Of Frames, Scripts, and Stories
by
Choong Huei Seow
Submitted to the Department of Electrical Engineering and Computer Science
in Partial Fulfillment of
the Requirements for the Degree of
Bachelor of Science in Computer Science and Engineering
at the
Massachusetts Institute of Technology
May 1990
Signature of.
Certified by
Accepted by
Aut
@ Choong Huei Seow 1990
The author hereby grants to MIT permission to reproduce and to
distribute copies of this thesis document in ,whole or in part.
Signature redacted
hor
/ Department of El rical Engineering and Computer Science
I 'May 21st, 1990
Signature redacted
Marvin L. Minsky
Donner Professor of Science
Thesis Supervisor
Signature redacted
Leonard A. Gould
Chairman, Department Committee on Undergraduate Theses
MASSACHUSETTS INSTITUTEOF TEChN01lOCY
SEP 18 1990
LIBRARIES
ARCHIVES
Of Frames, Scripts, and Stories
by
Choong Huei Seow
Submitted to the Department of Electrical Engineering and Computer Science
on May 21st, 1990
in partial fulfillment of the requirements for the degree of
Bachelor of Science in Computer Science and Engineering
Abstract
An essential part of research in Artificial Intelligence (AI) involves the development of
systems which can exhibit common sense and understanding. Such systems are proposed
to be built upon platforms containing rich and different domains of everyday knowledge.
This paper discusses the issues which dominate the framework of such systems. Such
issues include the frame problem, the recognition of patterns and analogies, organization
and usage of knowledge domains, etc.
Thesis Supervisor: Marvin L. Minsky
Title: Donner Professor of Science
2
Acknowledgements
I would like to express my sincere thanks to my thesis supervisor, Marvin
Minsky. The opportunity to work with Marvin has alway been very excit-
ing, fun, and stimulating. I wish to also thank all my past UROP project
supervisors, Dr. Andy Boughton, Professor Arvind, Professor Tommy Pog-
gio, and Professor Steven Pinker, for the breadth and variety of perspectives
and good discussions. My thanks to the members of the Learning and Epis-
temology Group of the MIT Media Laboratory, for initiating my curiosity
in how we learn and think. To the MIT Artificial Intelligence Laboratory
and all its members, thank you for the unusual, informal, and great research
environment.
To my family, my heartfelt thanks for their continuous support, love, and
encouragement. To Anne and Bob Berg, thank you for your love and support,
and for your kindess of inviting me to experience life as part of an American
family. My friends, thank you for being there when I needed it most.
Chapter 1
Introduction
1.1 Intelligence
What is intelligence ? The dictionary definition of intelligence is the capacity to appre-
hend facts and propositions and their relations, and to reason about them. The human
ability to exhibit intelligence has been an area of scientific interest, in particular, scien-
tists have been studying the implementation of intelligent "behavior" in artificial systems,
namely computers.
1.2 Motivation
A major part of a human's daily life involves the role of understanding. Our ability to
grasp new information, learn new concepts and ideas, are all essential parts of under-
standing. It seems that intelligence can only exist if we have the ability to understand.
We learn new knowledge when we understand the information which is presented to us.
At this point, we still do not know how the process of understanding really takes place.
However, research related to the process of understanding has been carried out in the
field of Artificial Intelligence in the past such as the work which have been carried out
on story-understanding models [Charniak, 1972][Wilensky, 1978], concept-learning pro-
grams [Winston, 1970], etc.
3
1.3 Al According to Marr
According to Marr, the implementation of a method to solve a particular information
processing problem should be left to be the final stage of the project [Marr, 76]. One
should begin with first identifying the information processing problem, and formulate an
abstract method for solving the problem. This method can be decomposed and described
by the following
* Algorithms which will be needed by the method.
* Functional descriptions of the algorithms.
* Information which is captured by this method.
Marr describes the above method as a Type I theory. According to him however, one of
the principal difficulties of the study of A.I. is that one can never be sure if the problem
which is to be solved has a theory of this type. He describes most A.I. programs which
have beeen carried out are of the Type II theory. While a program of this theory type
solves the proposed problem, it often performs the solution in a very clumsy and fragile
manner. Thus, any candidate with a Type Iltheory will attach greater importance to the
performance of the program. As Marr explains further, even though a Type IItheory can
be successfullly implemented, one cannot still assume that a simpler underlying method
does not exist. Marr states that a large AI system which has been built without a
Type I theory would not be considered substantial, thus this system will not have any
importance significance to AL.
This thesis discusses the underlying processes of understanding, in particular how
it works, and how it is used in intelligent systems. This thesis will focus on the issues
which need to be addressed in the design of systems which organize large and different
domains of knowledge, systems which exhibit "common-sense knowledge". Most of the
thesis will focus on the heuristics of understanding, and the development of a framework
for building such systems of knowledge.
4
1.4 The Role of Knowledge in Intelligence
Intelligence is being able to learn and understand.
1.4.1 Problems of Knowledge Representation
The issue of modeling knowledge is tricky. Workers in the field of A.I. have been trying
to build models to represent knowledge but have had limited success. A fairly common
problem is the amount and variety of knowledge we are most likely to use in the system,
how much is needed and what levels of detail are needed ? One might argue that we need
to represent explicitly, a large number of assertions to correctly represent an event or a
concept. But, just how much is needed ? Do we want to be bogged down with details
of knowledge which is not relevant to us at the time where we reference this knowledge
structure? Also, if you make a comparison between AI models and the human memory
model, you will notice that the human memory model poses a constraint on the amount
of processing capability. We cannot recall everything at one time, but yet, we are able
to recollect relevant information.
1.4.2 Useful Knowledge is Important
A good model of knowledge representation should represent knowledge which is useful
to the problem-solver in an explicit manner. However, most models of knowledge repre-
sentational which have been developed have a fairly limited domain of problem-solving.
Once you move out slightly from this domain, the whole system becomes very fragile.
How do they compare this to the human "model" of representing knowledge? We learn
new things every moment, concepts, experiences, skills, etc. The things which we learn
are very different, do we really have a single model of knowledge representation ? Does
this indicate that we do not possess a single model of knowledge representation? Or is
one model of knowledge representation sufficient for performing all cognitive skills which
we have ?
In comparison to human cognition, we seem to use the many different domains of
knowledge that we have in a wide variety of applications. At times, humans seem to
5
be able to make "connections or "links" to events which by themselves seem to be very
different in content. The main question arises, and remains an important one, "How
do we organize and integrate these domains of knowledge?" The following section will
describe an area of cognition which is in need of a theory of knowledge integration.
