An Examination and Confirmation of a Macro Theory
of Conversations
through
A Realization of the Protologic Lp by Microscopic
Simulation
A thesis submitted for the degree of Doctor of
Philosophy
by
Paul A Pangaro
Department of Cybernetics, Brunel University
May 1987
Please review the notes that accompany
the index for this document.
Abstract
Conversation Theory is a theory of interaction. From interaction (the theory
asserts) arises all individuals and all concepts. Interaction, if it is
to allow for evolution, must perforce contain conflict, and, if concepts
and individuals are to endure, resolution of conflict.
Conversation Theory as developed by Pask led to the protologic called Lp
which describes the interaction of conceptual entities. Lp contains injunctions
as to how entities can and may interact, including how they may conflict
and how their conflict may be resolved. Unlike existing software implementations
based on Conversation Theory, Lp in its pure form is a logic of process
as well as coherence and distinction.
The hypothesis is that a low-level simulation of Lp, that of an internal
and microscopic level in which topics are influenced by "forces"
that are exerted by the topology of the conceptual space, would, in its
activation as a dynamic process of appropriate dimension, produce as a result
(and hence be a confirmation of) the macroscopically-observed behavior of
the system manifest as conflict and resolution of conflict. Without this
confirmation, the relationships between Conversation Theory and Lp remain
only proposed; with it, their mutual consistencies, and validity as a model
of cognition, are affirmed.
The background of Conversation Theory and Lp necessary to support the thesis
is presented, along with a comparison of other software approaches to related
problems. A description of THOUGHTSTICKER, a current embodiment of Lp at
the macro level, provides a detailed sense of the Lp operations. Then a
computer program (developed to provide a proof by demonstration of the thesis)
is described, in which a microscopic simulation of Lp processes confirms
the macroscopic behavior predicted by Conversation Theory. Conversation
Theory thereby gains support for its use as a valid observer's language
for every-day experience, owing to this confirmation and its protologic
as a basis for psychological phenomena in the interaction of conceptual
entities of mind.
Acknowledgments
There are many individuals who must be thanked for their help in the research
and production of this dissertation. First and foremost is my thesis advisor,
mentor and friend Dr Gordon Pask whose intellectual and spiritual life have
been the greatest influence in my career.
To the many individuals who made up System Research Ltd over its long existence
my thanks to them must be anonymous. My appreciation is especially strong
for those who suffered the pressures of its research programme and research
conditions and who may or may not be individually identified for their very
tangible contributions to Conversation Theory. Elizabeth Pask provided emotional
support and personal expression of a kind that is rare in the world and
without which I could not have persisted at System Research.
Mr Colin Sheppard of the UK Admiralty Research Establishment (ARE) provided
contract support for the construction of THOUGHTSTICKER at a time when its
subtlety and power could be seen only as a concept. He must be acknowledged
and thanked for continuing the type of crucial and discriminating support
championed by Dr Joseph Zeidner, who during his tenure as Technical Director
of the US Army Research Institute supported the work of Pask for its own
sake. Mr Dik Gregory also of ARE provided intellectual support and contributions
to the construction of THOUGHTSTICKER during its development.
Dr Jeffrey Nicoll while Director of Research at PANGARO Incorporated constructed
the complex innards of THOUGHTSTICKER and hence conquered both the Symbolics
environment and the work of the Pask on the subject of Conversation Theory.
He has also contributed to its formal and theoretical side. He provided
undaunted moral support for my efforts on the dissertation and continues
to be an important collaborator and close associate.
Mr Peter Paine as my dual in PANGARO Limited has provided strong support
and has allowed me to take the time and resources to complete this dissertation,
often to his own disadvantage when timescales and responsibilities of contracts
were very great.
Others who provided moral support without which I could not have completed
are Herbert Brun, Patricia Clough, Graham Copeland, Karen Rose Elder, Michael
Granat and Symbolics Education Services, Christina Gibbs, Heather Harney,
Kevin Kreitman, Shelby Miller, Abe Raher, Vivian Scott, Louis Slesin, Ricardo
Uribe, Eric Wolf, and especially Heinz von Foerster, for his foundations
for Conversation Theory and the untiring vitality he has expressed to me.
I. Preface
I.1 Structure of the Dissertation
In the writing of this dissertation, while covering the necessary points
on the main issues of the thesis itself, I was encouraged by my colleagues
to insure that I had provided the following elements:
- Background on Conversation Theory itself.
- A personal history of my involvement with Conversation Theory, including
why I had adopted it as an approach to the problems that interested me.
- An indication of my own contributions to the field.
Upon review of the drafts up to a certain point, I thought that Point 3
had not been adequately expressed. I believe this had been the case in part
to keep a dispassionate tone in a scientific work. Also, I specifically
did not want to overstep a basic humility by giving attribution to myself
as a single individual where the ideas were so much the combination of past
efforts and more recent expressions of individuals other than myself. It
is particularly awkward to make such differentiations in the context of
a cognitive theory which emphasizes the ever-shifting definition of "individual"
based on beliefs rather than biological identification. The theory also
discourages attribution by providing a detailed model of how new ideas can
arise only from the seeds provided by others, in a dimensionality of time
that is neither linear nor fully ordered. However I can state that the core
of the thesis is entirely my own, namely, that the addition of the process
component to software manifestations of Conversation Theory provides a confirmation
of important, predicted and otherwise unconfirmed cognitive features. The
software written to prove this by demonstration is solely mine.
Point 2, concerning a personal history of my relationship to the Theory
and the context in which I adopted it, is fulfilled from a personal perspective
in the next chapter, and from a software development perspective in Appendix
C. It has been rewarding to reconstruct the personal side and to express,
albeit post hoc, how my career has proceeded from the ideas rather
than vice versa.
Point 1, concerning background, is less direct because the story to be told
cannot emerge as a simple narrative. The subtlety and scope of the meanings
require a hermeneutic circle. This cycle of interpretation is expressed
in the body of the dissertation as starting with my personal interest in
the Theory and related techniques (Chapter II), moves to the Foundations
of the thesis in Conversation Theory (Chapter IV), and uses the software
of the Theory to explain its elements and procedures in detail (Chapter
V). Then (Chapter VI) the limitations of past software is revealed and the
true operations emerge by the addition of the process component of the Theory.
The summary (Chapter VII) is a recapitulation of the main points and possible
extensions. This is followed by Appendices, with the technical details of
the implementations, Bibliography, Glossary and Figures. As an Annex to
this dissertation the THOUGHTSTICKER User Manual (Pangaro et al 1985
in the Bibliography) is attached for completeness.
The explication of Conversation Theory within the text is thus achieved
only upon completion of the cycle whereby the symmetries and aesthetics
noted in the beginning are achieved by an innovative approach to implementation
which fully explores the central tenets of the Theory.
I believe that all of the requirements are therefore fulfilled and I hope
the result is an effective examination of both the original Theory, and
its confirmation and extension represented by this dissertation.
I.2 A Context for AI and Cybernetics Terminologies
With the surge of interest in the field of Artificial Intelligence (AI)
primarily due to technological advances since 1980, certain concepts have
gained acceptance and comprehension within a wide audience of researchers
and software development projects in academia, industry and government.
Because of this, terms such as "knowledge elicitation", "knowledge
representation" and "machine reasoning" now have common meanings
and provide a background in which discussions in those communities may take
place (Barr & Feigenbaum 1981).
Each of these ideas had been given full treatment with detailed meaning
and context for interpretation within Conversation Theory (CT) as developed
by Pask and others (Pask 1976a, Pask 1980a), well in advance of their recent
uses within AI. Unfortunately, up to the 1980s, CT received wide exposure
only within the fields of cybernetics and computer-aided instruction. Within
those spheres and as illuminated by CT, many core concepts of epistemology
and human discourse were given tangible meanings that both reflect a common
sense usage and a precise and (within a cybernetic interpretation
of the term) scientific meaning. The terms "individual",
"conversation", "agreement" and "understanding"
are prime examples of this (Pask 1975a).
AI, engaging many more researchers and hence research publications, is perforce
divided in opinion and much more fragmented in technique than CT. (It is
an editorial comment to note that the fragmentation is evidence of the lack
of coherence and direction in the field.) Thus a perverse situation has
arisen where consistent and agreed meanings within Conversation Theory cannot
be explained using terms from AI without both distortion and ambiguity.
And, common every-day terms cannot be used unqualified to describe Conversation
Theory without losing a freshly-new and yet scientifically-strict meaning.
Therefore, although I will often draw on the metaphors of Artificial Intelligence
I will endeavor never to do so without immediately providing the significant
difference to the realm of Conversation Theory.
In general and to avoid constant qualification, references to AI do not
indicate the entire field of Artificial Intelligence, but rather those areas
within AI that relate to the subject of this thesis, namely, knowledge representation
and machine intelligence. None but the most obstreperous proponent of AI
will object to this usage.
II. Background
II.1 A Context for Adopting Conversation Theory
In 1976 I was engaged in software research projects centering around the
use of highly interactive computer graphics systems which in the present
day are taken for granted on any home video game; in those days such equipment
was extremely rare as it was only just being developed. The Architecture
Machine Group, a research facility at the Massachusetts Institute of Technology,
was producing innovative hardware systems for the creation of new types
of media environments: large screen displays, many simultaneous auxiliary
displays, touch panels and tablets for input of commands, graphics and even
gestures. The work of this laboratory has influenced a generation of workers
in the field of interactive computing. Its name must not be confused to
imply so narrow a field as mechanical architecture; rather, it was concerned
with the influence of mechanical and electronic and digital artifacts on
all aspects of the "built environment."
II.1.1 The Needs of Man-Machine Interaction
My background and interests at that time were centered on the issues of
man-machine interface (MMI) specifically for the creation of computer graphics.
These visual results might be static or dynamic, but always for the purpose
of expressing ideas, whether to oneself (as an aid to the process of design)
or to others for the purpose of communication. At MIT I had already had
the privilege of access to the newest and most powerful computer graphics
systems anywhere; what I felt was lacking was a powerful framework in which
to express the problems of MMI.
It was my conviction that to make a machine produce images representative
of abstract ideas:
- There should be a close connection between the formulation that the
user conceives on the one hand, whether in diagrams, pictures, movement,
etc., and the gestures made to the machine, whether in typing text,
programming, drawing, whatever, on the other.
- All of the power of "programming" should be available to the
designer/user, in the sense that procedures and conditional branching could
be used to great advantage, for general modelling as well as the conveniences
of repetition and variation.
The combination of these two ideas, and exposure Pask's protologic, led
to my design for a visual programming language for simulation-based graphics
of great expressive power (implemented by a research team and described
in Pangaro, McCann, Davis & Steinberg 1977, and Pangaro 1980).
One common paradigm of the era was that the human's task was to tell the
machine what was required. I however felt that this was not a complete image;
that it was also the requirement of the machine to tell the human what could
be done. These requirements were not fixed or ordered in time. They would
vary depending on the background and needs of the human as well as the machine.
The system's capabilities evolved (although not in the same time frame)
in parallel to the human's, as new versions of system code or new capabilities
were made available. Hence requirements would emerge over time, rather than
be done "all at once" at the start of the interaction. It seemed
essential to me that insofar as needs and knowledge evolve so must the interaction.
Therefore it was clear to me that a kind of teaching/learning communication
was necessary, and one which was symmetric: both the human user and the
mechanical machine had to both teach and learn.
II.2 Available Models before Conversation Theory
Obviously the interaction between human and machine was much more limited
than that between human and human; but I imagined that since one limiting
requirement (in some important aspects, perhaps the main one) was that of
the human and hence the human to human model might be a useful place to
start. Surely there was enormous history, cultural and scientific, technical
and artistic, on that subject.
