage in a conversation, however, cer- 
tain aspects of their domain model become more impor- 
tant than others. This importance is more than just a 
highlighting of those things that have been explicitly 
mentioned. Rather, certain things that are somehow 
related to those things explicitly mentioned in previous 
discourse are also highlighted. In fact, an orientation on 
the domain is usually established. 
In section 8 I will introduce a notion of object 
perspective that will enable this highlighting effect of 
previous discourse to be incorporated into the user 
model. Thus when the user model is accessed, certain 
things in it will be highlighted while other things will be 
suppressed. I will show how this variable highlighting of 
the user model can explain why the response to a 
particular misconception by a particular user may vary. 
2 RELATED MISCONCEPTION WORK 
The method of correcting misconceptions outlined 
above should be contrasted with the way that miscon- 
ceptions are handled by AI systems today. For the most 
part misconceptions have been left to the Intelligent 
Computer Aided Instruction systems, which basically 
use an a priori listing of misconception-response pairs 
(see, e.g., Brown and Burton 1978; Stevens, Collins, 
and Goldin 1979; Stevens and Collins 1980; Woolf and 
McDonald 1983; Woolf 1984). The major problem with 
these systems is due to their inability to reason about 
the misconception itself, they are completely at a loss 
when faced with a misconception absent from their a 
priori listing. 
The work of Sleeman (1982) on inferring defective 
algebra rules (mal-rules) is based on the observation 
that the a priori listing of misconceptions is a difficult, if 
not impossible, task. Sleeman proposes on-line infer- 
ence of mal-rules based on the answer the student has 
given to a particular problem. Although Sleeman's work 
is a major improvement over the a priori listing ap- 
proach, it still has several problems. First, there is no 
measure of how reasonable or likely a particular mal- 
rule is. In addition, once a mal-rule has been inferred, 
no indication is given concerning how the misconcep- 
tion should be corrected. 
The work of Kaplan (1979) and Mays (1980) is closer 
to the work described here in that they were concerned 
with handling and reasoning about whole classes of 
misconceptions, thereby giving the system the ability to 
handle a potentially infinite number of misconceptions. 
Kaplan and Mays were concerned with responding to 
Computational Linguistics, Volume 14, Number 3, September 1988 53 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
certain types of misconceptions in the context of a 
natural language interface to a database system. They 
worked on detecting and correcting such misconcep- 
tions based on domain independent linguistic cues from 
the user and an enhanced model of the domain. For 
instance, the query "Which faculty take courses?" 
indicates a presumption failure. A truthful response of 
"none" to this query would confirm the user's errone- 
ous belief that faculty can take courses. Mays suggests 
correcting a query like the one above by 1. denying that 
the "take" relation can hold between faculty and 
courses, and 2. describing all correct alternatives which 
can be reached by abstracting on each of the objects and 
the relation involved. This method would produce the 
following kind of response: 
R: I don't believe that faculty can take courses. 
Faculty teach courses. Students take courses. 
Although responses such as this would probably be 
helpful to the user, they have the potential (given a 
complicated domain) for being overly verbose and con- 
taining information that the user does not care about. 
One goal of this work is to provide a more pointed and 
natural response to these same kinds of errors. 
3 KNOWLEDGE AVAILABLE 
The work being done here is in the context of a natural 
language interface to a database or expert system. It is 
an attempt to define a module of an interface that could 
generate a cooperative response to a misconception-- 
the kind of response that would be generated by a 
helpful human conversational partner. In this work, a 
misconception is defined to be some discrepancy be- 
tween system beliefs and user beliefs (as exhibited 
through the conversation). Upon encountering such a 
misconception, the assumption is that the system 
knowledge is correct, and therefore the job of the 
misconception corrector module is to attempt to bring 
the user's knowledge into line with the system's knowl- 
edge. 
The scope of this work is limited by several assump- 
tions about the kind of knowledge available. 
• The system's model of the world contains an object 
taxonomy with attribute/value pairs attached to the 
objects. 
• The system has available to it a user model that 
includes the user's beliefs about the world.' This is 
what the system takes to be the user's model of the 
domain. Although the content of the user model and 
the system's model of the world may differ greatly, 
the user model is in the same form as the system's 
model of the world. Thus while both of these models 
contain an object taxonomy with attribute/value pairs 
attached to the objects, the set of objects in the 
taxonomies and the way these objects are classified, 
as well as the particular attribute/value pairs associ- 
ated with an object, may vary. This model of the user 
may be updated as the conversation progresses. 
• The .,;ystem has available to it certain pieces of 
contextual and discourse information that serve to 
highlight the user model. This highlighting (explored 
below) is gained from a new notion of object perspec- 
tive and from a record of items and attributes which 
have been explicitly focused on in the discourse. 
Given the kind of information assumed in the system's 
and user's models of the world, there are two kinds of 
misconceptions that may occur: misclassifications (a 
user may classify an object wrong) and misattributions 
(a user may give an object an attribute/value pair it does 
not have). For each of these kinds of misconceptions, a 
small number of response strategies 2 were found in the 
transcript study. These were abstracted into response 
schemas. In section 4, examples of the response strat- 
egies found for misclassifications will be examined. For 
each strategy an abstract specification of the content 
will be given. Next we will look at what beliefs about 
the user might have prompted the use of each strategy, 
and characterize these beliefs in terms of the structure 
of a highlighted user model. With this pairing of user 
model structures and response schemas, a misconcep- 
tion can be responded to by looking in the highlighted 
user model for one of the user model structures and 
instantiating the associated schema. Section 5 examines 
response strategies for misattributions. The sections 
following that will concentrate on the highlighting from 
object perspective. A new notion of object perspective 
will be defined and it will be shown how object perspec- 
tive aids in generating context sensitive responses to 
misconceptions. 
This paper is concerned with reasoning on the user 
model to decide how to respond to a misconception. It 
is not concerned with inferring the user model; it is 
assumed that the user model is already available. The 
emphasis in this work is on using the user model, in a 
domain-independent fashion, to respond to a miscon- 
ception in a manner similar to a human conversational 
partner's response. 