1.5 Connectionist Models
1.5.1 Perceptrons and PDP
The field of artificial intelligence has experienced many different forms of growth and
change. In the 1950's the theory of perceptrons and connectionist models gained popu-
larity among the workers in this field, and "suffered" a great setback upon a detailed
study carried out by Minsky and Papert [Minsky, 68] which described the computational
limitations of perceptrons. The revival of the popularity of connectionist models was
initiated by the work of Rumerlhart and McClelland [McClelland, 86] who successfully
described and implemented a connectionist model consisting of multiple layers of neu-
ral units which was "capable" of generating past tense forms of English language verbs.
While this connectionist model did show forms of performance which were comparable
to human ability, one should treat the success of the implementation of this connection-
ist model with caution. Although such a network model may actually work for a very
particular cognitive skill, it does not neccessarily imply that these forms of cognitive
architecture will be successful for other domains of cognition. I believe that scientists
should carefully evaluate the potential and functional extents of neural architectures be-
fore setting out to build neural networks to perform various cognitive skills. We have
yet to learn in detail the true dynamics of neural networks (their detailed behaviour is
still unknown), yet a a considerable number of researchers in the field are setting out to
build more complex and larger neural network architectures to perform more complicated
cognitive skills. In short, while the work carried out by McClelland and Rumelhart on
multi-layed neural networks has brought new insight and scope to revitalize the field of
connectionism once again, researchers have failed to recognize that a careful evaluation
6
of the prospects of neural networks should be carried out before building more complex
forms of the networks.
Some of the limitations of connectionist architectures currently being developed is
that they have cannot effectively store past states of the network, such that they could
be automatically retrieved later for cognitive usage. Can we then safely conclude that if
connectionist models of cognitive processes were to be successful, they would either be
implemented in a system where other models of cognitive architecture are present too ? I
posed this question since that if connectionist models are poor at maintaining knowledge
of past states, it thus cannot be a good model for the human memory "system" as a
standalone.
1.6 What's Holding Back Computational Vision
The ideas of Marr described in the previous section were greatly influenced by his pi-
oneering work in formulating a framework for solving the vision problem. Marr was
convinced that a mathematical model of vision was correct since there had been success-
ful implementations of early vision processes, such as Horn's shape-fMm-shading theory
for object reconstruction, etc.
At this point, we do not still understand how human object recognition works. Com-
putational vision workers argue that it is composed of various "early vision" processes
such as edge-detection, stereopsis, color vision and segmentation, etc. While most of the
"early vision" modules have been successfully implemented, there still remains the dif-
ficult and complicated problem of integrating these processes into a model of high-level
visual recognition. Have we progressed in the domain of object recognition from the
implementations of "early vision" modules ? I must admit that I have the tendency to
solve problems using a "bottom-up" design methodology. This design methodology is
sound if the overall design structure of the entire high-level system has been carefully
analyzed and the low level modules implemented according to the "global" high-level
design. Instead, vision scientists have implemented "early low-level" vision modules and
are currently working their way upwards to solve the problem of intermediate and high-
7
high-level OBJECT RECOGNITION
vision AND PERCEPTION
Intermediate behavour
early vision Shape from
processes Color vision Stereopsis shaefo
shading
Figure 1.1: Segmentation of the Vision Tree
level vision processes. I would not be entirely surprised if the future system of a high
level visual process such as visual recognition is not entirely dependent upon these "early
vision" modules. While the study and implementation of such processes are important,
it should be done in the manner and perspective of how such processes combine, interact
and contribute to more abstract and complex processes. Consider Figure 1.1 as a rough
representation of the current situation that current workers in vision are facing. While
there has been successful implementations of such "early vision" modules, there still re-
mains the open question of integrating these vision modules to construct higher-level
vision theories such as object recognition and perception. The main problem which I
believe is plaguing further development is due to the fact that the workers are using a
mathematically based, computational approach to develop a theory of object recognition.
While mathematical theories has been successfully used to solve early vision problems,
I do not believe that they can hold as the very basis for the development of high-level
vision processes.
1.6.1 Object Recognition Needs Knowledge
There have been numerous cases where AI scientists who began working on developing
a model or a system which "successfully" imitated a particular cognitive function. Some
of the projects were mainly motivated by an interest to solve particular problem, for
example "How do humans recognize objects ?" So we set out on a project, trying to
solve the problem from an engineer's perspective (as a large number of workers in AI
8
Process B
Teacher Process A Student
Figure 1.2: Initial Learning Environment
have done). After a period of substantial effort, such a "system" is completed. The
system performs "quite well" implementing a simple human cognitive mechanism, for
example a system for navigating robots in an environment. The design ideas behind
the development of the system was carefully planned and implemented. However, the
problem of such a system is that its domain of application is very limited. How is the
implementation of such a system relevant to Artificial Intelligence ? Minsky argues that
the key essence of developing intelligent systems lie in the domain of understanding how
we organize domains of knowledge which we refer to as being "common-sense" knowledge.
1.7 Looking Further Ahead
Previously, it was discussed that a successful system would have to have a good model
of representing knowledge. However, such a model is not the only essential criterion
for the successful implementation of a intelligent system. It is also essential for such a
system to possess powerful learning mechanisms. Such learning mechanisms may be built
out of more primitive mechanisms of reasoning, differentiation, etc. Do infants initially
rely upon supervised learning to obtain new concepts and knowledge. The reliance on a
supervisor for the concept learning is then slowly decreased until the period where the
learner has understood the concept which was presented. Once this understanding has
taken place, this individual can apply the knowledge of this new concept whenever the
appropriate situation arises.
Refering to Figure 1.2, this diagram gives a general description of the interaction
between the teacher and the student. The student is initially assumed to have very
little or no background of the concept to be presented. The teacher is also assumed to
have the goal of wanting the student to understand and learn the new concept. This
9
process proceeds further with the teacher presenting the student with information which
is relevant to the concept. Such a presentation of information is consistent with the
viewpoint of a good knowledge representational model, which was discussed earlier as
being a essential part of Al systems. The teacher should explicitly show examples where
the concept is correct, and slight variations of it where the concept is almost right. This
propogation of information is referred to as Process A in Figure 1.2.
Such a methodology was followed by Winston's model for learning structural descrip-
tions, essentially by presenting models which were correct or which were near misses of
the correct examples [Winston, 1970]. While Winston's model was one of the successful
AI systems which were built, I still see a strong potential of such a model failing. For
instance, the order of which the examples are presented to the student observer is very
important, since different conclusions about the structure to be learned can be obtained
via different combination sequences of the same examples. Here, the sequence of informa-
tion input has caused the failure, not the representation of the structural examples. It is
very unlikely that the student learner can learn the correct structural description. From
the viewpoint of the teacher, the teacher must take the following heuristics for teaching
the student learner a new concept/idea
* What forms of knowledge does this student learner have ?
* What forms of reasoning mechanisms does it have ?
* How can I present examples of correct and near-misses of a concept in a fashion
such that the student learner can infer the correct representation of it?