II.2.1 Shannon's "Information Theory"
From the scientific community, Shannon's communication theory (Shannon &
Weaver 1964) seemed to be the only direct foray into this problem, especially
in that it was named to address this very problem. The conception here is
that communication involves a channel between entities playing roles (perhaps
alternately) as "sender" and "receiver." The concern
of the approach is to control the uncertainty with which a "message"
is transmitted across the channel. Transmission is defined as the correct
receipt of a sequence of encoded data which makes up the message. Variation
in the noise of the channel determines a statistical measure of "goodness"
of the channel. Much can be said by communication theory about the redundancy
required to insure a given and desired level of certainty about the datums
[sic] getting through unaltered.
Given the robustness of the formulation and the major concern for insuring
the accurate (indeed "perfect") data required for computers to
operate (especially in the era of the 1950s when the limits of performance
of vacuum tubes were being reached) this approach was a landmark for many
of the problems in communications and computers.
Application of this model to human conversation however is fraught with
compromise and difficulty:
- The data are exactly that, data: objective encodings or symbols
that stand alone, require no context, and are either one symbol or another,
unambiguously.
- The class or alphabet of symbols is a fixed set and cannot be expanded;
the ability to recognize one from another is predicated on the need for
both the sender and receiver to have agreed on the fixed set beforehand.
- The redundancy described exists within the encoding scheme as applied
to symbols; there is no bearing on the redundancy realized by the interpretation
of the message as a whole.
These objections individually and together remove the utility of the approach
for application to human discourse. This work remains a foundation in branches
of "communication technology", true, but at the practical level
it serves little more than to express some technical issues associated with
bit transfer in hardware channels. Weaver (in Shannon & Weaver 1964)
admits to multiple layers of interpretation to the problem of "communication
theory":
- Level A. How accurately can the symbols of communication be transmitted?(The
technical problem.)
- Level B. How precisely do the transmitted symbols convey the desired
meaning? (The semantic problem.)
- Level C. How effectively does the received meaning affect conduct in
the desired way? (The effectiveness problem.)
Weaver then says, "...[communication theory] admittedly applies ...
to problem A, namely, the technical problem of accuracy of transference
of various types of signals from sender to receiver."
CT was conceived specifically to handle Level B. Unfortunately, Weaver's
characterization of Level C does not account for "second-order cybernetics"
where the recursion over interaction produces coherent systems of belief/behavior/language
(Maturana 1978) and hence it is not comparable to the goals of CT. At the
end of the day, CT would encompass the issues referred to by Weaver in Level
C, but only in a larger context of society and culture.
It is these further interpretations of communication beyond Level A, concerned
with the "semantics" (their term) of communication, that for me
was the issue. Linguistics of the period was centered on Chomsky, who expressed
the capability as absolute and pre-existing (Chomsky 1968); this approach
did not seem to hold hope for aiding an interaction that I saw as incremental,
evolving and flexible beyond what might be programmable in genetics. Semiotics
and related work was not specific enough to provide any hints about how
to write code. Neither would psychology, so concerned with the "objective"
and "scientific" as to avoid any admittance of qualities in the
study of cognition that we call human.
II.2.2 Artificial Intelligence
Much had been said by this time about "intelligent machines";
the field of AI had already been through a number of cycles from promise
to difficulty to redefinition of promise (Feigenbaum and Feldman, 1963;
Minsky 1968; Nilsson 1980). Despite the difficulties the focus of the field
remained (and does to this day in the latter 1980s) on the "intelligence
in the machine" [sic]; little or nothing is said about
communication with such an intelligent machine or between
man and machine. And this was for me the precise focus of need as I formulated
it then: solutions to the problems of communication must be part of any
solution for machines of intelligence.
AI has always seemed to me based on an over-confidence in Turing computability
(Minsky 1967). This has been supported in arguments by Jerry Lettvin in
which he specifically ties the AI community to work by McCulloch and Pitts
on the equivalence of simplified threshold networks to Turing computability
(Lettvin 1985).
The coupling of these two mathematical results unfortunately allowed the
AI community to avoid questioning its foundations, based in the presumption
that the power of Turing mathematics is supreme (a mistake Turing himself
did not make, as I learned by examining his unpublished works at Kings College
Cambridge). This over-confidence has prevailed until recently when AI, physics
and cybernetics were united in new work to extend the definition of computability
(Deutsch 1985). These extensions were presaged by Pask (e.g. Pask
1958).
Born and raised on a mathematical formalism and whatever technological capabilities
that followed, AI could not see that it could not see its limitations (to
paraphrase von Foerster). Cybernetics was simultaneously proceeding from
an epistemological basis of what can be known and, especially in CT, moving
to theoretically sound and practical formalisms on the nature of knowing.
More detail on how CT accomplishes this is given later in this chapter.
II.3 Reasons for Adopting Conversation Theory
I was introduced to Conversation Theory first in the form of Pask himself,
who was consulting for the Architecture Machine Group. Pask had influence
there by critiquing research programmes, inventing metaphors and providing
a rich interconnection with other workers in many fields, which he brought
to a Group concerned with increasing the bandwidth (my term) of interaction
between human and machine-based systems. One tangible result was collaboration
of the entire group (and fortunately myself included) in the production
of a major work called Graphical Conversation Theory (Negroponte 1977).
This was a research proposal submitted to the US National Science Foundation,
which would interpret CT in light of the newest and most powerful computer
technology. (Alas it was not funded.)
These interactions led me to the study of Pask's papers, frequent visits
to his research laboratory, and eventually to collaboration on research
projects. It was this collaboration under contract to US and UK establishments
that funded the implementation of THOUGHTSTICKER described in this thesis.
It is a subtle task to separate out a set of personal, individual reasons
for my becoming interested in CT, or for using it as the basis for endeavors
in computing, or for using it as the foundation of a research dissertation
in cybernetics. With the understanding that any such delineation is for
descriptive purposes only, here follows an attempt to linearize what must
be, as its origins in CT would tell, a set of reasons that are ultimately
holistic and hermeneutic.
II.3.1 Symmetry across Individuals
CT restores a symmetry to the modelling of all interaction. No hierarchy
exists between, for example, teacher and learner; both must "teach"
and "learn" from the other in order to achieve communication or,
as is preferred within CT, conversation (see the Glossary for a definition).
These interactions are considered to be "I/you" referenced, because
one individual treats the other as of equal rank, in that the language is
one of command and question; the other individual has options and may or
may not respond, cooperate, etc. This provides an aesthetic as well
as an ethical formalism (Pask 1980b).
II.3.2 Symmetry within Individuals
CT models discourse within individuals by levels which are symmetric to
each other and to those in other individuals. The model stratifies any language
of discourse into distinct levels (at least from the perspective of the
observer) and creates dependencies between these levels. Thus, a "higher"
level determines the actions at the level "lower" to it.
These interactions are considered to be "it" referenced because
no choice or response is allowed by the "lower" level (Pask 1975b).
One consequence is that given a level considered to be one of "method",
the lower level is where that method is carried out. Thus, the lower level
is an "environment" for the "higher" level. Consider
that the environment may be an external world of actions, or merely further
levels of cognitive activity. This symmetry provides aesthetic satisfaction
as well as an implication that computation can encompass actions in a world
of physical objects as well as mental constructs. This interpretation served
as the basis for my design of the Expertise Tutor, a software prototype
developed under contract to the UK Admiralty which contains precisely this
multiple level of discourse and access for the user. It is the first system
of which I am aware which makes this distinction of levels both explicit
and accessible to the expert and user alike (see further description in
Appendix Section C.7).
II.3.3 Subjectivity and Objectivity in the Same
Framework
The above two points provide a brief description of the "conversational
framework", a structure in which scientific observation can be made
and descriptions of interaction may be given. The framework provides for
both "objective" interaction, where no interpretation is made
(relative, as always, to an observer) and "subjective" interaction,
where any result can be seen only from a context within the interaction
of the two (or more) individuals.
Scientific discourse has always insisted in "objective" enquiry.
It is this very insistence which has kept psychology out of the realm of
mental activity (by its own admission). CT provides, I think uniquely, a
framework in which objective, "hard-valued" measurement can be
performed in the domain of mental activity. For example, a hard-valued,
objective and scientific meaning for "agreement over an understanding"
is obtainable within CT (Pask 1975c). Because of this, the requirements
of MMI for the transfer of information about a system's capabilities and
a user's desires can be specified.
II.3.4 The Language of Conversation
The interactions described must occur in a language, and here is the crux
of any framework. If the language is a "natural" one, such as
English, immediately any machine interaction is disqualified. Although it
may appear that our present-day interaction with computers is "in English",
in fact English words and phrases are used merely as tokens to indicate
constant and pre-determined meanings. No interpretation is involved and
hence the use is not of "natural" language.
It may therefore seem that, similarly, CT is inadequate for any advances
in MMI. This is not the case for two significant reasons:
- CT as a framework is adequate for any language so long as it is one
of question and command (von Wright 1963). It may be gesture or dance, visual
or aural, images or imagery.
- The use of language tokens can be kept at a mechanical level within
the software, with the user providing a "semantic" value to it.
- Interpretation is brought in when a user relates topics together to
formulate the "meaning" of the relationship (as in the CT construct
of a "coherence", detailed in Section V). The activity is basically
hermeneutic and the meaning arises in the circular interpretation and use
of tokens by the user.
II.3.5 Generality of "Individuals"
Interactions occur across an interface, among individuals. The distinction
among individuals is made by the observer, who asserts the existence of
the interface. The individuals so distinguished and the observer can be
considered as duals of each [sic] other. The emergence of a distinction
among individuals comes at the moment of distinction between self and other,
who are the same type of individual.
In CT, individuals are "P-Individuals" or psychological individuals,
rather than a simple reduction of physical individuals. Thus a single human
may be modeled as consisting of many P-Individuals, different at different
times, for varying purposes, evolving in the course of experience. This
would encompass the requirement for a design of user interface software
which is adaptive to the changing needs of the user, in a variety of guises
and situations.
Similarly, the same framework can be used to model an interaction between
a human and machine. The specifics of processes that are available within
each are clearly different; however they can be specified and the resulting
needs for mutually-understood interaction are achieved.
II.3.6 Cognitive Bases of CT
CT was developed not out of whole cloth but from a history of empirical
research on the nature of interaction, conversation and understanding (Pask
1975a, Pask 1975c). The theory which resulted therefore incorporates its
origins into its terminology and structures. The terminology has a great
deal of "common sense" appeal and the theory provides explanatory
support for many everyday events (including forgetting, remembering, mnemonics,
confusion, ambiguity, uncertainty, and conflicting desires). This is particularly
evident when contrasted with competing theories (cf. Minsky 1986).
II.3.7 Mediation of Language by a "Knowledgebase"
Knowledgebase is a term which is used with abandon in the field of AI to
mean a structure internal to a computer which "contains knowledge"
and which can be manipulated, perhaps to make inferences or deductions,
in software. CT maintains primary interest in a "knower", while
the "knowledge" cannot be held as independent from such an individual.
CT defines related structures in its dual called Lp (pronounced "L-sub-P"
and explained in detail in the course of the text). Completely consistent
with CT and all of the points described above and below in this text, Lp
is a class of well-specified processes that operate on a class of well-specified
structures that can be adequately computed in present-day, serial digital
computers. ("Adequately" is a point taken up later and the distinction
between various levels of simulation of Lp is a central point of the entire
thesis.)
II.4 The Emergence of this Thesis
Given of the above, it seemed extraordinary that there should exist a theory
of aesthetic elegance, simple formal symmetries, based in cognitive behavior,
and with a detailed calculus of knowing that could be programmed.