The methods described here have been implemented 
in the ROMPER system (Responding to Object-related 
Misconceptions using PERspective). The system takes 
as input a specification of the information that is incon- 
sistent with the system's model of the world, the current 
perspective (described below), and a record of past 
focus. It produces a formal specification of the re- 
sponse. This response specification is passed into the 
Mumble system (McDonald 1980), which, using a gram- 
mar and dictionary written by Robin Karlin (1985), 
produces an actual English response. 
The implemented system works on the financial 
securities domain. In order to show the generality of 
this approach, two different domains will be used in this 
paper. The motivation for the system's method of 
choosing a response strategy will use examples from the 
domain containing whales and fish. In the last sections 
54 Computatiional Linguistics, Volume 14, Number 3, September 1988 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
of the paper (those pertaining to the new notion of 
object perspective), the examples will be taken from the 
domain of financial securities, since the ideas can be 
better motivated with a more complex domain. 
4 MISCLASSIFICATIONS 
One kind of misconception concentrated on in the 
transcript study was that involving the classification of 
an object. From that study three major correction 
strategies, which could be supported with the kind of 
knowledge available in our knowledge base, were ab- 
stracted. These are used to illustrate the type of analysis 
advocated; no claim is being made about the complete- 
ness of this set. The first strategy is exemplified by the 
following dialog. 
U: I thought whales were fish. 
R: No, they are mammals. You may have thought 
they were fish because they are fin-bearing and 
live in the water. However, they are mammals 
since, (while fish have gills) whales breathe 
through lungs and feed their young with milk. 
Let us first examine the content of this response. It can 
be seen as consisting of three parts. In the first sentence 
we have the denial of the incorrect information and the 
statement of the corresponding correct information. 
The remaining sentences comprise the justification for 
the denial and correction given. 
The content of this response can be abstracted into 
the following: 
MISCONCEPTION = X is-a Y 
RESPONSE = 
I. X is-NOT-a Y 
2. X is-a Type-of(X) 
3. X is like Y because both share attributes-of(X) N 
attributes-of(Y) 
4. BUT X has attributes-of(X) -- attributes-of(Y) 
5. WHILE Y has attributes-of(Y) -- attributes-of(X) 
The justification is contained in (3-5) of the rule. Notice 
that it can be seen as a concede/override pair. First, 
information is conceded that actually supports the mis- 
conception. This consists of attributes that whales and 
fish have in common. These attributes are a potential 
source of the misconception since an object that shares 
many attributes with a superordinate is likely to belong 
to that superordinate class. One way to view this 
conceding of intersecting attributes is as R acknowledg- 
ing the apparent support for the misconception lest U 
think that that part of his/her domain model is wrong. 
The response goes on to override the conceded 
information by offering attributes that differentiate 
whales and fish. These are attributes that whales have 
but fish cannot, and attributes that fish have that whales 
do not. 
By analyzing the justification in this way, not only do 
we get an abstract representation that can be used by a 
generation system, but we get some insight into when it 
would be reasonable to use such a response. It makes 
sense to use a response that first concedes a similarity 
between the misclassified object and the object that was 
wrongly given as its superordinate (from now called the 
posited superordinate) and then overrides that con- 
ceded information, when such a similarity is a possible 
source of the misconception. This might be the case 
when the user model shows that the user believes that 
such a similarity exists. Thus a generation system may 
generate its response by instantiating the above rule if 
the user has misclassified the object X as a Y, and an 
analysis of the user model shows that the user believes 
X is similar to Y. Because of this reason for using this 
response strategy, it has been termed the like super 
response strategy. 
A second kind of strategy is exemplified by the 
following dialog. Imagine a recent conversation cen- 
tered on sharks being large fish that people are typically 
afraid of. Following that, the conversation turns to 
classifying whales. The user says: 
U: I thought whales were fish, too. 
R: No, they are mammals. You may have thought 
they were fish since they are like the fish, sharks, 
in that both are large aquatic creatures and both 
scare people. However, whales are mammals 
since, (while fish have gills) whales breathe 
through lungs and feed their young with milk. 
Again, this response can be seen as consisting of a 
denial, a correction, and a justification. The denial and 
correction are contained in the first sentence. As in the 
previous example, the remaining sentences comprise 
the justification in the form of a concede/override pair. 
The conceded information, while dependent on at- 
tributes, differs from that given above. The abstract 
content rule can perhaps best illustrate this difference: 
MISCONCEPTION = X is-a Y 
RESPONSE = 
I. X is-NOT-a Y 
2. X is-a Type-of(X) 
3. 3Z (Z is-a Y /~ X is like Z because both have 
attributes-of(X) A attributes-of(Z)) 
4. HOWEVER X is-a Type-of(X) because X has at- 
tributes-of(Type-of(X)) -- attributes-of(Y) 
The major difference between this response and the like 
super response is that instead of conceding a similarity 
between the misclassified object and the posited super- 
ordinate itself, a similarity between the misclassified 
object and a descendant of the posited superordinate is 
conceded. This response is reasonable if the user came 
to the misconception by reasoning: "I don't know how 
to classify whales, but I do know how to classify sharks. 
Computational Linguistics, Volume 14, Number 3, September 1988 55 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
Sharks are fish and whales are very similar to sharks. 
Therefore it is reasonable to believe that they are 
classified the same." Evidence for this kind of reason- 
ing might be a similarity found in the user model 
between the misclassified object and some descendant 
of the posited superordinate. Finding such a similarity 
would justify the use of this like-some-super strategy. 
A final strategy found for misclassifications is exem- 
plified by the following: 
U: I thought whales were fish. 
R: No, they are mammals. 
This is the degenerate no support strategy, which con- 
tains no justification for the correction. At first glance 
the use of this strategy seems rather puzzling. Since I 
have assumed that the misconception is important, one 
would expect a full response to be given. However, no 
justification was included in the response. 