Following Winston's model of learning structural descriptions from examples, the teacher
must present the student learner with the correct sequence/combination of examples
in such a manner that the learner system could correctly derive the intended model's
structural description. However, this heuristic discussed is not commonly experienced
with everyday world situations. The sequence of which we receive knowledge about
particular areas of the environment can be quite disorganized. However, we still have the
ability to reconstruct all the information needed to learn a new concept. What allows
10
System
Information
ADF BNDFGH 245R fsdf34B Structures
time
Figure 1.3: Streams of Information
us to do have this ability ? Figure 1.3 represents a more accurate situation of everyday
information processing.
You don't know what you're looking for most of the time. If you were, it would
be much easier to find a solution.
Obviously, this is quite a serious deficiency. How do we cope with such a situation when
information is flowing into the system in an incoherent manner. We shall now turn to
focus on several other AI models which were developed in the area of understanding.
1.7.1 Representations and Cognitive Skills
Humans use many different forms of information representations for facilitating our cog-
nitive skills. When we try to find a solution for a particular situational problem, such
as trying to find a path to go from Location A in State M to Location B in State N, we
often use a form of a roadmap (a spatial map) to facilitate us in finding a solution. Work
carried out by Kuipers [Kuipers, 1976] on spatial cognition describes different models of
representations for representing spatial knowledge. Compare the map's model of spatial
representation with an "identical" representation such as a complete written descrip-
tion of the same entire geographical area covered in the situation. While both forms of
representations can be semantically "identical", the latter form of knowledge representa-
tion does not serve its purpose in efficiently aiding human spatial cognitive skills. This
11
leads me to the following general direction. I believe that the constraints imposed by
human limitations in information processing capabilities caused us to develop models of
knowledge representations which provide "excellent performance" for performance com-
pensation. The forms of the knowledge representation models that we develop are most
likely to be influenced by the type of knowledge being represented. If you had followed
the example just given, the written form describing the spatial environment was not
suitable for usage in trying to arrange a journey between the two geographical locations.
It is quite possible that the nature of human knowledge representation models have
been influenced by the the functions and limitations of human cognitive skills.
1.8 Summary
I would like to end this introductory chapter with the intention of leaving the reader
with a sense of what the remaining sections of this thesis will be focusing on. First,
I have introduced certain aspects of AI which are considered to be key issues, such as
the problem of organizing knowledge, and the notion of understanding information. The
following sections of the thesis will remain focusing on the problem of organizing large
domains of everyday common-sense knowledge.
12
Chapter 2
Organization of Knowledge
2.1 Introduction
As described previously, there remains a problem of integrating the different domains
of knowledge that a system (and even humans) may possess into a body which is both
useful and effficient. These domains of knowledge are by comparison, extremely large,
since they contain all the facts facts and knowledge that the person has to date. While
there have been systems which have been developed in the past to contain domains of
knowledge, such systems are considered to be too specific in their applications, and their
domains of knowledge considered to be too small in comparison to a general description
of eveyday knowledge which is contained in individuals. A good solution to the process of
integrating knowledge of different domains will be benefitial to both builders of systems
which exhibit common-sense knowledge.
2.2 Recognizing Knowledge Structures
2.2.1 Abstraction of Events
Can we build a representation tree which describes the event Walking ? If we take a
not-too-short temporal subabstraction of this event, we will still get an event classified
as Walking. What will happen if I took a much smaller temporal slice of this event.
Would this small slice of this temporal event still be considered the same event, e.g. if I
13
were to show you a freeze frame of an event which displays a set of states, would it still
represent the same event. What cues and features do we use to describe an event? If the
freeze frame we had taken previously included only states which display an arm and a
leg moving, do we actually classify this set of states as being the event Walking?
Furthermore, there are certain classes of events which are very difficult to represent
and describe. For instance, how would one describe the event JohnlsDrivingA Car in a
set of assertions ? If we observe John, he could be experiencing a large number of smaller
events which are happening throughout the period of this event. I strongly believe that
such an event can only be described over a course of a certain period of time, i.e. there is a
minimal interval during which this event takes place over, and to consider that the event
actually happened, this interval should be completed. What follows is that when we take
a temporal slice during this interval, we do not consider that temporal slice as an event
of the same type also, but rather, we should assert that that event of JohnIsDrivingA Car
is in the midst of the process.
2.2.2 Important Properties of Events
There exist a large set of events which contain properties which will be called static. These
so-called static properties are the ones which are constant and are of little significance
or importance to the entire structure of the event. For instance, is the property John's
car is blue in color in the event JohnisDrivingA Car of any important significance to the
event ? For most properties like these which are part of the world which the event takes
place in, I would expect these properties to be static, assuming that event is the only
"event" dynamically taking place at that time. However, some of the properties which
are of significant importance to the event can themselves be static. Thus I came up with
the following questions which would probably the major design issues to be considered
while building this "mechanical" system.
* How does a system determine if a property of an event is of importance to the event
itself ?
- What would the representation look like then ?
14
* Can a particular property of a subevent which is considered to be of little impor-
tance to the subevent be of much greater importance to a "parent" event ? If so,
how would this be detected, and how would the status be changed ?
A system which exhibits intelligence will have to use various different domains of knowl-
edge that it has. However, a significant problem which is facing the symbolic AI re-
searchers is that the symbolic AI systems exhibit very poor performance in the area of
pattern recognition and the retrival of knowledge. Such mechanisms are considered very
crucial, at first glance humans seem to use such mechanisms always for problem solving.
The following sections introduce a wide variety of abilities which are dependent upon
pattern-recognition.
2.2.3 Building Analogies
One of the fundamental aspects of intelligence that humans display is the ability to draw
analogies between various forms of skills, events, and knowledge. A similar mechanism,
if implemented in a AI system, should be able to draw similar analogies between events
which happen. In order to draw such analogies, the mechanism should be presented
with knowledge which is represented in a form which possess certain similar structures.
However, in order to perform the task of drawing analogies, the mechanism will have to
examine the knowledge structures at varying levels of "detail". In the course of searching
for analogies, a more detailed analysis of subabstractions of the event should be done.
In the case of temporal events, we would have to examine and identify the temporal
relations which occur between subabstractions in an event. To an experienced individual,
the partial set of states presented in the freeze frame is sufficient for the conclusion that
the event Walking was actually being depicted in the freeze frame. Consider this case,
where the individual with no knowledge of what Walking is, can this individual learn the
concept of Walking which has just been presented to him ? Consider that the human
being has quite a wide domain of varying forms of knowledge which is obtained, is it
quite possible to state that we are not aware of the how we represent new information
in memory. Thus, this question arises, should an AI system be even aware on how its
15
knowledge-based library of information be stored, and if not, how do we represent such
information in a manner which is explicit and beneficial for the "intelligent engines" in
this system to use. A genuine reason for an interest in this question, is that a large
knowledge-based system will eventually contain a massive library of assertions. The
mechanisms of the system which use this library of assertions must be able to use the
information efficiently. Furthermore, there will probably be a wide and varying variety of
different levels of information. It is important that the various mechanisms of the system
which manipulate this library of assertions are able to "figure out" what is going on,
what was stored in the database, and how to use these knowledge structures accordingly.