Preceding sections describe the conditions under which I was introduced
to CT, especially in the context of MMI. My interests had always been considerable,
however, in the nature of cognition and communication, and how computers
may enhance or otherwise influence these daily human activities. I have
often heard others working within CT and cybernetics say that they had some
affinity or intuition for the ethos beforehand; upon introduction, the expressiveness
of the framework was immediately apparent. The simple elegance of CT to
describe mental events, and its coherence with the arts and humanities (Pask
1968, Pask 1976b) as well as sciences (Pask 1979), were a constant source
of interest for me. My initial entry from the perspective of MMI grew into
a general interest in its tenets. The desire to influence an entire field
with the sweeping power of CT by producing computer-based implementations
of its ideas became (and remains) very strong.
All of the above reasons drew me into the world of Conversation Theory and
each successive revelation within it confirmed to me its power and utility
for my interests, both theoretical and practical.
Any such software based on CT, to be useful in commercial applications (namely,
every-day use) and to be of sufficient power to influence the world's view
of MMI according to the ethics of cybernetics and CT, would require considerable
investment and relatively single-minded course of activity.
Since my first exposure to CT, I have devoted considerable time to designing
and coding software systems based on its tenets; details of my own and others'
contributions can be found in Appendix C on the history of THOUGHTSTICKER.
(That it also required the creation of a company framework is a detail of
management and of politics.) It was in the course of development of this
software that two issues converged: the need for maintaining the process
component in the simulation of the calculus of Lp; and the evolving display
of the Lp structures for the sake of the user. This discussion is taken
up in detail in Chapter VI.
This thesis returns emphasis to the importance of the process component
of Lp in any research and development centered on CT. In a very real sense,
without process the theory is lost, as one of its trilogy of features is
missing (the others being distinction and coherence). This is due to the
central role that process plays; for example, that CT states that memory
is not recall of a static configuration but a dynamic recalculation or reproduction.
All of the formal expressions of relationships with CT contain production
arrows which are not merely transformations from state to state, but continuous
processes whose continued execution and persistence is the given
cognitive element (topic, memory, concept). Hence the process component
is key, and one contribution of this thesis is the reinstatement of that
component to software manifestations of CT.
III. Review
III.1 Knowledge-Based AI Systems
Once I made the shift from principles of MMI to general theories of cognition,
it was necessary to ask the question as to whether, in all of the techniques
developed in the field called AI, some of its ideas or software results
might be appropriate, useful, or better than those of CT.
III.3.1 Semantic Nets
Semantic nets (Quillian 1968) appear on first review to be closely related
to Lp structures; an early question often posed in discussions about CT
with AI researchers is, How are Lp structures different? Because of this
apparent (but not actual) close association, a brief description follows.
(Details of Lp structures can be found in Chapters V and VI.)
Semantic nets consist of nodes and links. Nodes refer to objects or attributes,
linked by arrows (or, in programming terms, pointers) which have values
in themselves. For example, [FRED IS-A BIRD] relates the nodes FRED and
BIRD by the link IS-A. There must be many types of links, covering ideas
such as ELEMENT-OF, HAS-PART, GENERALIZATION-OF, EXAMPLE-OF, and so forth.
If in doubt, you simply create a new link willy-nilly.
The ability to create new links seems to provide for a general scheme without
boundary. However this very generality is its downfall. The class of link
types becomes very large and it rapidly becomes apparent that any subtlety
or power of the scheme is simply shifted one level into the operators that
the links represent. The nature of the computation contained in the links
(such as generalization, inference, and so forth) is not well-specified.
It is beyond the scope this text to explore this question with too many
specifics; however, there are clear differences which can be briefly listed
and which provide justification for choosing Lp above semantic nets in my
research efforts that followed the initial enquiries:
- Semantic Nets emerge primarily from programming constructs; they have
some common sense appeal but no basis in cognition or empirical research.
- There is no theory of knowing which shows that semantic nets are minimal,
necessary, sufficient, or even useful representations of human knowing.
- Further refinements (Brachman 1979) to the approach have added considerable
complexity but without achieving major advances or overcoming the objections
put forth even from its proponents (Brachman 1985).
Lp, as shown in this thesis, has none of these disadvantages, and considerable
advantages.
III.3.2 Frames
One major development (in a sense "on top of" semantic nets) is
that of Minsky's Frames (Minsky 1975). This approach accumulates semantic
relations (in the sense as shown just above) into frames of knowledge that
are related to contexts of interpretation; for example, while in a restaurant,
while going to a play, etc. This refinement handles cases of "default"
knowledge, where the scheme can attempt to fill in about items not explicitly
explained (a simple form of generalization). Unfortunately the old problems
remain; each of the above objections to semantic nets could be paraphrased
to apply to frames. Even more one is given the feeling that these are structures
that are conveniently computed by programming languages such as LISP, and
hence their popularity within AI.
Minsky has most recently revived his concept of the "Society of Minds"
theory of cognition (Minsky 1986). (This idea and others contained in a
paper called "Consciousness" were circulated privately in the
MIT AI community in the 1970s.) Basically, the society of minds puts forth
the idea that a mental organization consists of many, possibly conflicting
sub-units. These smaller units each require resources to be computed and
provide competition for the limited resources. The approach is intended
to address (what I will call) "post-Freudian" problems. These
are Freudian, because they deal at the psychological level identified by
his followers as the concern of Freud. They are also "post-" because
they are the interpretation of Freud rather than Freud himself.
Minsky offers engaging argument but neither theory nor confirmation of his
ideas. In fact, considered as metaphor and ignoring the Freudian overtones,
Society of Minds is quite consistent with CT's modelling of the "P-Individual"
being the unit of perspective within a mental organization, conversing (and
competing and conflicting) with other P-Individuals in the same organization.
Pask however provides details of:
- How the P-Individual is composed, namely, processes that can be modeled
by Eigen functions.
- The means by which they converse, namely a language capable of question
and command.
- How the interaction can be modeled, namely the "conversational
paradigm" (Pask 1975b).
- A detailed model of the structures that make up the transactions, i.e.
the Lp calculus.
- How conflict and its resolution can be modeled, via Lp and the
operations described in the main body of this thesis.
- How a continuing process of "saturation" occurs, forcing the
interaction of otherwise independent cognitive structures which in turn
creates new structures or reveals conflict, ambiguity, confusion.
Given the specificity of CT and the delightful but vague and unfulfilled
images of Society of Minds, the decision to use CT to attack problems of
my interest was a simple one. In fact, it was just such a formulation that
caused me not to pursue my work in a doctoral programme at MIT to which
I had been accepted, in order to pursue the line of research described in
this dissertation.
III.3.3 Expert Systems and Rules
Expert systems have received major attention most recently. These utilize
"production rules" in the form of "If...Then..." statements.
For example, "If the temperature is above 50 Celsius and the smoke
detector has been set off, conclude there is a fire in the room." Such
statements are said to represent the knowledge of experts, and to provide
the means to model how experts actually make decisions. Statements are processed
together to create new conditions that "fire" other rules, which
fire yet further rules, etc:
"If there is a fire in the room set off the fire alarm and the sprinklers."
Given that some rules represent desired conclusions that are distinguished
from others, the system is said to "decide." Alternatively, the
expert system can work backward from conclusions to necessary pre-conditions
and thereby diagnose initial causes.
PROLOG is a programming language designed to process these descriptions
of "knowledge", and the general approach has its origins in first-order
predicate logic. Comments about semantic nets still apply: the approach
is not based on cognitive theory or empirical studies of human knowledge;
the scheme is not known either to be sufficient or necessary to explain
human cognition; and extensions do not solve fundamental problems with the
approach. Expert systems have recently become a popular means to approach
the problems of training, wherein tutorial strategies are encoded as "If...Then..."
rules: "If the student has failed test A and test B then conclude topic
X not understood." At some point there should be a general recognition
of this fashion as no better than an intricate but equally ineffective form
of training as programmed instruction (in the same way that programmed instruction
is now widely recognized to be an impoverished technique of computer-aided
instruction; see Section III.3.)
Other AI approaches are more tangential to the requirements of a cognitive
approach to software and MMI design. Work in natural language parsing is
still focused largely on translation and getting knowledge "into"
a knowledgebase (Barr & Feigenbaum 1981). Shank's work on Scripts (Shank
& Abelson 1975) has some interest in communicating with users, but although
he takes an increasingly iconoclastic view of other approaches within AI
(Shank 1980) his alternatives are still within AI's limitations. Winograd
is the closest to a cybernetic view but his publications do not provide
a sufficiently tangible alternative to begin coding (Winograd & Flores
1986).
III.2 Related Work in Cybernetics
Despite its popular associations with robots, cybernetics does not of itself
refer to computers. A surprisingly small amount of work in cybernetics overlaps
with, or has produced approaches to, MMI or conversational software.
III.2.1 Foundations
At the theoretical level, related work in cybernetics has generally been
a precursor to CT and/or provided a foundation upon which CT could provide
the specific results that it does (for example, von Foerster 1960). The
emergence of second-order (retold in von Foerster 1985) and reflexive interpretations
of science (Bateson 1960) provides the beginnings within cybernetics of
an approach to systems that is both scientific and subjective. However
these foundations require interpretations, in both empirical studies and
detailed formulations, before they can be translated into tangible prescriptions
for action, which came only with Pask.
III.2.2 Laing
The utility of a reflexive view of interaction (again on the theme of the
problems of MMI as discussed above) is most effectively presented in Laing
1966. The interpretation in the context of conventional MMI would be something
like "I [the user] know what functions the system knows. The system
knows nothing about me." A more advantageous approach which I desired
would be something like "I [the user] know that the system knows what
I know about the system." This could perhaps be extended to incorporate
goals, as in "I [the user] know that the system knows my goal is to
..." Thus the user could proceed with greater confidence and efficiency.
Laing thus provides a metaphor of desire, but nothing detailed on which
to base a software approach.
III.2.3 Personal Construct Theory
In terms of software, the work of Kelly in the extraction of grids of constructs
for purposes of explicating knowledge otherwise internal to a knower (Kelly
1966, Bannister & Mair 1968) is closely related to the interests of
CT. This has powerful implications as seen in practical and modern software
implementations (Personal Construct Theory and the software Pegasus, in
Shaw 1980; and MAUD software, Humphreys 1975).
In these latter two cases, the software is used as a means of extracting
constructs internal to the knower, and in a form which is self-consistent.
This exactly parallels one of the intentions behind Lp, where the names
of topics and the relations that they are contained in are delineated and
named by the user. The software, again as in Lp, is used to reflect back
to the user on the implications of the constructs and their structures,
as for example in cases of ambiguity and contradiction (Humphreys 1980).
These other approaches both preserve the subjective quality of the "extracted
knowledge" and emphasize the self-consistency of the result.
However, Lp additionally provides a framework that is based on the epistemology
of observation, empirical confirmation of the utility of its references
to individual learning style (independently confirmed by Marante & Laurillard
1981, and Bogner 1986), and an extended set of operations which encompass
many more events that are recognized as cognitive (Pask 1983).
III.3 THOUGHTSTICKER and Computer-aided Instruction
Computer-aided instruction (CAI) has been widely available on computers
since the advent of minis and micros, starting in the 1970s. Largely accepted
as useful tools for training by computer, some criticisms have arisen over
the years (see Kearsley 1977 for a view inside the field, and also Pask
1972). The following section presents a self-contained explication of how
THOUGHTSTICKER can be applied to the problems of a user learning from computers,
in direct comparison to existing software training approaches. THOUGHTSTICKER
represents a complete revision of all existing techniques.