This last response is an example of a response in a 
situation where no evidence was found in the high- 
lighted user model for any of the preidentified kinds of 
faulty reasoning. If R could not figure out what led to 
the misconception, then R could not include justifica- 
tion to refute the support! The most that R could do in 
this situation was deny the incorrect information and 
state the corresponding correct information. This is 
exactly what was done in the response above. For this 
reason the response has been termed the no support 
response, and it can be used when the user model 
contains none of the evidence that would key off any of 
the other correction strategies. 
5 MISATTRIBUTIONS 
Misattributions are the second kind of misconceptions 
that may arise given the kind of knowledge assumed. A 
misattribution occurs when the user gives an object an 
attribute/value that the object does not have in the 
system's domain model. 
The transcript study revealed that a misattribution 
response could also be broken into three parts. In this 
case, however, the justification took the form of offer- 
ing another statement that would change what the user 
said in a small way in order to make it a correct 
statement. This is illustrated in the following dialog, 
which is an example of the first kind of correction 
strategy found. Imagine that we are discussing sea 
animals and have previously discussed sharks and their 
attributes. We move to whales. 
U: I thought whales had gills. 
R: No, they breathe through lungs. Were you think- 
ing of sharks? 
The response might go on to discuss the similarities and 
differences between whales and sharks. 
Notice what was done in this response. We have the 
denial and the correction as before. This time the 
justification offers another object that is similar to the 
object involved in the misconception but for which the 
misattributed attribute/value pair is true. This content is 
captured in the following rule: 
MISC, ONCEPTION = X has attribute Y with value V 
then RESPONSE = 
1. X NOT have attribute Y with value V 
2. X has attribute Y with value W 
3. 3 Z (Z has attribute Y with value V), did you mean 
Z? 
This kind of strategy would be effective if the user's 
misconception results from either confusing the object 
being discussed with some other object with the misat- 
tributed property, or through an incorrect analogy with 
such an object. Mention of the object in the case of a 
confusion prevents continuation of the confusion. In the 
case of an incorrect analogy, mention of the other object 
alerts the user to the inappropriateness of analogy and 
also prevents the continued usage of the analogy. No- 
tice that in either case, a failure to mention the second 
object will probably bring protest on the part of the 
user. For instance, suppose that the preceding miscon- 
ception was arrived at because the user was confusing 
whales and sharks. If the expert just responds, "No, 
they breathe through lungs", it is likely that the user 
will counter, "But you just told us that they had gills!" 
By mentioning the confused object, the user becomes 
aware of the confusion. 
It makes sense to use this wrong object strategy 
when a possible source of the misconception is either an 
object confusion or the usage of a bad analogy. Evi- 
dence for this would be the existence of some similar 
object from the highlighted user model that has the 
misattributed property. 
A second kind of correction strategy used for misat- 
tributions is shown here: 
U: I thought whales had gills. 
R: No, whales breathe through lungs. Did you mean 
you thought they had fins? 
This is very similar to the last response except, instead 
of offering another object, the strategy offers a similar 
attribute which makes the user's statement correct. The 
content rule for this response follows: 
MISCONCEPTION = X has attribute Y with value V 
then RESPONSE = 
I. X does not have attribute Y with value V 
2. X has attribute Y with value W 
3. 3Q=IU(X has attribute Q with value U /k simi- 
lar(Y,Q), did you mean X has attribute Q with value 
u'?) 
This strategy, termed the wrong attribute strategy, is 
used when there is a similarity of attributes within the 
56 Computational Linguistics, Volume 14, Number 3, September 1988 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
highlighted user model. (The no support strategy also 
occurs in the case of misattributions.) 
6 USER MODEL ANALYSIS 
Given the association of user model configurations with 
response strategies given in the previous sections, we 
can come up with a method for deciding how to respond 
based on the type of misconception along with an 
analysis of the highlighted user model. Basically the 
user model analysis looks for one of the preidentified 
configurations and, if one is found, suggests instantiat- 
ing the associated strategy. The following rule captures 
the way that the ROMPER system implements what has 
been discussed so far. 
IF misconception = "X is-a Y" 
THEN 
IF similar(X,Y) 
THEN instantiate like super schema 
ELSEIF 3Z (Z is-a Y)/~ similar(X,Z) 
THEN instantiate like-some-super schema 
ELSE instantiate no support schema 
ELSEIF misconception = "X has attribute Y with 
value V" 
THEN 
IF 3Z ((Z has attribute Y with value V)/~ 
similar(X,Z)) 
THEN instantiate wrong object schema 
ELSEIF 3Q3U ((X has attribute Q with value U)/~ 
similar(Y,Q)) 
THEN instantiate wrong attribute schema 
ELSE instantiate no support schema 
Notice that each of the tests for instantiating a schema 
hinges on the similarity assessment of two objects. 
These assessments must be context dependent. Be- 
cause of this, it is crucial that the user model analysis be 
done on a user model highlighted by previous discourse 
and that the similarity metric take advantage of this 
highlighting. The highlighting and similarity metric used 
by ROMPER will be discussed below. 
This method for correcting misconceptions suggests 
a model of natural language generation that is similar to 
that put forth by McKeown (1982) but which differs 
from McKeown's model in several ways. 
Both McKeown and this work concentrate on deter- 
mining the content and textual shape of a response. 
McKeown is concerned with responses to questions 
about the structure of a data base. Upon encountering 
such a question McKeown first delimits a relevant 
knowledge pool using fairly simple mechanisms. This 
relevant knowledge pool contains that information from 
the knowledge base that could possibly be included in 
the response; the actual generated response need not 
exhaust this pool. Next, based on the goal of the 
discourse (as determined by the question type) and a 
characterization of the relevant knowledge pool (again, 
a simple test) a response schema is chosen for the 
response. The response schema dictates the textual 
structure of the response. This schema is filled by 
stepping through it and matching its predicates against 
the relevant knowledge pool. A focusing mechanism is 
used to mediate between choices arising during this 
process. 