In the domain of building systems for understanding stories, the role of analogy is
considered to be very important in the process of describing the thematic content of the
story. For instance, the Broadway musical "The West Side Story" is very much based
upon the story of "Romeo and Juliet". An observer who has read both stories will have
no trouble finding an "analogous flavor" relating them, although if the reader analyzes
both stories are a lower level of detail, he will find a fairly large number of differences
between the two stories.
2.2.4 The importance of inference
The ability to make appropriate inferences is important in any intelligent system, be it a
natural or an artificial one. As mentioned in the previous section, the reader of a story
must be able to deduce and make certain inferences based upon the story being read.
By using inferences, a system can learn "much more", in the sense that it can make
implicit inferences which lead to new ideas and concepts, by also depending on prior
default knowledge.
2.2.5 Consistency of Knowledge Structures
Another question which stimulates curiosity with is the notion of maintaining informa-
tion consistency. Assuming we have a very large database of knowledge which we have
accumulated across a long period of time, e.g. our own experiences that we have had
16
stored in memory, how do we maintain a consistent view of the experiences we had, in
temporal terms ? In my opinion, I don't believe that we actually maintain total temporal
consistency of events which had occured in our memory. Let me give an example which
was given by Minsky [Minsky, 90], that shows inconsistency among answers which we
may give to questions about famous history events.
Person A: Which event do you think happened first, the American
revolution or the French revolution ?
Person B: The French revolution.
Someone who has studied history will know that the American Revolution had taken
place earlier (in 1776) than the French Revolution (in 1789). Obviously, Person B was
wrong in choosing the latter. Why did this person choose the wrong answer? I carried
out this experiment with a number of people, and among those who stated the wrong
answer mentioned that France seemed much 'older' than the United States. Let me now
proceed with this argument. I boldly state that in this case, the number of persons who
initially given the correct answer will decrease. The reason, I believe, is that many of
these individuals had plainly used the dates of the occurences of both events to help them
in making a guess. Now, if we hypothetically erase both dates from their memories, they
will have to resort to using other heuristics. If we look back at the group who seemed to
think that the revolution in France had taken place earlier since it was an 'older' country.
Here, they have made the error in judgement, by using a simple heuristic of temporal
reasoning. There is no doubt that an error was made, but the decision was made within
a short period of time. Is the knowledge that we represented in our memory incorrect,
or did we use an incorrect heuristic to come to the conclusion ?
2.3 A Flavor of Knowledge Structures
This part of the thesis will cover the motivations which led to the development of the
frame-system theory by Minsky [Minsky, 75] and another theory of knowledge structures
quite similar to frame-systems, the notion of scripts and plans developed by Schank
17
Jane's birthday party
relation links
Music Laughter
default assignments
Birthday cake Balloons Friends of the nodes
Figure 2.1: Frame representation for Jane's birthday party
[Schank, 1977]. Both these theories of knowledge structures are considered to be signifi-
cant and essential in building AI systems which exhibit common-sense reasoning.
2.4 Frames
The idea of frames was first proposed by Minsky [Minsky, 75]. The frame theory is con-
sidered to be an key essence in the construction very large and rich databases which
contain encylclopedic domains of knowledge needed in a system used for common-sense
reasoning. In particular, the main idea of the theory was to construct such a database
which encoded real-world knowledge in a structured and flexible manner [Shapiro, 1987].
A frame is considered to be a structure for representing knowledge of stereotyped situa-
tions such as a children's birthday party, being in the living room of your Aunt Mary's
home in New Jersey, etc. Such a frame structure allows a computer system which is pro-
cessing input from the external environment to exhibit coherence and understanding in
similar situations, and would allow the system to access information which is needed for
novel situations which are not anticipated prior to processing. Attached to such frames
are default assignments of information which are present in those situations, and also in-
formation on usage and functions of that particular frame system. The theory of frames
was constructed with the essence that knowledge had to consist of "larger and structured
chunks", and that their procedural and factual contents were linked in intimate manners.
18
2.5 Charniak's Model of Story Understanding
2.5.1 Charniak's Demons
The following is a definition of "demons" given by Charniak [Charniak, 1972].
demons Facts which are introduced by "concepts" occuring in the story are called
"demons" since in many cases they must wait for further information. In such
cases we can think of them "looking" for the appropriate fact. So "not being
willing to trade" might put in a demon looking for another offer.
Mary was invited to Jack's party. She wondered if he would like a kite.
Charniak [Charniak, 1972] was interested in learning how we intepret text sentences,
in particular how the processing of a prior sentence effected the way we "processed"
the sentences that followed it. Charniak suggested that the processing of a sentence
or an event was immediately followed by the activation of recognition-agents known as
"demons". The activated demons then will proceed with analyzing further data input
(the remaining sentences of the story), and will dynamically alter the further processing
of text sentences. Charniak's system introduces "concepts" while processing sentences
in a story. These concepts influences the the processing of future sentences of the story.
In this system, "demons" are introduced into the story whenever the proper concept
(relative to a particular demon) has been mentioned in the story.
1. He plunked down $5 at the window.
2. She tried to give him $2.50, but he wouldn't take it.
3. So when they got inside, she bought him a large bag of popcorn.
Consider the example above which is given as an example of how we "activate" concepts
while processing text sentences. If we first read the first sentence, we will probably invoke
a concept of horse race frame [Shapiro, 1987]. However, if we carry on reading the second
sentence, we may conclude that the event was an attempt to return change. However,
with the processing of the third sentence, that the whole sequence of event was probably
describing a date.
19
2.5.2 Filling in Details
Consider the following short story (given by Miikkulainen and Dyer [Miikkulainen, 89]).
John went to Leone's. John asked the waiter for lobster. John gave a large
tip.
Based upon our own personal experience, we can probably come up with a few other
details and events to the story.
John went to Leone's. The waiter seated John. John asked the waiter
for lobster. The waiter brought John the lobster. John ate the lobster. The
lobster tasted good. John paid the waiter. John gave a large tip. John left
Leone's.