III.3.1 Intelligent" Training
THOUGHTSTICKER is an intelligent software system for training and information
management. The system is "intelligent" in the sense that it mediates
between an expert knowledgebase and a user to provide some of the features
of human conversation: a shared vocabulary, history and context of the dialogue.
It is the most effective system of its kind available on any hardware.
THOUGHTSTICKER was developed as an enhancement to conventional computer-based
training (CBT) and provides substantial improvements to CBT in:
- Ease of use, for both courseware creation and delivery of training
- Management of the courseware creation process
- Sensitivity to individual learning style
- Training efficiency and effectiveness, especially in complex tasks
- Flexibility to encompass job-aiding and advising, as well as training.
The software consists of two independent parts: the means for creating the
knowledgebase (the Authoring Module); and for giving access to the knowledgebase
(the Tutoring Module or Browser). Both are conceived and implemented as
generic solutions that can be tailored to the specific requirements of the
application, its users, the target hardware and interactive media (including
videodisc, CD-ROM, graphics and sound). THOUGHTSTICKER is attached easily
to existing application software and simulations for a complete training
solution.
III.3.2 Background of the Term "THOUGHTSTICKER"
The term THOUGHTSTICKER refers to software based on a cybernetic approach
to the problem of measuring understanding in human conversations. In the
1970s, THOUGHTSTICKER was developed at Pask's laboratory as an extension
of Pask's studies of the 1950s and 1960s in human learning and individual
conceptual style. These studies culminated in a comprehensive approach to
educational technology (the CASTE system) that has been widely influential
in educational theory and computer-aided instruction.
The term THOUGHTSTICKER was coined by Pask to mark the maturation of a general
approach to knowledge representation whose elements reflected cognitive
structures. The name itself emphasizes that in order to converse we must
externalize our thoughts into a tangible form for ourselves and for others.
Using a computer as the medium for this conversation means that thoughts
must temporarily take a static form in the computer, before becoming dynamic
again as they are interpreted by a user. THOUGHTSTICKER models mental structures
with a few simple but powerful constructs that:
- Capture the author's or expert's precise approach to the subject matter,
but still
- Allow the user to learn the subject matter according to his or her conceptual
style.
THOUGHTSTICKER software is the medium for the conversation, not a participant.
The power of THOUGHTSTICKER derives from:
- A theoretical basis in cybernetics and learning theory. The advantages
of Conversation Theory as a model for learning have been supported by experiments
in cognitive style. THOUGHTSTICKER is derived directly from these ideas.
- Evolutionary development in application to complex training problems,
including those with training in the performance of a task. Extensions for
job aiding and expert advising have also been demonstrated.
III.3.3 Existing Applications
Prototype knowledgebases have been constructed by me and my colleagues in
PANGARO Incorporated, and in the subjects of AI and cybernetics, naval strategy,
introduction to computer usage, and word processing.
For the Behavioural Science Division of the UK Admiralty, THOUGHTSTICKER
has been integrated into an Expertise Tutor, consisting of a naval simulation
and expert knowledgebase (described in Appendix C). The Tutor provides tactical
training as well as basic rules and operations of the game. This system
is effective because it provides the user with equal access to descriptive
knowledge (elements, relations, goals), prescriptive knowledge (methods,
tactics), and the environment (the simulation itself).
For the US Army Research Institute, a videodisc interface controlled by
THOUGHTSTICKER has been developed to demonstrate training of a vehicle identification
task.
Most recently a prototype training course has been developed for Symbolics
Education Services. (Symbolics, Inc. is the manufacturer of advanced software
engineering and Artificial Intelligence workstations; the most advanced
implementation of THOUGHTSTICKER runs on this hardware.) Derived from an
introductory, paper-based workbook written by Education Services, this course
presents the basic components of the Symbolics computer, concepts of symbolic
processing, and how to use certain features of the machine such as the editor
and command processor. The learner can immediately practice what is to be
learned via the "hands-on" capability: in the course of
learning about the editor (for example) the editor window is automatically
displayed and commands may be tried step-by-step by the learner concurrently
with their presentation in the training material.
III.3.4 The User Experience
THOUGHTSTICKER facilitates the user in any training and information
management activities by:
- Allowing a mixed-initiative dialogue so that the user may either give
the system control, or direct the conversation based on immediate
needs (e.g. uncertainty or current goal).
- Producing distinctly different actions and responses for different individuals,
based on the background, purposes, context and cognitive style of the user.
Thus the user is provided with more focused and efficient interaction than
conventional computer-aided instruction and information management systems.
These results can be achieved because THOUGHTSTICKER "models the user"
throughout the interaction, creating a history with each individual that
is maintained even across sessions. Because this user model is the basis
of all actions by THOUGHTSTICKER, the interaction has more of the qualities
of human conversation: context, focus, and shared vocabulary.
III.3.5 Comparison to Computer-Aided Instruction
The following two pages contain a brief, "side-by-side" comparison
of conventional computer-aided instruction techniques and THOUGHTSTICKER.
Conventional Computer-Aided Instruction THOUGHTSTICKER
Based on concepts of "programmed Based on a cognitive theory of human
instruction" developed in the 1950s conversation developed over the period
and substantially unchanged since then. of 1955 to the present, and affirmed
in empirical studies.
The subject matter is given a Uses a robust knowledge representation
pre-ordained sequence in which it is scheme to provide a true
to be learned; there is no other knowledgebase; all conceptual
structure to the material. dependencies are represented in a
network structure with no fixed paths.
All users are treated identically, and Sensitive to an individual's cognitive
thereby are presumed to have the same style, modifying responses accordingly.
cognitive learning style.
The author of the subject matter makes Sensitive to individual variation in
assumptions of prior knowledge of the user's prior knowledge and can be
user; very little variation of tuned by a variety of user profiles
material is possible despite differing (for example, naive computer users;
backgrounds in the user population. experienced computer users but not of
this particular type; users of another
particular vendors' hardware).
Additional questioning by the user is User is free to ask questions and
limited or not allowed. Remedial explore throughout the knowledgebase
material is offered to the user upon at any time. The user helps direct the
supposition of reasons for user's remedial dialogue, which is derived
failure and usually from a static from a combination of user's focus,
model based on averages or likelihood. the structure of the knowledgebase,
and the history of the interaction.
The comments about computer-aided instruction can stand as generalizations
across a number of commercial products because they characterize a field
which is substantially homogeneous. Although specific features of
training packages vary, the instructional model and the organization of
the subject matter does not.
The driving force of the interaction is the user's interests and uncertainties.
THOUGHTSTICKER has specific features that help the user discover these interests
and uncertainties, and then explore or resolve them. By allowing such strong
initiative on the part of the user, THOUGHTSTICKER provides an effective,
efficient and supportive training experience.
III.3.6 THOUGHTSTICKER's Training "Knowledgebase"
THOUGHTSTICKER is constructed as a generalized information management system.
Its internal database, called a knowledge representation or knowledgebase
in modern parlance, supplies a flexible, "relational" format that
is suitable for any subject matter.
To describe the format briefly (to be detailed later in Section V): topics
are defined and associated in relations by the author or expert.
These objects together define a network or mesh of "knowledge"
and thus determine the structure of the knowledgebase. There are no pre-defined
types of relation; the author is free to create relations as desired. THOUGHTSTICKER
contains training heuristics, many concerned with the user's purpose and
conceptual style, for moving over this structure. The conditions which determine
the action of these heuristics are:
- The User Profile: A preset stereotype of the background of the trainee.
The author pre-determines what classes of users are expected to interact
with THOUGHTSTICKER. For example, these classes may represent a particular
range: novices at a particular task, individuals with some exposure to comparable
tasks, and experts. The User Profile can be styled by the author as a single
default state, or chosen from a descriptive list by the user, or determined
with highest accuracy and detail from a pre-test. Given such a User Profile
for a particular user, the choices THOUGHTSTICKER makes are more directed
to that individual's level. However, the Profile is only a starting basis
and the two mechanisms described next provide further refinement of THOUGHTSTICKER's
actions.
- The User History: A tracking of all actions and results since the user
started, whether at the present session on the machine or in the user's
history with THOUGHTSTICKER over time. The history consists of, among
other details, a record of terms used by the user and the system, topics
and explanations shown, and the current context of conversation. This shared
history is used by THOUGHTSTICKER at each moment to choose an explanation
or a new focus of attention. The result is more directed for the user and
hence more efficient and satisfying. The disk requirement for storing this
User History is modest.
- The User Model: A representation of the user's conceptual learning style.
As in the User History, the User Model influences THOUGHTSTICKER's choices
at each moment, but by applying criteria associated with the user's preferred
modes of learning. For example, these may include a preference for examples
before general descriptions; or preference for thoroughly completing current
areas of learning before touching on new areas; or preference for graphics
over text. The User Model can be configured by the author, the user, or
by the results of a pre-test. It can even be modified on the fly, provided
the user is imposed upon to give feedback on the effectiveness of explanations.
In addition, the User Model may include the broader components of the user's
purpose. Thus THOUGHTSTICKER can respond differently if the user
wishes to learn the entire subject, or the performance of a specific task,
or a single precise command name.
III.3.7 Aids to Authoring
It is widely reported that the major expense in using computer-aided instruction
is the cost of "authoring" the material, that is, creating the
subject matter that the learner is to see.
Conventional computer-aided instruction provides basic utilities for the
creation of text and graphics to be assembled into frames for the user to
view. In addition, features for managing the user's records, keeping statistics
across groups, etc., are generally available.
Like conventional CBT, THOUGHTSTICKER can provide any "management"
functions relevant to a particular site; for example, tracking a student
population, creating output reports, or collecting feedback on the effectiveness
of any aspects of the course. These requirements are best defined for the
specific needs of a CBT application, and tailored accordingly.
Unlike CBT, THOUGHTSTICKER is exceptionally strong in providing tools for
creation and maintenance of the knowledgebase. The power of its environment
for providing such features, utilizing the bit-map display, menus, the mouse,
etc., is unrivaled. The author uses a full-feature editor to create
text material to be integrated into the knowledgebase. Graphics functions
or particular devices (such as videodisc) can also be provided for specific
training areas. A variety of tools provide views of the resulting structure
and show the implications for the learner. In addition, semi-automatic tools
are used to convert pre-existing, machine-readable text of the subject matter
into THOUGHTSTICKER data files.
Conventional computer-aided instruction systems provide authoring tools
that are basically passive so far as the content of the presentation to
the learner is concerned. THOUGHTSTICKER provides a number of active
tools that facilitate the authoring process:
- THOUGHTSTICKER suggests key topics by which to represent the explanation
in the knowledgebase; it searches the text as provided by the author, looking
for variations and similar terms in the current author's, as well as other
authors', knowledgebase.
- THOUGHTSTICKER checks the existing knowledgebase of all authors and
reports how its contents relate to the new statement. It suggests how the
statements might be related (identical, containing, contained, etc.).
- In certain cases THOUGHTSTICKER can detect a possible conflict between
statements (technically speaking, it does this not by the semantics of the
text but the structures of the knowledgebase the text expresses;
- THOUGHTSTICKER does not yet contain natural language processing). The
system offers a series of methods to resolve the conflict depending on the
structures: statements may be declared "not accepted", they may
be merged with others, distinctions may be added, etc.
- In all cases the author's input is tagged to that author and other key
parameters such as time of entry. Some THOUGHTSTICKER user interfaces provide
the identity of the author at all times; others display it when the distinction
is required. Any authors' denials of a statement are also so tagged, and
hence many-valued disagreement and consensus may be stored. (A denial is
the modification of a statement relative to a user, as to whether that user
accepts the statement as valid or not. This applies the user's own statement
as well as to the statements of other users.) In this way, local extension
or modification of the contents of the knowledgebase is easily achieved
while still preserving the original.