The schemas used by the ROMPER system are more 
complicated than those advocated by McKeown. In 
effect, ROMPER's schemas are responsible both for 
determining the textual shape of a response and for 
determining what information from the knowledge base 
to include in the response. Thus they are applicable to a 
much more restricted generation problem (e.g., re- 
sponding to a misconception of a particular type). 
Because of this, the test for determining which schema 
to use can be much more specific than the tests em- 
ployed by McKeown. 
7 HIGHLIGHTING AND OBJECT SIMILARITY 
We claim that in order for the above strategy for 
correcting misconceptions to work, the similarity metric 
that is used to assess object similarity must be affected 
by the preceding discourse. 
To date, most AI systems do not assess object 
similarity in a way that is context dependent. Several 
systems that do assess object similarity (Rumelhart and 
Abrahamson 1973, McKeown 1982, Carberry 1984, 
Weiner 1984) use a metric based on distance in some 
space. Most often, this space is the generalization 
hierarchy. Basically, two objects that have a common 
immediate superordinate (i.e., are siblings in the hier- 
archy) are seen as very similar, while objects whose 
lowest common ancestor is several levels up in the 
hierarchy are seen as quite different. 
One problem with this metric arises when objects can 
be classified in more than one way and there are several 
lowest common ancestors of the objects being com- 
pared. A decision must be made about which of these 
lowest common ancestors should be considered since 
the similarity assessment of the objects might vary 
widely as a result. For instance, a treasury bond and a 
corporate bond may be assessed as being very similar 
since they have a common immediate superordinate of 
bonds. On the other hand, both of these objects can be 
classified along other dimensions. If we look at a 
treasury bond as being a treasury issue, which is a type 
of US Government security, it seems quite different 
from a corporate bond, which is a corporate security. 
A second major problem with a similarity metric 
based on distance in the generalization hierarchy is that 
it is context invariant; contextual information has no 
way of affecting the assessments. As shown by Tversky 
(1977) and others, human judgments of object similarity 
have been found to shift both when the set of objects 
under discussion are altered (e.g., a violin and an 
electric guitar may be judged quite similar when in a 
group with a clarinet and an oboe, and may be judged 
quite different when the other members of the group are 
Computational Linguistics, Volume 14, Number 3, September 1988 57 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
a cello and an electric bass), and when the salience of 
attributes are altered (e.g., in a group containing a red 
triangle, a blue triangle, and a red square, the red 
triangle might be judged similar to the blue triangle 
when attribute shape is stressed, but may be judged 
similar to the red square when attribute color is 
stressed). 
One metric that avoids these problems was intro- 
duced by Tversky (1977). Tversky's metric, rather than 
relying on distance in some space, is based on the 
common and disjoint features of the objects involved. 
The metric, termed a contrast model, allows context to 
be taken into account in several places. 
Suppose we have two objects a and b where A is the 
set of properties associated with object a and B is the 
set of properties associated with object b. Tversky's 
measure can be expressed as: 
s(a,b) = Of(A fq B) ~ af(A -- B) ~ /3f(B ~ A) 
for some 0, a, and/3 -> 0. 
In the above equation 0, a, and /3 are parameters 
which represent the importance of each piece of the 
equation. The function f maps over the features and 
yields a salience rating for each. In essence, the contrast 
model states that the similarity of two objects is some 
function of their common features minus some function 
of their disjoint features. The importance of each par- 
ticular feature involved (determined by the function j') 
and the importance of each piece of the equation 
(determined by 0, a, and/3) may change with context. 
Although Tversky discusses in general terms how 
these functions might be set, he gives no concrete 
methods for doing so. For instance to set 0, a, and/3 he 
turns to the relative prominence of objects a and b in the 
discourse. The more prominent an object is, the more 
its attributes should have an impact on the similarity 
rating. Thus, finding the relative prominence of objects 
a and b in the discourse would help set these values. If 
a is relatively more important, then functions 0 and a 
should be greater than/3 resulting in the attributes of the 
more prominent object having a greater influence over 
the similarity assessment. While I would conjecture that 
information about the focus of the discourse (Grosz 
1981, Sidner 1983, Grosz, Joshi, and Weinstein 1983) 
might give an indication of an object's prominence and 
would therefore be useful in setting the values of 0, a, 
and/3, in this work I have assumed a value of 1 for the 
0, a, and 13 and have concentrated on setting the f 
function. 
We turn, then, to the problem of finding a value for 
theffunction: the measure of salience for each property 
of the objects involved. Other work, such as Carbonnell 
and Collins (1970) and Weiner (1984), has hand-encoded 
salience values for attributes of individual objects di- 
rectly into the knowledge base, permanently setting the 
ffunction. This approach is not sufficient for setting the 
f function for Tversky's metric, since it is crucial that 
theffunction be able to change with context. In order to 
make this happen, the salience values computed by f 
must change with context. 
To see this, consider our ability to explicitly mention 
an attribute to increase its salience. In the example 
earlier with the red and blue triangles and the red 
square, if the request for a similarity judgment had been 
preceded by "Look at the pretty colors of these 
objects;", the red triangle and red square would have 
probably been judged to be more similar than the red 
triangle and the blue triangle. Thus we see that explicit 
mention of an attribute in a discourse is one way that the 
f function might be affected by previous discourse. 
Explicit mention of an attribute is only one way in 
which the salience of the attributes may change dynam- 
ically. Another aspect of dynamic salience comes from 
the point of view, or perspective, applied to the domain. 
For instance, a building can be referred to as being an 
architectural work, for example, or as being someone's 
home. The two different views of the building cause 
different sets of attributes to become salient. Notice 
that this set of attributes is in addition to the attributes 
that have been explicitly mentioned in the discourse. 