How easy should it be to activate demons ? How long should they remain active ? How
many of them should be activated ? Charniak studied how questions which were related
to a story could be answered by a story understanding system. It seemed that in order
to do this, there was a need for a substantial amount of knowledge which is not directly
contained in the story. Charniak's model of children's story understanding was one of
the first AI systems developed which followed a trend in AI representation theory to use
"large chunks" of knowledge [Charniak, 1972]. Charniak's model is considered to be an
early idea of what Minsky defines as being a frame system [Minsky, 75]. In comparison
to Winston's model for learning structural descriptions, Charniak's model was quite
different. Winston's system lacked a rich containment of default knowledge. Instead, it
relied on examples of correct structural descriptions, and examples which were considered
to be near misses of the structural descriptions by the teacher supervisor. In Winston's
case, the student learner had a goal, that is to find the correct structural description
for a object block. However with the proper sequence presentation of carefully chosen
structural examples, Winston's system could deduce the correct default properties of the
block structure. On the other hand, a system which is build to understand a story does
not necessarily have definite goals to work on. Thus when such a system is processing
a story, it is difficult for the system to deduce indirect consequences of such a story
20
without having default assignments of values. However, there is a strong resemblance
of Charniak's system and Winston's system. Each system, if given the proper sequence
of events and information can build and construct a general idea of what was being
presented to them. This is quite true in individuals who write stories. Such individuals
write stories in a fashion such that readers are able to grasp the ideas which the authors
are to present. Most importantly, the writers do take into account the reader audience
that the stories are intended for. In this sense, the writers themselves conform to a type
of default structure for the type of audience the stories are intended for.
2.5.3 Good ideas from Winston's system
Winston's system for recognizing structural descriptions from examples exhibits some
heuristics which are considered to be excellent, and should be used in conjunction with
the frame theory. Winston's system was able to learn the correct description of an
object by systematically adding and deleting features/relations that were present in the
objects or the near-misses. Thus after going through a series of near-misses and correct
descriptions, the system was able to create a correct structural description of the object.
The system exhibited a behavior which is considered to be a refinement process. At
each stage while the system was processing the structural descriptions given to it, it
selectively constructed a database containing the object's structural description by using
pattern matching or a graph-matching process.
2.5.4 Groupings
Winston's system exhibited the ability to form generalizations about the underlying
important features of a certain structure. For instance, the system could create a gen-
eralization for a Tower structure, which it converged with the generalization of a three
block object in the BLOCKS world with the following characteristic relations
Clear(A) (2.1)
Between(x, (A, C)) for any number of x (2.2)
On(C, Table) (2.3)
21
The method of learning the general description of a Tower in Winston's system can be
viewed as being very similiar to the creation of a frame description. In this case, the three
most important structural features which categorized the object Tower was captured by
the system. These structural descriptions can be viewed as being the relational links
between slots in a frame description.
However, Winston's system was fortunate and successful in the sense that it had an
excellent teacher who could provide it with good examples of what were considered to
be correct samples of objects, and near-misses of object samples. In addition to this, the
teacher has also provided only key descriptions (data) which it considered to be essential
in learning the structural description of an object. Such a problem is evident while
constructing a system which processes information from the outside world, a process
considered to be similar to human perception. For example, consider a system which
was built for object recognition. If we were just to consider the amount of raw data
which is to be processed from the visual image, the amount of computation needed will
be inherently be large, expensive, and slow. In such a case, I would render that the
frame theory will provide an excellent solution since it eliminates the need to process all
the raw data which is being fed into the system. A system which uses the frame theory
as a platform will only need to selectively choose the input which the particular frame
structure considers "relevant" for common-sense reasoning.
Although Winston's system displayed some considerable success in learning the struc-
tural properties, it relied on the teacher to provide it with features which were considered
important to the learning of structural descriptions. In other words, the teacher had cho-
sen a representation which had exploited the important aspects in the problem of learning
structural descriptions. In the case of Winston's system, the teacher explicitly empha-
sized the spatial and geometric relational links between the objects and their features.
Winston himself acknowledged the possibility of complications if the number of oject
features and their relations grew larger in size, than the ones he had used for the system.
In the case of more complex and larger sets of features and links, the system would have
a much more difficult time in matching the features between the different examples given
to it. I would expect that the number of examples of correct and near-misses which have
22
I
to be presented to the system in order for it to learn could easily increase by an expo-
nential factor. I have reason to believe that such an expection is a lower-bound estimate
on the scalability of the problem since we assume that all features are "equal" in nature.
There are strong reasons to believe that more complex features are in existance within
the framework of a certain problem. The notion of "intelligence" is one good example of
a very complex feature which humans possess. How does one categorize "intelligence"?
Can we teach a system to recognize the existance of "intelligence"? The Turing test is
one of the classic tests which were proposed by AI researchers that could be used to test
if a machine was intelligent. A machine which passed the Turing test is thus said to be
indistinguishable from a human being, relative to another human being's observation.
2.5.5 Folklore Beliefs on Relational Links
Human languages seem magical and mysterious. How can the manipulation of a formal
set of symbols and notations convey an seemingly infinite group of information ? The
ways of which we express these notations (in forms such as voice tones, emphasis, etc.)
seem to also convey various different meanings and intepretations. At this point, let us
take a look again at Charniak's model of understanding children's stories. Why wasn't
this system as successful as Winston's model for learning structural descriptions from ex-
amples ? Besides the point of not having a supervisor, Charniak's system attempted to
implement natural language processing in the context of understanding children's stories.
In such domains of natural language, there are strong reasons which indicate the task
of analyzing the semantic content of sentences is very difficult and ambiguous indeed.
Such stories are very much different from the problem of learning structural descriptions.
The latter is very well-defined, with laws of geometry and spatial relations governing the
problem. On the other hand, Charniak's system tries to understand children's stories.
The laws of processing, creating, and understanding the semantic contents of such sto-
ries is very subjective and ill-defined in comparison to the processing of the traditional
informative text found in science articles, financial reports, etc. Can we actually build a
correct semantic representation of the former ?
23
___.J
2.5.6 Inherent Difficulty of Relational Links
One of the principal difficulties which can observed in data intepretation lies in the
domain of the relational links. It is quite evident that the context in which we intepret
data from the environment lies in the way which we introduce the relational links which
connect the various domains of information available from the input. Once again, we
take a look at Winston's system and see why his system displayed success in learning
structural descriptions. In the case of the examples used, it only contained prepositions
which "captured" the important aspects of learning structural descriptions. However,
this is not quite true if a teacher which can give the "correct hints" is not available. In
such cases, how do we generate and assign relational links to the data which we receive
then ? This problem is very evident in problem domains which are very complicated
(many control variables and parameters), a good and relatively humorous example are
the dynamics of the stock market. No one seems to be able to correctly predict the
trends of the market with great accuracy. A same analogy could be applied to other
domains as well, if we cannot learn and process the relational links which exist between
different domains of the data, we may never learn the concept being presented. Could
this problem be very similar to ones Piaget observed in young children then ? Children
of the ages 3-5 years old cannot grasp the principle of conservation. In this case, the
same amount of liquid is poured into two different containers. The first container is a
thin and tall container, while the second container is shaped very low and wide instead.