- To stimulate the author to add further structure and material to the
knowledgebase, THOUGHTSTICKER will propose new structures which do not yet
exist and which, if instated by the author, will not conflict with
existing structures. This process can be focused by having the author indicate
areas to extend or areas to avoid. Alternatively, THOUGHTSTICKER can suggest
areas that are "thin" compared to others; in this way the author
is encouraged to achieve a uniform level of detail.
Unique to THOUGHTSTICKER, the combination of these features make the process
of creating the subject matter much more efficient. In addition, multiple
authors, possibly at different sites, can contribute to the same knowledgebase
without interfering with each other. The original knowledgebase can be augmented
and tailored to differing needs at different locations.
III.4 Related Software Systems
THOUGHTSTICKER, Lp and dynamic graphics displays of knowledge representations
have implications for the domain of software systems as they are now presented
in both commercial systems and research programmes. The work of this dissertation
presents innovations in these areas, briefly discussed in the following
sections.
III.4.1 Database Management Systems
The concept of a database is a simple one: to store and index data in a
form for swift and convenient retrieval and update.
Approaches to database management come in various forms; the most flexible
of which is the most complex to implement but also the most general and
most useful. These "relational database" concepts find mature
implementations in modern, commercial database programs available on computers
from large to small. Their power derives from a complete flexibility in
how the data is indexed: truly relational systems can be indexed
on any entry in the database. This corresponds to THOUGHTSTICKER's capability
for every topic object to be accessible directly, and for any relations
that topics exist in to be used as a means to move from relation to relation.
Details of implementation aside, THOUGHTSTICKER is functionally a complete
relational database.
Consider that database connections are arbitrary and unconstrained;
THOUGHTSTICKER provides structures that model cognitive relationships. Hence,
the result may be considered a "knowledge representation" or "knowledgebase"
rather than a mere database. Of course in both cases the data contained
must be interpreted by a human to make it alive with meaning and become
true "information"; however in the case of THOUGHTSTICKER, the
structure reflects contextual relationships that are valid in the construct
of the creator or author of the structure.
For the user, the nature of the two systems (relational databases and THOUGHTSTICKER)
is completely different. Database require that the statement by the user
be a "well-formed expression" which can be syntactically parsed
and interpreted logically. For example,
(AND (OR (subject = cybernetics) (subject = protologics)) (type = thesis))
would retrieve all records on the subject cybernetics or protologics which
are a thesis. Modern systems often employ pseudo natural language input
schemes, whereby the same search could be performed by typing, literally,
Show all records of the subject cybernetics or protologics, that are
a thesis.
THOUGHTSTICKER in its present forms is not tailored to perform precisely
this type of search. However, it can be used in such a mode by two of its
mechanisms:
- It can use features of the entries, such as type of entry and contents
of the relation to prefer or exclude some entries over others.
- The course of the conversation includes a context of previous requests
which are used by THOUGHTSTICKER to determine what data is retrieved.
Note that the relational database search is independent of all previous
and future searches; it is without context. THOUGHTSTICKER, by contrast,
builds a history of interaction by tracking all requests and modifying subsequent
responses. For example, first a request for cybernetics as a topic would
recall all such available entries (i.e. relations and their models;
see Section V for full explanations of these terms). A second request for
the topics protologics would first provide those relations which overlapped
or were close to cybernetics; thereafter, entries that were individually
related to one or the other. Finally, a request for entries on the topic
thesis would first retrieve entries that are in cybernetics or protologics.
Thus the retrieval is not a one-shot and without context, but rather an
emerging purpose that is created by the history of requests on the user's
part. The result is less immediate. (Of course the conventional search patterns
could be added as a capability to THOUGHTSTICKER, making it a subset of
relational database systems.) However, for application to research where
a fixed answer is not sought but rather a picture is to emerge over a series
of refined retrievals (which database retrieval usually is) THOUGHTSTICKER
holds great promise as a revision to the nature of database management.
The free form manner in which statements are added into the database and
the lack of restriction on "keys" are substantial improvements.
III.4.2 "Thought Processors"
There was a brief flurry of interest in commercial personal computer markets
for programs that were erroneously dubbed "thought processors."
In fact, each of these were merely word processors with a fixed format for
creating outlines. With the appropriate command, a given line in the outline
could be expanded to contain sub-lines. The process is fully recursive.
The resulting outline would make an excellent basis for writing a full document;
hence the claim of "thought processing", which here means instead
to help plan the writing process.
THOUGHTSTICKER is rather more like a true thought processor because of its
power in cognitive modelling. The structures that result and the process
of conflict resolution are a strong partners in the thinking process. In
fact, THOUGHTSTICKER is to modelling thought processes as word processing
is to writing. No other commercial or research software can make such a
claim.
IV. Foundations of Conversation Theory
This chapter provides background and the bases of the argument of the thesis.
A very brief synopsis of this chapter was the content of the Abstract.
IV.1 Interaction and Conflict
Conversation Theory is a theory of interaction. The minimum psychological
observable is that of an interaction between two distinguishable entities,
the distinction of which is made by an observer (Pask 1975b, Pask 1975c).
The role of the observer and the interaction are so inextricably linked
that they are duals; one does not exist without the other (Pask 1976c, Pask
1980c). Hence, from interaction arises all individuals, all distinctions
and therefore all "conceptions." These make up (or "inhabit")
the organization of systems as a whole.
The existence of a distinct entity is an observer phenomenon that is consistent
with other distinction logics (Varela 1975). The persistence of an entity
is the result of a convergent process rather than, for example, the physical
existence of a mass (von Foerster 1977).
A range of possible types of interaction arise within and among systems.
Trivial interaction is that which is consistent to and meshes smoothly with
the existing organization and therefore merely reinforces that organization.
Information introduced into a system which is not "novel" is an
example of this (Pask 1980b). The crucial case is when the information introduced
is novel so far as the system under scrutiny is concerned. Hence,
if interaction is to allow for change and evolution of organization, it
must perforce consist of occasions where processes (by definition, programs
that are executed in one or more processors, Pask 1980b) are not mutually
consistent and do not smoothly mesh and where the organization is in danger
of change. This is conflict. Without conflict, the organization cannot
undergo change.
One outcome of conflict can be destruction of the organization. Alternatively,
if concepts and individuals are to endure under the influence of conflict,
it is necessary that conflict be resolved, with the accompanying persistence
of organization albeit a modified one.
IV.2 The Requirements of Representation
Conversation Theory arose in the context of learning environments where
the subject matter to be learned required a representation outside of the
human subject matter expert. The independence of knowledge (or more precisely,
"knowables") from a knower is an absurdity which is often mooted
for the purposes of practical implementation in current digital machines,
and for the sake of discourse. Hence it is a simple error to lose this point.
Current AI research, of course, is predicated on the possibility of knowables
without a knower and the nature of this contradiction is not always acknowledged
(as it is in Dreyfuss & Dreyfuss 1986, and Winograd & Flores 1986).
CT at all points re-affirms the role of the knower.
In addition, other aspects of this epistemological stance of CT imposes
certain requirements on the needed knowledge representation, requirements
which AI has generally not benefitted from. Communicability, stability (memory),
ambiguity and its resolution are all central to cognition and a knowledge
representation based on CT must encompass these issues.
IV.3 The Rise of Lp: Coherence, Distinction and
Process
The needs of a knowledge representation as constrained by CT led Pask to
invent the protologic called Lp. The term "Lp" arose in context
where CT had already been concerned with descriptions in a language called
"L". Pask's work had previously involved a formalism containing
"L", a symbol standing for any true language (natural, spoken
languages as well as the language of dance, gesture, or signs). The requirement
was that L have the capacity for questions and commands as well as statements
and possibilities. (Classical mathematics and predicate logic does not;
see von Wright, 1963 and more recent echoes in Winograd & Flores 1986.)
Because the representation underlying cognition was more primitive than
that language (in the sense that all languages could be modeled with a common
structure and kinetics), Pask added the "p" subscript, meaning
"proto" (meaning "primitive" or "original",
as in a substrate). Lp is therefore a substrate or structuralism [sic]
on which would rest a logic or language to carry the richness of
human discourse.
Lp describes the interaction of conceptual entities by providing rules that
constrain the interaction of these entities and hence model their evolving
organization. These interactions are described at the level of concepts,
that is, as the interaction of topics in relations that form conceptions.
Lp contains injunctions as to how topics can and may interact, including
how they may conflict and how their conflict may be resolved.
In its pure form, Lp is a logic of process as well as coherence and distinction:
- Process, in that all entities are the result of
- interactions, where an entity is that which is stable and recognizable.
- Distinction, in that there arise in the course of the interaction of
processes, entities that did not exist before and that are distinguishable
from one another by further processes.
- And coherence, in that conceptual entities "cohere" together:
their dynamics are such that their process interaction creates stabilities,
themselves conceptual extensions of the original.
Due to their characteristics (such as their kinetics, leading to their stability)
Lp entities imply models for memory, uncertainty and innovation.
IV.4 The Distinction of Micro and Macro
The attribution of a term such as micro or macro is made by an observer
relative to some purpose. In the context of software simulation, it refers
to the "grain" at which elements are chosen as primitive, and
the relations between elements are simulated by procedures which related
them.
To choose a level of description and to name it "macro" is equivalent
to stating that there will exist some elements of the database which will
be considered indivisible atoms and processes below a certain level will
be asserted rather than acted upon. In the present case of THOUGHTSTICKER
software as described below, the topics of Lp will be considered as atoms,
and their relationship will be asserted to be dynamic but represented
as a static structure. This is not a condemnation; it is nothing more
than a proper declaration of the status of the database elements, and it
serves to clarify the observer's intentions in the nomenclature of declaring
it to be "macro."
For some purposes, such as tutorial representations of subject matter as
detailed below, such a "macro" description of topics is sufficient,
because the relationship among topics is to be activated by the user, and
the mere existence of their relation is sufficient from the perspective
of the software.
For other purposes, this level of grain of the simulation may not
be sufficient. In particular, it cannot represent the true implications
of CT as a model of the dynamics of mentation. It requires a process interpretation
of the structures of Lp. This can be achieved by an increasing series of
more detailed simulations. The first of this series retains topics as atoms,
but provides a process relationship between them (this being the main topic
of the remainder of the thesis). A second would be to break the topics down
into sub-components, thereby exposing their "internal" structure
to scrutiny. This is unnecessary for demonstration of the thesis and is
not explored further. However it is appropriate to comment that such an
extension of the simulation would be necessary to provide additional evidence
in support of the more subtle implications of Lp, in its power to model
generalizations of concepts and their creation in abduction.
IV.5 Static, Macro Representations of Lp
All previous software programs based in some way upon Lp operations have
used a description of Lp that encompasses coherence and distinction only.
The level of description of these programs has been that of the topics
(considered as indivisible atoms); and a level "higher" than the
topics themselves, namely, a level of topic relations.
The topics are represented as static elements in a database; they exist
not by nature of a process which is executed but rather because of a configuration
of 0/1, binary data in a static software structure. Similarly relations
are static aggregates of tokens associating (either by means of pointer
structures or common names) the topics they relate. It may be tempting to
consider that in order for these entities to be used by the digital machine,
a "process" in the form of a program is executed by the digital
machine to access them, and that this is sufficient to achieve "process
interaction." However, this misses the crucial point that the topics
and their relations in a true Lp processor are embodied because they
are executed as processes, and their attributes and interactions arise
from execution; not because they are being accessed as a static token. The
qualities of the entities are the result of execution and not simple reference
to a list of static attributes. Process interaction can only be simulated
in a serial digital machine by pairwise checking; in a true Lp processor,
the medium in which the processes are executed (here unspecified) also affords
the means for their interaction.