From an architectural work point of view, attributes like 
the architect's name, date of building, and particular 
architectural features become salient. On the other 
hand, from the home point of view, attributes like the 
kitchen size, number of bedrooms, and living space 
become important. If we can find a way of modeling 
how these "precompiled" sets of attributes become 
highlighted in a discourse, we will have a principled 
method for setting the f function needed for Tversky's 
similarity metric. The next section discusses how this 
highlighting can be modeled by a computer system. 
8 OBJECT PERSPECTIVE 
The notion of point of view or object perspective has 
been noted by other researchers in artificial intelligence. 
Perspective's ability to explain the changing attribute 
salience has been attributed to a limited inheritance 
mechanism (see, e.g., Grosz 1977; Bobrow and Wino- 
grad 1977; Tou et al. 1982). An object viewed from a 
particular perspective is seen as having one particular 
superordinate, although in fact it may have many. The 
object inherits properties only from the superordinate in 
perspective. Therefore different perspectives on the 
same object cause different properties to be inherited 
(and therefore highlighted). 
While explaining object perspective via a limited 
inheritance mechanism is intuitively appealing, it is 
unable to handle some effects which intuitively should 
be handled by object perspective. The first has to do 
with the availability of object attributes. A limited 
inheritance mechanism makes attributes inherited from 
superordinates other than the one in perspective un- 
available. This seems a bit too strong. When we discuss 
a building as an architectural work, I may comment on 
the number of bedrooms the building has. While you 
58 ComputatJional Linguistics, Volume 14, Number 3, September 1988 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
may think my comment irrelevant to the current con- 
versation, you would still be able to understand it and 
even evaluate its truth or falsity. In the limited inheri- 
tance account of object perspective, however, this 
would not be possible. As far as the system would be 
concerned the number of bedrooms would not even be 
an attribute to the building. 
A second problem with the limited inheritance ac- 
count of object perspective has to do with deciding what 
attributes the superordinate in perspective has. The 
superordinate itself may have multiple classifications 
and thus potentially multiple perspectives. Thus, in 
order to figure out what attributes a particular concept 
should inherit, we must figure out not only what per- 
spective it is being viewed from, but also what perspec- 
tive the perspective superordinate is being viewed from, 
and so on. But this seems to be much more work than is 
necessary. 
Another problem is that a limited inheritance mech- 
anism explains the perspective for a single object only. 
However, during the course of a conversation it is 
usually the case that more than one object will be 
discussed. When this happens, usually the same kinds 
of things are discussed about the objects. In essence, a 
particular highlighting of attributes (or point of view) 
seems to be in force during the conversation. Yet, this 
highlighting is applied to different objects--some of 
which may not even have the same superordinates. 
What seems to be happening is that the conversational 
partners are viewing an entire group of objects from the 
same perspective. A limited inheritance mechanism 
cannot account for this unless each of the objects under 
discussion can be said to have the same (immediate) 
superordinate. 
A final effect that is not accounted for by the limited 
inheritance mechanism, yet seems to hinge on the view 
being taken on the domain, has to do with the height- 
ened importance of some objects during a discourse. 
Like the importance of attributes, the relative impor- 
tance of some objects in the discourse cannot solely be 
accounted for by explicit mention. Some objects are 
more likely to be mentioned and discussed in a dis- 
course than others. For instance, when discussing a 
particular building as an architectural work, I might 
reasonably mention the library down the street that was 
designed by the same architect. On the other hand, I 
will probably not mention my apartment. Along the 
same lines, in discussing that building as a home, my 
apartment is a likely candidate for mention in the 
conversation. Although this effect seems to be in some 
way tied to the notion of object perspective, the limited 
inheritance mechanism does not address this issue. 
We want to retain the dynamic highlighting of "pre- 
compiled" groups of attributes. Instead of the limited 
inheritance mechanism, we propose that the following 
account be used: 
1. Instead of tying perspective into the generaliza- 
tion hierarchy of objects as has been done in the 
past, the new notion of perspective is independent 
of that hierarchy. Perspectives that can be taken 
on the objects in the domain will be defined and 
will sit in a structure that is orthogonal to the 
generalization hierarchy. 
2. A number of perspectives are available for any 
domain of discourse and any given domain object 
may be viewed from any one of several perspec- 
tives for that domain. 
3. Each perspective comprises a set of attributes 
with associated salience values. It is these sa- 
lience values that dictate which attributes are 
highlighted and which are suppressed. 
4. One such perspective is designated active at any 
particular point in the discourse) 
Our solution is that any object that is accessed by the 
system is viewed through the current active perspec- 
tive. However, instead of dictating which attributes an 
object inherits, the active perspective affects the sa- 
lience values of the attributes that an object possesses 
(either directly or inherited through the generalization 
hierarchy). The active perspective essentially acts as a 
filter on an object's attributes. By raising the salience of 
the attributes, it highlights those attributes which have a 
high salience rating in the active perspective. By low- 
ering the salience of the attributes, it suppresses those 
attributes that are either given a low salience value or do 
not appear in the active perspective. 
By defining object perspective in this way, we have 
retained the desirable results of the limited inheritance 
account of object perspective while avoiding its prob- 
lems. In addition, since any object accessed by the 
system is viewed through the active perspective, we 
gain the feeling of perspective on the entire domain. The 
object importance aspect of perspective is gotten by 
saying that those objects that contribute attributes 
which are highly salient to a perspective are important 
while that perspective is active. 
We propose that theffunction in Tversky's metric be 
set by taking into account the salience values derived 
from the active perspective. This would yield an f 
function that is context dependent and would help the 
similarity metric exhibit many desirable properties. 
9 MODELING A DOMAIN WITH PERSPECTIVES 
In some natural language systems, a model of a partic- 
ular domain includes a usual object taxonomy contain- 
ing all of the objects in the domain and all of the 
attributes associated with those objects. We will show 
an example of building a domain model with perspec- 
tives. In order to do this, one must build the domain 
model as usual. In addition, the perspectives that can be 
taken on the domain objects must be defined. The result 
of viewing the domain model through the perspectives 
will be shown. 