Almost all of the children in this age group will testify that the first container has a
larger amount of liquid than the second one.
It seems quite clear the children do not have the ability to recognize and observe
the relational function which connected the initial frame to the final frame. What was
very important in this relational function of Pouring Water into the Container was that
the amount of water from the initial frame was the same as the final frame. Instead
of observing the transformational function, the children had instead just focused on the
comparison of the two final frames (referring to the water in the first container, and the
water in the second container after the pouring of water). During this comparison, it
is quite possible that children were relying on the relational links and the observational
24
I
data solely present in the final frames. In such a case, the children would have used the
relation of Higher Level meaning "more" water to make their judgement. Thus they had
failed to observe the importance of the conservation of liquid in the transformational
function/relation. One more example is another observation made by Piaget on young
children. The following is a reconstruction of the experiment conducted by him.
The little girl was playing with a toy cat. There was a small blanket which
was near them. I came closer, and took the cat away from the little girl. I
placed the cat under the blanket while the girl was looking. The girl started
crying when she couldn't find the toy..
In this case, we could apply a similar explanation for the observation. The child was
not able to grasp the fact that the toy had not disappeared, but was merely under the
blanket. Further, this could be due to the fact that the child had only processed the
information of the final frame, which did not explicitly indicate the prescence of the toy
under the blanket. In this case, the child had ignored (or failed to recognize) the "impor-
tant" characteristic of the transformational function Hide The Toy Under TheBlanket, the
characteristic that the object would still be in the environment of reference in the final
frame, only that it was under the blanket this time around.
This phenomena also exists as a problem in a number of AI systems which have been
developed. This problem of not being able to use the transformational relations between
frames is evident in the STRIPS problem solver. In the case of STRIPS, its mechanisms
run problem-solving methods on only the current state of events which appear in the par-
ticular frame which it is processing, and it does not have any database which contains the
history of past events. Thus, it is quite possible that the system may repeat procedures
which have (in the past) been carried out for problem-solving, and end up have the same
results, on the negative side this results as failures to solve the problem. Furthermore, if
such a system does not reach its success goal, it cannot resort to using heuristics which
in the past history have been known to be successful. STRIPS seem to act like the young
child given in the example above. It probably cannot realize that the transformational
event of putting the toy under the blanket maintains that the toy is under the blanket.
Instead, it processes the final frame, where the notion of the toy being under the blanket
25
is not present explicitly (if you assume that the vision mechanism is the one providing
the input from the frame).
2.6 The Problem of Frame Recognition
"Things that which are so important to us are easiest to retrive."
Frame recognition is one important problem in the whole essence of the frame theory
which has yet to draw any solid answers. As examples which were given in the previous
sections, we did not have any trouble activating the frames for the date event. However,
the method on how we perform this skill is still not known clearly. I have reason to
believe that the following "things" are important towards building frame recognizers.
These "things" are just as essential towards the whole notion of the frame theory.
* refinement process of learning
" organization of knowledge and relations
" intentions/goals of parties
2.6.1 The Problem of Too Many Frames
Frames are a collection of knowledge structures which were constructed as a result of
events, concepts, and experiences that we have accumulated throughout time. While the
concept of frames seem natural to us, a side effect of having such forms of knowledge
structures is their size and numbers. It is quite possible that we may have a typical
long-term memory storage which contains a million different types of frames. The large
number of frames is accounted for by the large number of stereotypical events, concepts,
ideas, etc. which an adult accumulates during an average lifetime. This relatively large
number of such structures poses issues which question the theory of frames. How do we
recognize frames which need to be activated from this large number of frames ? In the
previous sections, a stereotypical situation of Aunt Mary's home in New Jersey (frame)
was given as an example. How is this frame chosen, recognized, and activated from a set
26
of one million other different frames ? There could be several ways which we recognize
objects, and they can be listed as follows
* data slots of frames
9 relational links of frames
* the combination of slots and relational links
A recognition mechanism based on only the first heuristic can be very simple. In this case,
we recognize and familiarize a situation with a certain stereotypical frame based upon
the base components which are taken from the situation/event currently being processed.
However, there are a number of disadvantages of this heuristic. This heuristic allows the
possible activation of a large number of frames residing in memory. For instance, if we
were to activate all possible frames which have an object like Car, the number of such
activated frames can be large. Another example in which this recognition heuristic fails
takes place in visual perception. Such an example is once again the case of features of a
face (eyes, ears, nose, etc.) which are presented in a non-facial form (all these features
are disjointed and separated). We obviously do not activate a face frame using this
heuristic, rather this observation points towards the validity of the second heuristic used,
the relational links of the frames.
The usage of relational links in frames for recognition purposes are more suited to-
wards recognizing possible analogies between the observed situation and a particular
stored frame. As the earlier example of West Side Story and Romeo V Juliet, the "simi-
larities" which exist between them exist at a higher level than the level of the base com-
ponents. In this case, the "similarities" in the relational links of these two frame events
led to the recognition that the West Side Story musical is merely a modern version of
Romeo V Juliet. Scientist classify this level of recognition as being more "abstract". At
this point, there seem to exist a difficult question which needs to be answered. Does
the ability to draw such abstract analogies a good indication that we understand the
functions and the structures of the relational links in a frame? Or can we actually draw
correct analogies without understanding or being aware of all the relational links in the
frame?
27
Looking back again at Winston's system, the problem of pattern recognition was
somewhat more easier. This was due to the paradigm model which was used for the
system. It assumed that the presenter (teacher) was correct with presenting the system
with correct and near-misses (incorrect) structural descriptions of objects.
At this point, it would be interesting to reintroduce an important feature of frames.
Knowledge which are contained in frames are most likely linked in a graph-like represen-
tation, rather than a feature vector form of representation. This conclusion is mainly due
to the observation that such facts and knowledge which are present in a frame structure
must be "connected" in certain manners, for instance, if A is an example of B, then such
a concept may be represented and connected in the frame with a IS-A link. The impor-
tance of these relational links in frames can be justified as the this following example
shows. Suppose I present you with two figures. The first figure contains an image of your
face. The second image will contain all the components of your face (e.g. nose, eyes, ears,
mouth), however they are all "jumbled" up. When presented with the two figures, we
can immediately recognize the former, but not the latter. This show that the relational
links between the components play a key role in the description and thus recognition of
a frame.
2.6.2 Prototypical Frame Features
This section will introduce the issue of prototypical features which may appear as default
slots in frames. This issue is very much related to the notion of scripts which will be
discussed later. Relating back to the frame which covers the event of Jane's birthday
party. In this event frame, we have specific default slots, such as party balloons, birthday
cake, presents, friends and relatives, etc. However, we have also left out some finer
details. For instance, when we mentioned balloons as being present in the event, we do
also have a general notion of what type of balloons we have present, what color they
have, etc. Curiously however, these "micro-features" are not considered to be important
to the frame, however yet we still have a certain "sense" of their prescene.