In existing software implementation of Lp, conflict is detected by a simple
counting and comparison scheme. The software makes reference to the static
data structures and conditions for conflict (described in Section V.4 on
THOUGHTSTICKER) are calculated.
It is important to realize that this calculation is performed by a program
that has available to it all necessary information of the organization of
the system as a whole. It is a privileged position which is akin to a global
or "god-like" view. It is therefore a position taken by a process
that is independent of the system itself (in the sense that an observer
is outside the system). Because it is independent, the calculation in this
form cannot be performed in this way by the system itself.
Because of this view, the level at which the dynamic interaction of the
topics is simulated, is here called "macro." For restricted applications
of Lp, such as in a knowledge representation scheme for training, this may
be sufficient.
IV.6 Deficiencies of the Macro
However, the macro position has two deficiencies: the fundamental tenet
of Conversation Theory, that of true process, is missed; and, conflict
does not arise internal to the system, but rather is computed external of
[sic] the system; that is, macroscopically. It may be seductive to
say that the existence of conflict can be denigrated and trivialized to
a mere artifact of the level of description and its historical origins in
subject matter representation. However the rise of conflict within
systems must be recognized for its power to model the initiation of distinctions,
and hence as a powerful engine for innovation arising within the organization
of a system.
Without a process component, there is no "available energy" for
the system, and further mechanisms would need to be hypothesized. With process,
the entire theory holds together in a consistent manner.
IV.7 Hypothesis: Theory Confirmation in Micro
Simulation
All theories consist of descriptions in a language. A description may imply
or produce a further description which in science is often called a "result"
of the theory. More precisely, such further descriptions are hypotheses
or hypothetical statements that are deduced from the body of the theory.
A hypothesis or "theoretical result" is usually compared to observations
of some environment and when correlations exist the theory is, in some part,
"confirmed."
The hypothesis put forth in this dissertation is that a low-level description
of Lp, that of an internal and microscopic level in which topics are influenced
by "forces" that are exerted by the topology of the conceptual
space, would, in its activation as a dynamic process of appropriate dimension,
produce as a result (and hence provide a confirmation of) the macroscopically-observed
behavior of the system manifest as conflict and resolution of conflict.
The resulting implementation would reify a system modeled in Lp by producing
a system whose topological space was constrained by the interaction of its
entities.
This has some similarities to the work in new "quantum computability"
(Deutsch 1985), which is another revision of "classical" computational
theory (i.e. that attributed to Turing). There as here the desire
is to achieve certain classes of computation which otherwise would not
be possible; in particular, computation which would not be possible
in any "Turing architecture" consisting of a finite state machine
and a tape (memory). This position is in sharp contrast to the historical
view of Turing computability as sufficient for any class of finite computation,
including that of brain (for an excellent discussion of the interactions
between these views see Lettvin 1985). Since the digital computer is based
on the Turing model, it was considered just a matter of engineering before
computers were smart like humans. The revision requires new hardware architectures.
IV.8 Dynamic, Micro Representation of Lp
To return to the original intentions of Lp as founded on process, a new
software model must be put forward to reify Lp structures. Unfortunately
there are fundamental limitations presented by present-day serial, digital
machines and a true Lp embodiment must await new architectures (which are
beginning to appear, see Section VII.3.4). However the basis for a new approach
can be set out now, and simulated in current hardware. The software written
for this dissertation, although a simulation, restores the process component
to Lp embodiments and provides a direction for future work in fully concurrent
machines.
In brief (as the details will be presented in Chapter VI), the individual
entities that exist within an Lp structure exist due to the execution of
a process rather than their existence in a static database. This can be
simulated within serial machines if the interactions (relations) between
entities (topics) are expressed as a continuous computation of relationships
within a topological space. These relationships are represented to the observer
as relative positions on a display screen.
The topics themselves are atomic units and not processes whose execution
result in stable (but dynamic) entities. For the purposes of practical implementations,
some level of "atom" must be chosen to begin the simulation. But
while this is the case, their relations as manifest in a graphical
display exist due to the interactions of processes. These processes individually
are the action and interaction of each entity within the organization of
the system under execution.
Interpreted graphically in this way, topics compute their positions relative
to their neighbors in relations that they share. The computation is performed
in accordance with Lp rules. The resulting positions, which may or may not
be stable, represent the conceptual relationships of the topics relative
to each other.
Within the simulation of the micro interactions of Lp, macro features of
CT should emerge. For example, for certain initial configurations ambiguity
or contradiction should be detected. This prediction is confirmed, as will
be shown in Section VI.5.
The next major Section presents a detailed view of THOUGHTSTICKER software
at the macro level, which is a necessary precursor to discussion of the
micro of Lp and results.
V. Conversation Theory Software
V.1 The Birth of "THOUGHTSTICKER"
As noted in Section III.3.2, THOUGHTSTICKER was invented by Pask and collaborators
at System Research Ltd in the late 1970s (Pask 1976a). Its development was
more or less coincident with the development of Lp in that the creation
of THOUGHTSTICKER (as a software manifestation) both fed on and was fed
by development of Lp (its formal and notational basis). However, and as
already noted, to be a full-blown logic for CT, Lp required a bifurcation
principle (described below). Pask has stated that this was available only
after the enquiries of Vittorio Midoro about the notation of analogy and
distributive coherence (an overlap of a single topic in more than one relation)
co-existing in the same diagram. Pask had already realized that some principle
was needed to explain how new structures arise from computation performed
from within a system. This, along with the principles of conservation, duality
and complementarity already formulated, would make CT a complete, "scientific"
theory. Midoro's enquiry led to the simple and elegant bifurcation principle
that shows how distinctions and hence new structures arise from within an
organization. Midoro was at the University of Genoa at the time, and hence
Pask has called the bifurcation rule the "Rule of Genoa."
It will be demonstrated that THOUGHTSTICKER in all of its software forms
represents the embodiment of Lp at a macro level, an argument that
is one foundation of this thesis. To further clarify what is meant by this
and to provide necessary background a detailed description of THOUGHTSTICKER
follows.
V.1.1 Raison d'Etre of THOUGHTSTICKER
The background to the THOUGHTSTICKER system may be seen as two,
concurrent, threads:
- The development of Conversation Theory as a scientific and psychological
model for knowledge and beliefs; and
- To "compute" knowledge structures inside of presently-available
digital machines, a goal that is analogous to the attempts of AI.
This chapter focuses on the former, although some comments on the latter
are necessary.
V.1.2 Represent What?
Both AI and cybernetics have encountered the same dilemma, albeit from quite
different paths:
- What is knowledge that it may be represented in a concrete structure;
and
- What is a representation that it may reflect the process of knowing?
Without question the issues here are very deep and are properly treated
in other places, the literature of philosophy being one (in a monograph
particularly focusing on CT and Lp, see also Nicoll 1985).
In the context of the use of computers in human decision making situations,
it can be shown why the issue arises at all by the aid of the following
parable: To help humans to perform calculations such as check-book balancing,
word processing and orbits of satellites, the computer must manipulate with
facility the elements of these domains, such as numbers, representations
of text, calculations under specified equations, and so on. Without this
capability, the computer would be useless for these tasks.
Similarly, for the computer system to provide any help, support, advice,
what have you, in the "thinking process" (alias the "decision
making" process) it must manipulate with facility the elements of the
domain: the knowledge of the user. This presumes, not unreasonably, that
the computer is to calculate a domain beyond mere numbers and measures.
The domain becomes the non-quantitative, non-specific and often inchoate
world of beliefs, conceptions and impressions. (See comments in Section
III.4.2 concerning so-called "thought processors.")
V.1.3 Attempts at Knowledge Representation: "Expert
Systems"
The goals of knowledge representation are easily said, but not so easily
done. The 25 year history of AI has attempted to deal with these issues
from the "bottom up": starting from the computer technology and
a reductionist view of mental processes. Some consider that the process
has borne fruit (Michie 1982; Feigenbaum & McCorduck 1983) while even
the most impressive of so-called "expert-systems" are limited
in the extreme (Duda & Shortcliffe 1983).
The "expert system" paradigm is one which considers that the "knowledge"
of experts may be captured by a manual process, and converted into a form
computable by present-day computers. This manual conversion is performed
by a "knowledge engineer" who codes the "expertise"
into rules which are easily calculated over by the digital engine. The tribulations
and disadvantages of such a presumptive approach have been discussed elsewhere,
in Pangaro & Nicoll 1983.
For our purposes here, it is sufficient to point out that it may be possible
to divide the global problem of the reification of knowledge into two stages:
- Capturing a representation which is useful to the user and to others
but which cannot be computed over by the digital engine, that is, the computer
does not know; and
- Elaborating the structure of representation so that the digital engine
may itself perform (e.g. decide) in a manner which reflects somewhat the
form as well as the content of the original human thinker. Humberto Maturana
has made the point (Maturana 1986) that going the route of representation
separate from ontology is a fundamental misunderstanding of the nature of
knowing and any attempts in that direction are doomed. His position is beyond
the compass of this discussion, as the center of this thesis are the issues
of demonstration and confirmation; hence representation is desired.
Expert systems and AI in general attempt the second, and more difficult
stage, first. The goal of pragmatic research programs (Sheppard 1981; Pask
1981) is the former, with a clear plan of extension into the latter, as
techniques and technology catch up to the more demanding requirements
of "machine intelligence."
This claim of capability (for it is only a claim until demonstrated in working
systems) is based on the substantial theoretical and experimental work which
Conversation Theory represents as embodied, within its limitations, in the
software system called THOUGHTSTICKER.
V.2 The origins of THOUGHTSTICKER
In some sense it is impossible to pinpoint a "first" implementation
of the concepts behind THOUGHTSTICKER as it is now discussed. Pask and his
associates produced many machines from the middle 1950s to the middle 1970s,
each of them contributing important ideas to Conversation Theory and its
fruition in THOUGHTSTICKER. In the late 1960s and early 1970s, a few machines
were made that were clear predecessors to THOUGHTSTICKER. These were the
CASTE and EXTEND systems (Pask 1975c). Each had an electro-mechanical interface
manipulated by the human subject, connected to software programs for purposes
of recording data and performing some calculations most conveniently done
in software.
V.2.1 The Demands of Course Assembly
The very need for a system to represent the structure of knowables grew
out of the problem of representing subject matter for environments for learning.
The ultimate structure of the representations grew out of the epistemological
foundation of cybernetics, in the form of Conversation Theory itself.
It is interesting to note that the need for a representation of subject
matter for teaching came before the theory of conversations or its strict
calculus of knowledge representation. Once CASTE was mature as a "Course
Assembly and Tutorial Environment", THOUGHTSTICKER was conceived as
a software aid to assembling the structures which would "hold"
the subject matter for tutorial purposes.
The distinctions between CASTE and THOUGHTSTICKER are a source of confusion
since recent usage of these terms has tended to imply different implementations
in different hardware but both based on Conversation Theory and Lp. For
the record, recent uses of the name THOUGHTSTICKER emphasize the knowledge
representation functions and the name CASTE emphasizes the tutorial heuristics.
However, any THOUGHTSTICKER demonstration usually includes some CASTE functions
for purposes of practical use and demonstration, while CASTE operates on
the structures produced by THOUGHTSTICKER. Hence either term implies the
other and neither can be independent.
EXTEND was a related software program which allowed for the user (whether
in the role of "teacher" or "learner") to extend the
subject matter representation.
Thus it was subsequent to CASTE, EXTEND, and even THOUGHTSTICKER that, with
the invention of a bifurcation principle, CT produced what Pask would consider
a complete, scientific theory capable of encompassing, minimally, the domain
of epistemology.