Computational Linguistics, Volume 14, Number 3, September 1988 59 
Katldeen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
Money Market Certificates 
Maturity: 3 months 
Denominations: $1,000 
Issuer: Commercial Bank 
Penalty for Early Withdrawal: 10% 
Purchase Place: Commercial Bank 
Safety: Medium 
Treasury Bills 
Maturity: 3 months 
Denominations: $1,000 
Issuer: US Government 
Purchase Place: Federal Reserve 
Safety: High 
Savings Instruments 
Maturitywl.0 
denominations--1.0 
safety---0.5 
yield---0.5 
Issuing Company 
issuer--l.0 
safety--1.0 
purchase-place--0.5 
yield--4).5 
tax--0.5 
Figure 2. Sample Perspectives. 
Treasury Bond 
Maturity: 7 years 
Denominations: $500 
Issuer: US Government 
Penalty for Early Withdrawal: 20% 
Purchase Place: Federal Reserve 
Safety: High 
Figure 1. Objects in "Flat" Domain Model. 
Figure I shows a small piece of a typical domain 
model. The domain is that of financial securities. Three 
of the objects from this domain, Money Market Certif- 
icates, Treasury Bills, and Treasury Bonds, are shown 
with the attributes they possess. In systems as they are 
defined today, a group of objects defined in this way and 
arranged in a generalization hierarchy would constitute 
the domain model. Ifa system were to access any one of 
the objects in the domain it would be given the object 
with the attributes as listed in the figure. 
In order to get a dynamic highlighting of the domain 
model, we must build, in addition to the object taxon- 
omy, a separate structure containing the perspectives 
that can be taken on the domain objects. This means 
that we must think about the different points of view 
that can be taken on the objects in the domain and 
compile sets of attributes from our model which capture 
the important domain concepts in that point of view. 
There will be attributes in each perspective that do not 
occur with all of the objects in the domain. At the same 
time, there will be attributes of individual objects that 
do not appear in a particular perspective. The perspec- 
tive simply captures the attributes that are important in 
a particular point of view. 
Figure 2 contains two perspectives that might be 
reasonable for this domain (here we are assuming 
salience values from low salience of 0 to high salience of 
1). The perspective of Savings Instruments highlights 
maturity and denominations, and somewhat highlights 
safety and yield. This indicates that when people are 
discussing securities as savings instruments, they are 
most interested in how long their money will be tied up 
and in what denominations they can save their money. 
The perspective of Issuing Company, on the other hand, 
highlights different attributes. When securities are dis- 
cussed from this perspective, things like the name of the 
company and the stability of an investment in the 
company become important. Other attributes of the 
securities are ignored (recall that attributes not men- 
tioned in the perspective get assigned a low salience 
rating). 
Notice how the objects look when accessed, depend- 
ing on which of the two different perspectives are 
active. For instance, through the savings instruments 
perspective the objects' attributes take on the salience 
values shown in Figure 3. Attributes are not shown in 
the figure that have 0 salience. When no static salience 
is inchlded in the domain model, the attributes of the 
objects derive their salience directly from values given 
in the active perspective. Attributes in the perspective 
that the objects do not have (e.g., yield) are ignored. 
Attributes of the objects not occurring in the perspec- 
tive are given the lowest salience rating. The very same 
objects look different when viewed through the other 
perspective. This is shown in Figure 4. 
The salience values derived from perspective can be 
used for many tasks. Section 11 will show how they can 
be used in conjunction with Tversky's similarity metric 
to generate context sensitive responses to misconcep- 
tions. 
10 CHOOSING THE ACTIVE PERSPECTIVE 
In order for the notion of object perspective to be truly 
beneficial, there must be a mechanism for choosing the 
active perspective based on previous discourse. While 
this topic is still very much open to investigation, some 
preliminary research has revealed several factors that 
might influence the choice of active perspective. 
Perhaps one of the most influential pieces of infor- 
mation useful in choosing a perspective is the user's 
current goal. In (McKeown, Wish, and Matthews 1985) 
the user's goal completely determines which perspec- 
tive is active. In their work each perspective that can be 
taken on the domain objects is indexed by potential 
goals. Thus once the system has determined what the 
user's goal probably is, it has also determined what 
60 Computational Linguistics, Volume 14, Number 3, September 1988 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
Money Market Certificartes 
Maturity: 3 months---1 
Denominations: $1,000~1 
Safety: Medium---0.5 
Treasury Bills 
Maturity: 3 months--1 
Denominations: $1,000--1 
Safety: High---0.5 
Treasury Bond 
Maturity: 7 years--1 
Denominations: $500---1 
Safety: High--0.5 
Figure 3. Objects Through Savings Instruments 
Perspective. 
perspective the user has probably taken on the domain 
objects. 
While the user's goal is a good source of information 
to use to determine the probable perspective, other 
factors may also influence this choice. These include 
the attributes and objects mentioned so far in the dialog. 
The mentioned attributes are obviously thought to be 
important and one would therefore expect them to be 
given a fairly high salience rating in the active perspec- 
tive. Thus the choice of active perspective can be 
narrowed down to those in which the mentioned at- 
tributes appear with high salience. 
By the same token, the objects mentioned so far in 
the dialog can also give a clue concerning the active 
perspective. One would expect that the active perspec- 
tive would deem these objects important. Therefore the 
system might look for perspectives that give high sa- 
lience ratings to many of the attributes associated with 
objects that have been mentioned in the discourse. 