28
__J
2.7 Scripts
The relational links in frames are very much related to the notion of script which were
proposed by Schank and Abelson [Schank, 1977]. They proposed that knowledge that
we have of everyday common events are organized into knowledge structures which are
labelled as scripts. Events which appear in scripts are considered as the primary entries
of this knowledge structure. These events are connected by causal links which relate
these events, in a serial-type manner. For example, the story of John dining at the resta-
raunt [Miikkulainen, 89] mentioned previously is considered to be a script. Scripts are
constructed with emphasis given to the ordering of events in accordance to stereotypical
events which are commonly encountered.
2.7.1 The Relevance of Scripts in Frame Theory
Scripts seem to be a naturally occuring event of the frame theory. In the theory of frames,
we have default slots and relations which describe a particular event or concept. If the
representation of the frame is constructed properly, the roles and functional descriptions
of the agents participating in the event will be captured. Thus, by using the functional
descriptions and roles of these "agents", we can successfully predict their sequence of
"role-playing" in the event. With the goals and intentions of the agents relative to a
particular frame made available for processing, one can construct a general-type script
to describe the event.
Like frames, some events which are considered to be unimportant to the particular
context which it is being processed, are not "activated".
2.7.2 What Demons Should and Shouldn't Do
The problem of having a very large number of possible frames for all events of possible
stories is related closely to the question of object-recognition systems which rely on a
knowledge base. Such an object recognition scheme will basically use its prior knowledge
(residing in the knowledge base) to perform object matching and recognition. Obviously,
if we had all possible variations of even a commonly known object, for instance a Auto-
29
I
mobile, the number of Car templates which are stored in the knowledge base is extremely
large. As a result, the process of'matching the currently observed automobile with an
entry in the knowledge base will be computationally slow, if carried out without any
search heuristics.
Charniak's model of story understanding would be ideal if all stories were written in
the same format as Charniak's model would expect [Wilensky, 1978], that is a knowledge
structure which was a frame type. However, such a system cannot adapt to novel situa-
tions. Such situations involves children's stories like the one give below which Wilensky
[Wilensky, 1978] and Charniak himself pointed out as being problematic.
Jack was going to paint. He washed the brush.
In this story, Jack is presumed to have washed the brush so that it was clean. Charniak's
system could not find this interpretation since its painting frame did not possess the
particular event describing the brush cleaning at the beginning of its activated frame.
One could argue then that this particular event could be added to the painting frame.
However, if we were to account for all the possible stories and events which could be
fed to the system, the number of such stories is enormously large. This would be of no
advantage since the system would have to search through a large database of frames to
find a correct match. Such a search is computationally slow and inefficient, and it is
most likely that we do not have such a methodology. However, I believe that the frame
theory is an strong example of an excellent Knowledge Representation model of AI, thus
the problem of having large sets of frames and irrelevant inferences should be addressed.
2.7.3 A Counter Argument
From the previous section, we obtained Wilensky's argument that Charniak's system
would fail upon the example given by him. There seems to be a solution to this problem.
Suppose we asssumed that the event of washing the paint brush was linked to the painting
frame via a relational link such as the following. In this case it is assumed that the event
of cleaning the paint brush (we call this event A) is "small" compared to the event of
30
Mlw
Understanding
Learning
~Analogy ??
Knowledge structures Inaleny ??
____________________Inferences ??
Pattern Matching ??
O o o Globs of data input
Figure 2.2: A Framework of an AI system ?
painting (we call this event B).
(2.4)
however, the relation is not symmetric, as shown below.
(2.5)
Thus, can one say that the reason the Painting frame did not "contain" the event of
washing the brush, is because our system could not trace it to the latter ? In the example
given by Wilensky to criticize Charniak's system, we can use the following explanation.
After processing the first sentence "Jack was going to paint. ", we did not arrive with the
activation of the brushing cleaning expectation, since we have no way of reaching out and
activating the latter, since there was no causal link present in that direction. However,
the latter event was still contained in the general frame of the painting event. If so, why
do we do this ? This problem is evident in the design of relational databases in computer
systems. Designers of such databases have pointers relating one entry to another via
the method of using pointers. However, since pointers only travel in one direction, there
cannot be any pathway which the pointed can go back to see who the pointer is.
31
I
2.7.4 How Large are Frames?
At this point, we don't really have a definite bound on the size of frames which can
be generated and used. However, it is possible that we use the following heuristics for
determining the size of frames.
* The type of event which is being represented
* The context of which the event appears in
* The information processing capabilities of a system
The first two heuristics are very closely related in determining the size of a frame. It is
not clear yet if there exists subframes which are used in building a larger frame. Frames
of different types can share a common set of slots/features, therefore, it is very difficult
and complicated to draw boundaries between frames of this type. Such an observational
difficulty poses the following question about the strength of the frame theory itself.
Do we actually have these frames in our knowledge bases, or do we use a certain set
of operations/transformations on a common database of raw knowledge such that the
result of these transformations ? If we were to assume that we do not use the latter
type of representation, we must then address the important issue which arises from the
common sharing of data. This issue once again deals with the importance of the relational
links connecting the data. How do we distinguish the different frames with ease, with
the difficult situation which arises with the sharing of common knowledge assertions. It
seems that the ability to do this is either caused by the following
* The frames which are constructed in memory do not share any particular common
knowledge, i.e. there are no intersections between the "frames".
e Alternatively, the mechanism which can recognize and activate the correct frame
must depend heavily upon the relational/structural links which connect the com-
mon sets of knowledge
The third heuristic mentioned is based upon the fact psychological observation that hu-
man beings have a limited capacity of information processing. This is based upon a
32
cognitive argument that we have a limited amount of short-term memory. With this
limitation in existance, we resorted to (by choice or by evolution) the construction of
frames which faciliated in the processing of information. The assistance which was pro-
vided by the formation of frames was quite similar to the concept of using heuristics for
problem solving. We used these heuristics to serve as "guidelines" to the direction which
we should head to. Similarly, we can activate a certain frame to facilitate the processing
of a certain situation.
2.7.5 Information Processing Methods
Winston's model of learning structural examples is a system which uses data-driven)
processing. In data-driven processing, the inferences that the system make are just
strictly dependent upon the input it receives, in the case of Winston's system, the input
are the sequence structural examples of an object. On the other hand, Charniak's model
is an example of top-down processing. In this type of system, the program will have a
knowledge structure activated. The system will thus have to determine how it will match
the data input it receives (in Charniak's system, this consists of sentences in the story)
with the "data slots" that are present in the knowledge structure. While matching takes
place between the data and the knowledge structure Thus top-down processing can be
described as a mechanism of prediction[Wilensky, 1978].
33
Chapter 3
Conclusion
In summary, there still remain open questions in the frame theory which need attention.