V.2.2 THOUGHTSTICKER Defined
A precise distinction between Lp, Lp software and THOUGHTSTICKER was originated
by C Sheppard and R (Dik) Gregory of Admiralty Research Establishment (ARE),
Teddington, UK. "THOUGHTSTICKER" indicates a user interface written
in software and connected to a software embodiment of Lp structures and
processes ("Lp software"), within the constraints of present digital
technology, which constraints are very great compared to the intention behind
the formal protolanguage itself ("Lp"). Multi-process, concurrent,
conflict-ridden as well as conflict-free computation are a few of the gross
omissions inherent in any present-day THOUGHTSTICKER. Even the proposals
of modern AI for non-von Neumann, many-processor digital hardware is not
capable of the proper processing that is required for Lp. Some of these
points will be encountered more fully in the ensuing argument, below.
Even given these real restrictions, the potential benefits of a THOUGHTSTICKER
are very great as applied in a direct way to database construction and retrieval,
computer-aided instruction, and a variety of tasks that involve multiple-authors
and/or multiple locations. Its practical implications for decision support
and machine intelligence are only implied and as yet unexplored.
V.2.3 THOUGHTSTICKER in its current forms
The specific history of THOUGHTSTICKER implementations is offered in the
Appendix C as it is tangent to the main thesis. For our purposes here it
is sufficient to describe current implementations.
At present there are two major embodiments of THOUGHTSTICKER in software
available for examination.
Microcomputer BASIC Versions
There are two versions running in the Apple II microcomputer with additional
hardware boards. One, called Apple CASTE, was developed largely for the
Admiralty Research Establishment, UK, with some modules and features added
for the US Army Research Institute (ARI), by PANGARO Incorporated. Another
is called C/CASTE, based on the original code of Apple CASTE and developed
for the US Army Research Institute at Concordia University under the direction
of Pask. The systems have as their strengths that they are self-contained
in available and inexpensive hardware, and are well debugged and documented.
Their limits are in the size of the database they may comfortably contain
and the restricted set of Lp operations they perform.
Both contain the basic Lp operations (up to but not including condense/expand
and generalization). Apple CASTE emphasizes the authoring and presentation
of text models for Lp entities, although a simple Apple graphics module
can be used. In contrast, C/CASTE emphasizes the multi-display presentation
of tutorial material including computer-controlled slides of the subject
matter and maps of the knowledge representations.
Both use CASTE in their names, relating them to the Course Assembly System
and Tutorial Environment, the system that preceded THOUGHTSTICKER and Lp,
because the focus of their use is the tutorial application of CT.
Symbolics LISP Versions
This is an extended version of Lp operations, including simple generalization,
bifurcation, and extended conflict resolution, running on the Symbolics
LISP Machine. The THOUGHTSTICKER code is manifest in a number of forms on
the Symbolics, including a series of user interaction frames for studying
the evolution of the knowledge representation; a "naive" interface
for users without knowledge of CT; and the Expertise Tutor, used to teach
a naval command and control task. The power for the system is very great
due to the power of the environment of the Symbolics, its speed, size and
efficiency of experimentation and implementation. The implementation surpasses
all previous versions in raw functionality and capability.
Further details of the history of software development of THOUGHTSTICKER,
including its successful integration into a complete CT system of discourse
(the Expertise Tutor), are found in Appendix C.
The explanations below will use the Symbolics version for its examples,
although the particular details of screens and menu functions are minimized
as they are specific to this implementation; the concepts are however general
in the context of CT.
V.2.4 Lp at the Macro Level
As defined above in Section V.2.2 and in distinction to Lp software and
Lp itself, THOUGHTSTICKER is a software program which provides access to
a set of Lp functions in software. The formal description of its functions
is that of Lp, which in turn is the dual of CT, a macro theory of conversations
from whence it arises. Although it cannot be a complete implementation
of Lp (for both technical and theoretical reasons) THOUGHTSTICKER is used
below to explain the operations of Lp. Details of the full operation of
the software can be found in Pangaro et al 1985, attached.
V.2.5 Uses of THOUGHTSTICKER
To elicit knowledge from users, software such as THOUGHTSTICKER may operate
in one of two modes:
- In the background, accepting input from a domain (as from a simulation
called HUNKS for the ARE , or from the Team Decision System (TDS) as developed
for ARI); or
- Directly, with the user and the elicitation software engaged in the
one-on-one interaction.
The second is the model used for the following description as it is the
more general case. The discussion focuses on THOUGHTSTICKER as an engine
for receiving and representing knowledge without direct concern for the
ultimate structure and its kinetics, the central issue of this thesis. However
it is essential to present in detail the meaning and interpretation of the
structures at the macro level (to the "user" and his or her "psychology")
and from that presentation make the case for the correctness of interpretation
of the theory at the micro level (to the "topics" and their atomic
interactions).
The frames used as examples below are from the "research" version
of THOUGHTSTICKER, that is a set of interaction windows constructed in 1983
and 1984 whose purpose is to allow the user to easily explore the knowledge
representation scheme behind THOUGHTSTICKER. A simpler, "Naive THOUGHTSTICKER"
is also available for users not wishing to be exposed to the internal issues
of the scheme.
V.3 Making Statements
In a one-on-one interaction with THOUGHTSTICKER, it is the user's responsibility
to take initiative in making assertions which THOUGHTSTICKER endeavors to
represent in its internal structures. It is THOUGHTSTICKER's responsibility
to conform the user's actions to its internal requirements and to engage
the user in a dialogue when there is conflict or disagreement between the
user and THOUGHTSTICKER. There are additional features of THOUGHTSTICKER
that provide stimulation to the user in hopes of initiating novel assertions.
The primary way in which the user adds to THOUGHTSTICKER's internal knowledge
representation is by making text statements. Further transactions are required
to characterize the meaning within these statements. All this is accomplished
by display screens, menus of functions, and the ability to point to sections
of the text to indicate words and phrases within the user text.
Figure 1 shows a Write Watcher "window", as a software screen
of the Symbolics is called. It consists of "panes", each of which
is surrounded by a border and contains elements such as text and symbols.
The user in the role of "author" begins by typing statements at
the keyboard, which are entered into middle pane as if into a word processor
which is contained inside of THOUGHTSTICKER. The usual English language
conventions of grammar and punctuation are followed to the discretion of
the user.
The first important distinction to make about interaction with THOUGHTSTICKER
is that the system performs no semantic processing. This means that
details of sentence structure and grammar are ignored; the text in its entirety
is recorded by THOUGHTSTICKER for later retrieval. This is not to say that
a natural language interface would not be to advantage at the THOUGHTSTICKER
interface; it is an elaboration which is hoped for in future development.
Let us presume that the user has made a statement which reflects his or
her belief about a particular subject. In this case, the statement is "To
represent knowledge is the goal of artificial intelligence programming."
The result of typing the statement is seen in the lower pane. Certain words
or phrases appear in that pane in bold type. This indicates that users have
previously distinguished these words and phrases to THOUGHTSTICKER as significant
and to be noted whenever they appear. These are the topics of CT.
V.3.1 Models, Topics and Relations
It is important to differentiate the various elements of the authoring process
in terms of the strict definitions of CT:
- The text of sentences which are typed by the author are models. They
are not modified by the system itself but are "executed" (that
is, printed onto the screen) as a manifestation of the object they model,
whether a topic or entailment (relation).
- Models are associated with entailments. An entailment is a grouping
of a particular type, for example, a coherence or an analogy.
- The elements of entailments are topics. These are the exteriorized elements
of concepts, which are themselves clusters of processes. Topics are the
public elements of concepts (whether shared among different individuals
or merely exteriorized into the THOUGHTSTICKER interface). In the present
THOUGHTSTICKER, topics are represented by words or phrases.
The implications of entailment will be discussed below but to keep this
exposition reasonably brief, exceptions and qualifications to statements
will be minimized. Some implications worth noting are: models may very well
(and in certain cases, should) be graphics or sound; topics are not
the same as words or phrases but are efficiently represented so; models
need not contain identical words/phrases as the topics of the entailment
they model.
THOUGHTSTICKER can extract from the text of the model topics' words and
phrases which it already has noted, as Figure 1 shows. Not all topics will
have necessarily been asserted to THOUGHTSTICKER, and new topics can be
added manually by the author. This is accomplished by pointing at the words
or phrases in the lower pane. THOUGHTSTICKER checks to see whether the text
is close to previous topics (for example, is the new topic "goal"
similar to previous topics, such as "goals" or "goal structures").
The user may indicate that these new topics are intended to be the same
as earlier ones, or to be kept distinct.
The resulting topics are displayed in a separate pane (second from the top)
labeled "Topics."
V.3.2 Instating Entailments
Up to this point, the author has made a statement into THOUGHTSTICKER in
the form of a few sentences of text. This is to serve as a model of an entailment
involving a set of topics, which the author has also indicated to the system.
The next step is to integrate this new knowledge into the pre-existing models
and entailments which THOUGHTSTICKER holds.
V.3.3 Coherence
As noted, commands to THOUGHTSTICKER are made by choosing items on a menu.
The label on the menu choice in Figure 1 is "Instate"; clicking
here will provide further choices, to instate the utterance into the database
as a coherence, analogy, or topic model. Referring to the term from CT,
a coherence is an entailment between topics which requires, first,
that each topic in the entailment "make sense" with its neighbors
in the entailment. Thus, the meaning of knowledge must be derivable
from Artificial Intelligence and goal; recall that the purpose
of the model is to represent this meaning.
However, within a coherence every topic must be supported and "producible"
from all of its neighbors in the relation; hence, artificial intelligence
may be explained as something which has as its goal the representation
of knowledge. Equisignificantly, a goal is seen to be made
from the entailment of knowledge and artificial intelligence;
the topic goal is their entailment.
This requirement for coherence between topics distinguishes THOUGHTSTICKER
from all other knowledge representation software available in the field
of AI.
There is the issue of choosing how to model a given utterance, namely, how
it should be instated. The requirement for mutually-producible topics, just
described, is the minimum test for coherence. Lesser conditions can qualify
as analogy. Topic models are used if they require no further breakdown of
detail, that is, if the topic itself is an "atom" of the subject
matter. (The "Naive" modes of THOUGHTSTICKER provide a semi-automated
approach to this, where questions about the utterance are used by the system
to guide the user through considering how to best model the utterance in
the knowledgebase.)
V.3.4 Subjectivity of Statements
A few observations are in order at this stage. One may argue as to whether
one of the topics should really be "artificial intelligence programming",
rather than simply programming. This, as everything about the process
which the author undertakes, is a matter of opinion. Nothing about the representation
is true, immutable, or correct. It is merely the belief of the author, in
the context in which he or she is making statements. Thus, THOUGHTSTICKER
is a system for reflecting subjective assertions, namely, beliefs.
One may also argue about whether "goal" is really
produced from the remaining topics. Again, this is a matter of the opinion
of the author, but unless the sense of production can be comprehended
by users other than the author, the results will not be so useful. THOUGHTSTICKER
contains specific features, described below, which endeavor to make the
knowledge structure as comprehensible as possible (but always, of course,
within the limits of the communication skills of the author). Without the
concept of coherence, however, THOUGHTSTICKER is merely a keyword database
retrieval system. With coherence, it is a unique system for testing
agreement between individuals.
As will be seen in later chapters, coherence may be seen as a set of forces
imposed by and acting upon topics. In THOUGHTSTICKER, these forces are computed
by external fiat (see next section); however CT requires that they be dynamic
forces acting within the relations themselves.
V.4 Contradiction Checking
THOUGHTSTICKER must insure that new assertions are "consistent"
with old ones, in order to maintain the coherence of the knowledge representation
as a whole. THOUGHTSTICKER uses the requirement for coherent entailment
as a basis for evaluating the internal consistency of the knowledge representation,
as follows.