In this section I have identified several factors that 
influence the choice of active perspective. While the 
success of other systems (McKeown, Wish, and Mat- 
thews 1985) has shown that a reasonable choice of 
Monrey Market Certificates 
Issuer: Commercial Bank--I 
Purchase Place: Commercial Bank--0.5 
Safety: Medium--I 
Treasury Bills 
Issuer: US Government--I 
Purchase Place: Federal Reserve---0.5 
Safety: High--1 
Treasury Bond 
Issuer: US Government--1 
Purchase Place: Federal Reserve---0.5 
Safety: High--1 
Figure 4. Objects Through Issuing Company Perspective. 
perspective can be made based on discourse goals, the 
nature of establishing and perhaps "shifting" perspec- 
tive during a discourse must still be investigated. Still 
unanswered are questions such as: When does a per- 
spective change? How long is a perspective active? Is 
there a relationship between a discourse unit (Grosz and 
Sidner 1985) and perspective? Is there any structure to 
the space of perspectives that would put constraints on 
moving from one active perspective to another? These 
questions must be taken up in future research on 
perspective. 
ll PERSPECTIVE'S INFLUENCE ON RESPONSES 
In this section we will show how using the active 
perspective to highlight the user model, the misconcep- 
tion correcting algorithm from Section 6, and the 
Tversky similarity metric can account for context sen- 
sitive corrections to misconceptions. Recall that in 
correcting a misattribution one of the correction sche- 
mas used by ROMPER called for a similar object to be 
offered as a possible object of confusion. A study of 
transcripts reveals, however, that this schema may be 
instantiated in different ways depending on the context. 
Consider once again the following dialogs that were first 
seen in the introduction. 
U: I am interested in investing in some securities to 
use as savings instruments. I want something 
short-term and I don't have a lot of money to 
invest so the instrument must have small denom- 
inations. I am a bit concerned about the penalties 
for early withdrawal. What is the penalty on a 
T-Bill? 
R: T-Bills don't have a penalty. Were you thinking 
of Money Market Certificates? 
In this case money market certificates were seen as 
being similar to T-bills and therefore included in the 
response. A different object might be used in a different 
context. Consider: 
U: I am interested in investing in some securities. 
Safety is very important to me, so I would 
probably like to get something from the govern- 
ment. I am a bit concerned about the penalties for 
early withdrawal. What is the penalty on a T-Bill? 
R: T-Bills don't have a penalty. Were you thinking 
of T-Bonds? 
The difference in these two responses can be explained 
by different perspectives being taken on the objects. 
Suppose we are given the objects, attributes, and per- 
spectives from Section 9. The dialog preceding the first 
example could lead to the establishment of savings 
instruments as the active perspective 4 since it mentions 
the perspective by name and explicitly mentions several 
of the attributes made important by the perspective. 
If ROMPER were given this information it would 
Computational Linguistics, Volume 14, Number 3, September 1988 61 
Kathleen F. McCoy Reasoning on a Highlighted User Model to Respond to Misconceptions 
proceed by attempting to instantiate the wrong object 
schema described in section 5. Recall that this schema is 
applicable when there is a similar object that has the 
property involved in the misconception. The system 
would collect all objects having the attribute in question 
and then test their similarity with the object involved in 
the misconception. In our knowledge base there are two 
objects that have the attribute involved in the miscon- 
ception: Money Market Certificates and T-Bonds. 
Assume that theffunction in the metric is set solely 
on the basis of the salience values given by the perspec- 
tive. Also assume that we have decided that two objects 
are highly similar if Tversky's metric returns a number 
greater than 0, and not similar otherwise. Applying the 
Tversky metric using the salience values attached by 
the savings instrument perspective (and assuming a 
value of 1 for 0, a, and/3) we get: 
s(T-BiU, MM-Cert) -- f(maturity, dehorn) - f(safety) 
= 2 - .5 = 1.5 ===> high similarity 
s(T-Bill, T-Bond) = f(safety) - f(maturity, denom) 
= .5 - 2 = -1.5 ===> low similarity 
With these calculations the system would choose the 
Money Market Certificate as the possible object of 
confusion and respond: 
R: Treasury Bills don't have a penalty. Were you 
thinking of Money Market Certificates? 
Contrast the above calculations with calculations that 
might occur given a different active perspective. The 
discourse preceding the misconception utterance in the 
second example suggests the active perspective of "Is- 
suing Company". Using the salience values attached by 
this perspective the similarity metric would produce the 
following calculations: 
s(T-Bill, MM-Cert) 
= f0 - f(issuer, safety, purchase) 
= 0 - 2.5 = -2.5 ===> low similarity 
s(T-BiU, T-Bond) 
= f(issuer, safety, purchase) -f0 
= 2.5 - 0 = 2.5 ===> high similarity 
In this case a reasonable response by the system would 
be: 
R: Treasury Bills don't have a penalty. Were you 
thinking of a Treasury Bond? 
As the examples show, changes in the active perspec- 
tive can account for the same misconception being 
responded to in two different ways. 
12 CONCLUSIONS 
This paper has described a new method for responding 
to misconceptions that relies on an analysis of a high- 
lighted user model to generate a response that is most 
likely to benefit the user. A number of strategies were 
abstracted from a study of transcripts. Each strategy 
was associated with a distinguished structure in the user 
model which could explain its use in a given situation. A 
system can use this pairing to decide how to respond by 
looking in the highlighted user model for evidence of 
one of the distinguished structures. The corresponding 
strategy can be used to respond. 
In addition, the paper has described a new notion of 
object perspective that is able to model one aspect of 
the dynamic highlighting of the user model due to 
previous discourse. It was shown how perspective 
could account for certain contextual affects on re- 
sponses to misconceptions. 
ACKNOWLEDGEMENTS 
I would like to thank Sandee Carberry, Julia Hirschberg, Aravind 
Joshi, Martha Pollack, Bonnie Webber, and the participants of the 
User Model Workshop for their helpful comments and discussions on 
various aspects of this work. Many thanks also to the anonymous 
reviewers for their constructive input concerning the style and content 
of this paper. 
Some of this work was done while the author was at the University 
of Pennsylvania and was partially supported by the ARO grant 
DAA20-84-K-0061 and the NFS grant MCS81-07290. 
NOTES 
1. While in general it may be necessary for the user model to contain 
more information about the user (e.g., goals and plans), these are 
not required for the restricted task set out in this paper. 
2. The dialogs contained in this paper were not taken directly from 
the transcripts. They are used to illustrate the kinds of responses 
found. 