These questions cover specific areas, such as the problem of organizing large domains of
knowledge. The problem of organizing large domains of knowledge can substantially be
reduced if a good set of heuristics can be developed to explore and "understand" the
relational links which exist between different knowledge structures. AI systems which
have displayed success, such as Winston's model for learning structural descriptions from
examples, were able to develop general sets of concepts through a process of refinement.
However, in the case of Winston's system, it relied upon a teacher to present it with
examples which emphasized the important relational links between different parts of the
information.
While it is crucial to recognize the importance of such relational links, it is still
a very difficult task to understand the structure of such relational links, ones which
represent a "concept". The idea of a concept remains mysterious, we seem to able to
grasp such high-level knowledge structures, but we cannot seem to be able to understand
fully the underlying composition and operations of the relational links. This problem
is more apparent in knowledge which are catergorized to be of "common-sense" nature,
such as the knowledge required to understand children's stories. Such situations are
much more difficult than building a system to understand scientific text. In the case
of scientific text, most of the text processing involves building a knowledge structure
which contains objective data. This differs greatly from the processing of understanding
children's stories. The text contained in children's stories have "thematic" contents
34
which are very subject indeed, dependent upon the goals and themes which the write
and the reader have. In short, there does not seem to be an objective method for learning
"concepts" which are presented in a child's story. These concepts have very subjective
interpretations, thus their network of relational links between the different knowledge
structures in the "concept" is very context-dependent. This is due to the observation
that concepts which are generated in context of common-sense knowledge seldom contain
objective forms of information. In particular, the relational links which are present in
the structural representation of a "concept" can be non-objective themselves (subject to
context). To build a successful system which can organize "common-sense" knowledge
from different domains, the questions which are expected to be of importance are the
following
* The objectivity of relational links
* The order of importance of such links
An important aspect of relational links which should be given attention are the temporal
relational links which exist between knowledge segments in a frame. It is still unclear
on how these relations are built, and how they are used in relation to common-sense
knowledge. Another issue which needs further study is the nature of accessing and cate-
gorizing relational links. This question is in part raised by the organizational hierarchy
of frames. How are the slots in the frames organized ? There does not seem to be a
definite manner for organizing the different "levels" of a frame. Rather, the manner of
which the structure of the frame which we use seem to be very context-dependent. Thus
it is quite possible that frames themselves are a form of graph representation which can
be described as "non-hierachial". In such a case, it is possible that all the slots which
exist in the frame can be accessed since there is no usage of "directional" pointers.
It is our hope that future work in AI will focus on studying and building systems
which exhibit common-sense knowledge. Current research in Al places emphasis on
problems which are very specialized and domain limited. These trends in AI research are
considered not to be of fundamental importance to the field of AL. Instead, we should
focus our efforts on solving problems which are considered by humans to be straight
35
forward and easy. Good solutions on how we solve these "easy" problems will provide
fruitful insights to the advancement of AL.
36
Bibliography
[Bobrow, 1975]
[Charniak, 1972]
[Dyer, 1983]
[Guha, 1989a]
[Guha, 1989b]
[Kahn, 1975]
[Kuipers, 1976]
[Lenat, 90]
[Marr, 76]
[McClelland, 86]
[Miikkulainen, 89]
Bobrow, Daniel G. and Donald A. Norman. Some Principles of Mem-
ory Schemata. In Language, Thought, and Culture: Advances in the
Study of Cognition. Bobrow, Daniel G. and Collins, Allan (editors)
Academic Press 1975.
Charniak, Eugene. Towards a Model of Children's Story Comprehen-
sion. MIT Artificial Intelligence Laboratory AI-TR-266.
Dyer, Michael G. In-Depth Understanding: A Computer Model of
Integrated Processing for Narrative Comprehension. MIT Press 1983.
Guha, R.V. and Douglas B. Lenat. The World According to Cyc, Part
2: Agenthood, Instituitions, and Agreements. MCC Techical Report
No. ACT-AI-453-89
Guha, R.V. and Douglas B. Lenat. The World According to Cyc, Part
3. MCC Technical Report No. ACT-AI-455-89
Kahn, Kenneth M. Mechanization of Temporal Knowledge. MIT
Project Mac MAC-TR-155
Kuipers, Benjamin. Spatial Knowledge. MIT Artificial Intelligence
Laboratory AI-Memo No. 359
Lenat, Douglas B. and R.V. Guha. Building Large Knowledge-Based
Systems: Representation and Inference in the CYC Project. Addison-
Wesley, 1990
Marr, David. Artificial Intelligence - A personal view. MIT AI Lab-
oratory Memo No. 355.
McClelland, John and David Rumelhart (editors). PDP: Explorations
in Parallel Distributed Computing Vol. I, II and III The MIT Press,
1986.
Miikkulainan, Risto and Michael G. Dyer. "A Modular Neural Net-
work Architecture for Sequential Paraphrasing of Script-Based Sto-
ries." Proceedings of the 1989 IEEE Conference on Neural Networks
pp. (11)49-56.
37
[Miller, 56]
[Minsky, 61]
[Minsky, 68]
[Minsky, 75]
[Minsky, 85]
[Minsky, 90]
[Schank, 1977]
[Shapiro, 1987]
[Simon, 1979]
[Van Baalen, 88]
[Wilensky, 1978]
[Winston, 1970]
Miller, G.A. "The magical number seven, plus or minus two: Some
limits on our capacity for processing information." Psychology Review
63:pp. 81-97.
Minsky, Marvin L. Steps Towards Artificial Intelligence. Proceedings
of the IRE Vol. 49, No. 1, 1961.
Minsky, Marvin L. and Seymour Papert. Perceptrons. MIT Press
1968.
Minsky, Marvin L. "A Framework for Representing Knowledge."
In Readings in Knowledge Representation R.J. Brachman and H.J.
Levesque (editors). Morgan Kaufmann Publishers 1985.
Minsky, Marvin L. The Society of Mind. Simon and Schuster, New
York 1985.
Minsky, Marvin L. Personal communications. 1990
Schank, Roger C. and Robert Abelson. Scripts, Plans, Goals, and
Understanding. Lawrence Erlbaum Press, Hillsdale, New Jersey.
The Encyclopedia of Artificial Intelligence Edward Shapiro (editor).
John Wiley 1987.
Simon, Herbert A. Models of Thought. Yale University Press 1979.
Van Baleen, Jeffrey and Randall Davis. "Overview of an Approach
to Representation Design." In Proceedings of the National Conference
on Artificial Intelligence, pp. 392-397, Minneapolis, MN, 1988.
Wilensky, Robert. Understanding Goal-based Stories. Yale University
Research Report No. 140.
Winston, Patrick H. Learning Structural Descriptions from Examples.
MIT Artificial Intelligence Laboratory AI TR-231.
38
---A