3. Saying that exactly one perspective is active is actually a simpli- 
fication. It may be the case that a number of perspectives can be 
active. In this case the resulting "active perspective" will be 
some function of the individual active perspectives. The exact 
nature of this function is an open research question. 
4. Note that the ROMPER system does not choose the active 
perspective--it is given as input to the system. This example is 
simply used to illustrate perspective's influence on misconception 
responses. 

REFERENCES 
Bobrow, D.G. and Winograd, T. 1977 An overview of krl, a knowl- 
edge representation language. Cognitive Science 1(1): 3-46. 
Brown, J.S. and Burton, R.R. 1978 Diagnostic models for procedural 
bugs in basic mathematical skills. Cognitive Science 2(2): 155-192, 
1978. 
Carberry. Sandra M. 1984 Understanding pragmatically ill-formed 
input. In lOth International Conference on Computational Lin- 
guistics and 22nd Annual Meeting of the Association of Compu- 
tational Linguistics, Stanford University, CA: 200-206. 
Carbonnell, Jaime R. and Collins, Allan M. 1970 Mixed-Initiative 
Systems For Training and Decision-Aid Applications. Technical 
Report ESD-TR-70-373, Electronics Systems Division, Laurence 
G. Hanscom Field, U.S. Air Force, Bedford, MA. 
Grosz, B. 1977 The Representation and Use of Focus in Dialog 
Understanding. Technical Report 151, SRI International, Menlo 
Park, CA, 
Grosz, B.; Joshi, A.K.; and Weinstein S. 1983 Providing a unified 
account of definite noun phrases in discourse. In Proceedings of 
the 21st Annual Meeting of the Association for Computational 
Linguistics, Cambridge, MA: 44-50. 
Grosz, B. 1981 Focusing and description in natural language dialog. In 
Webber B, Joshi, A, and Sag, I. (ed.), Elements of Discourse 
Understanding, Cambridge University Press, Cambridge, En- 
gland: 85-105. 
Grosz, B. and Sidner, C. 1985 Discourse structure and the proper 
treatment of interruptions. In Proceedings of the 1985 Joint 
Conference on Artificial Intelligence, IJCAI85, Los Angeles, CA. 
Kaplan, S.J. 1979 Cooperative Responses From a Portable Natural 
Language Database Query System. Ph.D. thesis, University of 
Pennsylvania, Philadelphia, PA. 
Karlin, Robin 1985 Romper Mumbles. Technical Report, University 
of Pennsylvania, Philadelphia, PA. 
Malhotra, A. 1975 Design Criteria for a Knowledge-Based English 
Language System for Management. Technical Report TR-146, 
Project MAC, MIT, Cambridge, MA. 
Malhotra, A. and Sheridan, P. 1976 Experimental Determination of 
Design Requirements for a Program Explanation System. Tech- 
nical Report RC 5831, IBM Research Center, Yorktown Heights, 
NY. 
Mays, E. 1980 Correcting misconceptions about database structure. 
In Proceedings of the 3rd CSCS1 Biennial Meeting, Victoria, BC, 
Canada. 
McDonald, D.D. 1980 Natural Language Production as a Process of 
Decision Making Under Constraint. Ph.D. thesis, MIT, Cam- 
bridge, MA. 
McKeown, K. 1982 Generating Natural Language Text in Response 
to Questions About Database Structure. Ph.D. thesis, University 
of Pennsylvania, Philadelphia, PA. 
McKeown, K.; Wish, M.; and Matthews, K. 1985 Tailoring explana- 
tions for the user. In Proceedings of the 1985 International Joint 
Conference on Artificial Intelligence, Los Angeles, CA. 
Pollack, M.; Hirschberg, J.; and Webber, B. 1982 User participation 
in the reasoning processes of expert systems. In Proceedings of 
the 1982 National Conference on Artificial Intelligence, AAAI, 
Pittsburgh, PA. 
Rumelhart, D.E. and Abrahamson, A.A. 1973 A model for analogical 
reasoning. Cognitive Psychology 5: 1-28. 
Schuster, E. 1982 Explaining and Expounding. Technical Report 
MS-CIS-82-49, University of Pennsylvania, Philadelphia, PA. 
Sidner, C.L. 1983 Focusing in the comprehension of definite ana- 
phora. In Brady, Michael and Berwick, Robert (eds.), Computa- 
tional Models of Discourse, MIT Press, Cambridge, MA: 267-330. 
Sleeman, D. 1982 Inferring (mal) rules from pupil's protocols. In 
Proceedings ofECAI-82, Orsay, France: 160-164. 
Stevens, A.L. and Collins, A. 1980 Multiple conceptual models of a 
complex system. In Federico, Pat-Anthony; Snow, Richard E.; 
and Montague, William E. (eds.), Aptitude, Learning, andlnstruc- 
tion, Erlbaum, Hillsdale, NJ: 177-197. 
Stevens, A.; Collins, A.; and Goldin, S.E. 1979 Misconceptions in 
student's understanding. International Journal of Man-Machine 
Studies 11: 145-156. 
Tou, F.; Williams, M.; Fikes, R.; Henderson, A.; and Malone, T. 
1982 Rabbit: an intelligent database assistant. In Proceedings of 
AAAl-82, Carnegie-Mellon University, Pittsburgh, PA: 314-317. 
Tversky, A. 1977 Features of similarity. Psychological Review 84: 
327-352. 
Weiner, E. Judith 1984 A knowledge representation approach to 
understanding metaphors. Computational Linguistics 19(1):1-14. 
Woolf, Beverly P. 1984 Context Dependent Planning in a Machine 
Tutor. Ph.D. thesis, University of Massachusetts, Amherst, MA. 
Woolf, B. and McDonald, D. 1983 Human-computer discourse in the 
design of a pascal tutor. In Janda, Ann (ed.), CHl'83 Conference 
Proceedings--Human Factors in Computing Systems, Boston, 
MA: 230-234. 