Upon the user's injunction of "Instate", the system compares the
proposed entailment against all previous entailments. Because THOUGHTSTICKER
does no semantic processing, all evaluation is done purely structurally,
by examining the topics in their entailments in the entire knowledge representation.
In essence, THOUGHTSTICKER searches for overlaps of the proposed entailment
with past entailments. If THOUGHTSTICKER determines a potential contradiction,
the user is warned as shown in Figure 2. A new menu has popped-up on the
screen, stating that the proposed entailment has conflicts in the mesh and
offering among other menu choices "Try to resolve conflicts."
The other menu choices refer to alternative interpretations of the same
text model. These are available if the user wishes to back away from asserting
a complete coherence. However by far the most interesting case is that of
conflict resolution, described in the following sub-sections.
V.4.1 Cases of Contradiction
The situations that THOUGHTSTICKER can detect are from the following cases:
- There is no overlap of topics: the new entailment is completely unrelated
to existing ones, and indeed contains topics which appear no where else.
The proposed entailment cannot be contradictory, but is unrelated. As, for
example in the present case, "Cybernetics is the epistemology of science."
(Intended topics are bold.)
- There is overlap on one topic only: in CT, this is called distributive
entailment, because the meaning of the overlapping topic is distributed
across more than one entailment. There is no structural contradiction and
hence the proposed entailment may be accepted. For example, "Artificial
Intelligence as a field was established in the 1950s by McCarthy, Minsky
and others."
- There is overlap on all but one topic; for example, there are some identical
topics present in both the proposed entailment and a previous entailment;
and two further, but different, topics, one in the proposed entailment and
one in the previous entailment. For example, "The goal of Artificial
Intelligence is to make computers smart like people." Within CT this
is the classic case of contradiction and requires some explanation, below.
- There is a common subset of topic names between the proposed entailment
and a previous one. Is the entailment with fewer topics a "conceptual
subset" as well, in that it is entirely contained in the larger one?
"Artificial Intelligence involves the embodiment of knowledge into
computers for the purpose of makeing computers smart like people..."
would be an example of this case.
- There is complete overlap: all topics in the proposed entailment are
contained identically in a previous entailment. Since THOUGHTSTICKER (as
noted) performs no semantic processing, the question arises: Are the entailments
truly identical? Or, was a new, different entailment intended by the author?
For example, "The goal of Artificial Intelligence is to program machines
to behave as if they contained the knowledge of human beings."
Of course, any of the cases of contradiction may exist with more than one
previous entailment.
In all of the cases cited, the procedure is one of comparison with past
entailments and (if necessary or desired) a resolution of the conflict based
on the author's intention. It is conceivable that any keyword retrieval
system could point out the condition of keywords in common, although without
CT as an underpinning, there would be no reason to draw any implications
from the particular structure in each case. THOUGHTSTICKER can provide specific
aid in interpreting and resolving the contradiction, as exemplified below
by the particular and perhaps most interesting case of classic contradiction,
where all but one topic in both entailments overlap.
V.4.2 Resolution of Conflict
Returning to the same statement example, THOUGHTSTICKER had detected a condition
of possible conflict between an existing entailment and the new statement
just made by the author. It is now up to the user and THOUGHTSTICKER to
modify the structure if resolution of conflict is desired.
THOUGHTSTICKER responds to the user injunction "Try to resolve conflicts",
shown in Figure 2, by offering a new window called the Resolver, shown in
Figure 3. This window shows the proposed entailment on the middle left,
with the previous entailment that it conflicts with, on the middle right.
The "Shared topics" is shown in a pane in the upper middle of
the window, "Artificial Intelligence" and "knowledge",
are in both entailments. "Goals" is present only in the
proposed entailment (left side) and "data structure" is
present only in the previous entailment (right side). The symmetric menu
choices on each side represent various procedures for the author to follow
to resolve the conflict; for example, to "Deny" one model or the
other. Other functions are aids to the user, for example "Undo"
returns to a previous state, and "Describe" gives relevant details
of the structure of the nearby database. As before, details of function
are available in Pangaro et al (1985).
As noted, the significance of the detection of conflict comes from the implication
of coherence: each topic is producible from the remainder of topics in the
same entailment. The present situation implies that the same topics (the
ones which overlap in both the proposed and previous entailment) may produce
either of two topics, thus:
- Artificial Intelligence and knowledge produce goals;
and
- Artificial Intelligence and knowledge produce data
structure.
Which is it? Either/Both? Each, but with qualification? Let us examine the
possible resolutions in detail:
- The non-overlapping topics are really the same topic: in this case,
goals and data structure were perhaps originally intended by the author
to have the same meaning. Here, it is not the case, although one can easily
imagine an author inadvertently using two different names for topics (for
example, goals and purposes, or data structure and internal representation)
while meaning the same thing.
- The two entailments should really be merged into one, relating all the
topics of both entailments. In this case, the result might be a model such
as: "The goal of Artificial Intelligence is to capture knowledge in
software data structures." Let us suppose for our purposes here that
this is not sufficient for the author, as the previous entailment was making
a slightly different point.
- One or more of the overlapping topics are not really a single topic
but are two (or more) as related by analogy. For example, Artificial Intelligence
is, in the proposed entailment, a field of programming; whereas in the previous
entailment, the statement is about the proponents of Artificial Intelligence,
the individuals themselves. A fine distinction, to be sure, but one which
must be accommodated within any knowledge representation scheme. THOUGHTSTICKER
would accommodate this by splitting Artificial Intelligence into Artificial
Intelligence programming and Artificial Intelligence proponents, as joined
by an analogy entailment, Artificial Intelligence. The overlap of topics
would now merely be a distributive entailment and the structure would be
coherent.
- Or, the author's intention is yet more subtle than any of the above,
and it was only through the author's process of semantic comparison that
the real intention is clear. In our example, this requires editing of both
of the existing models and a modification of the topics in the corresponding
entailments. Choosing "Modify" results in the appearance, as in
Figure 4, of a new pane which is used to modify the proposed entailment.
Figure 5 shows the same for the previous entailment. These smaller panes
are analogous to the original Write Watcher window and allow text editing,
choosing of topics, as well as the ability to examine previous uses of topics
in other entailments.
The outcome, where the text models and the topics in the entailment have
been changed, is one where there is no longer conflict within the entailments
we have been dealing with. This is indicated in Figure 6 by the new menu
choice "Local Resolution" in the top middle of the window, which
when chosen accepts the two entailments in their modified form. However,
the resolution is only local; this means that other difficulties
may exist between the new entailments and previous ones. The procedure
is thus recursive.
Contradiction checking by THOUGHTSTICKER does not in itself result in a
definite judgment that resolution is mandatory. The judgement is entirely
the user's, and the user may decide to perform a resolution or not, depending
on the purpose of the resulting knowledge representation.
V.4.3 To Resolve
In art, contradiction and its dual, ambiguity, are often used for conscious
effect. For psychological modelling, also, the existence of contradictory
structure may be appropriate. The important idea is that THOUGHTSTICKER
can represent what the author wishes, and the flexibility to provide
for any belief is one of its strengths. For tutorial purposes (and
for distributed planning and decision making), surely a self-consistent
structure will be appropriate.
V.5 Analogy
Like coherence, analogy is a type of relation within CT. A complete discussion
of analogy would require space far beyond what is practical here, as it
is the foundation of coherent relationships between topics and the basis
for condense/expand operations (where the evolution of analogies leads to
the creation of independent organizations of structures and possibly to
innovation). A synopsis is provided below rather than omit the topic but
it is not complete in the implications of analogy to Lp.
V.5.1 The Form of Analogy
At a level of modeled structure, analogy consists of a group of topics,
a similarity term and one or more difference terms. The similarity
term indicates how the topics are similar, while the difference term indicates
how they are distinguishable, i.e. distinct.
At all places at the interface, THOUGHTSTICKER allows for the user's choice
of analogy or coherence in instating a relation. Since current Lp software
does not dynamically interrelate coherence to analogies, the contradiction
checking is not modified by the existence of analogies; in future this should
be the case. Likely forms of resolution could be determined by the software
itself using analogical structures, and proposed to the user as candidates.
In an automatic Lp processor, each proposal could be executed concurrently
with the "richest" paths instated as new structures.
V.5.2 The Relation of Analogy and Coherence
Analogy is the most primitive form of relation between topics. Consider
that an analogy (at least) relates the (say) two topics that produce a third.
If, in addition to that production, one of the two producing topics and
the produced topic can also produce the second producing topics, then a
further and distinct analogy exists between those topics. The addition of
the final production (the other producing plus produced produces the first
producing) produces a condition that is recognized as the requirement for
coherence. Put another way, the existence of the necessary set of analogies
is coherence.
The rise of analogies is the necessary pre-condition to the existence of
coherence.
V.5.3 Analogy and Distributivity
The distributive case refers to a single topic overlapping in two or more
coherences. This is a very common event as a large structure of relations
could not easily exist without such a means to overlap relations. This status
of distributivity is important for a variety of reasons within Lp.
Consider the topic on which there is an overlap. From one perspective, the
topic is an atomic unit that applies to the (let us say in this example)
two coherences that it is present in. Put another way, the two relations
overlap on the similarity of the topic in both relations. However,
since the topic is produced in different ways in the two relations, some
differences must exist that could be extracted out of the two means
of producing the topic in the individual cases of the two relations.
Clearly the distributive case contains within it an analogical relation
centered on the topic in common to the relations.
Some comments are made in Section VI.4 concerning the role of analogy in
the microscopic simulation of Lp.
V.5.6 Adding Coherent Relations: Saturation
The philosophy of THOUGHTSTICKER is to provide the user with feedback which
is provocative to the authoring process. Contradiction checking is one such
feedback process, which indicates the way in which new statements relate
to previous ones. There is a converse situation, in which the system could
suggest entailments which are not yet present in the knowledge representation,
but which nonetheless would be permissible according to the rules of checking
contradiction.
Saturation is an operation whereby THOUGHTSTICKER suggests new entailments
to be made. This is analogous to a conversational partner asking for more
information about the relation of existing topics in the conversation, but
in new combinations. The challenge is to propose new combinations for which
the author is likely to want to, and be able to, provide models. The term
"saturation" implies that the entailments among topics are being
filled in or saturated, making a richly interconnected network of relationships.
In a large domain, the number of combinations of topics is very large and
most combinations would not be sensible to use as the basis for new models.
Arbitrary combinations chosen by the system would be absolute nonsense nearly
every time. The exceptions might be the seeds for innovation, as when very
different ideas are juxtaposed, making a new and unforeseen entailment (which
is the entire concept behind DeBono's "lateral thinking"). At
issue is the efficiency of the entire process, and the likelihood of useful
suggestions.
THOUGHTSTICKER must contain additional mechanisms for "focusing"
the saturation process to minimize absurd suggestions and to stimulate the
author in an efficient manner. Experience has shown that a combination of
these techniques results in an effective authoring process.
The first means of focusing the suggestion process is to use contradiction
checking to avoid new entailments that would conflict with existing ones.
THOUGHTSTICKER produces a possible combination of topics and checks the
possible entailment against existing ones. If a conflict exists, the suggestion
is discarded and a new possibility is generated combinatorially.
A second means to focus the saturation process is for the author to specify
what range of topics to choose from in the composition of new entailments.
The author indicates one or a few topics to start from, and requests THOUGHTSTICKER
to gather all topics which touch upon those topics in existing entailments.
The process may be repeated, reaching further out from the initial topic(s)
as far as the author wishes (with the limiting case of the inclusion of
every topic of the mesh in new suggestions). This is equivalent to
asking the system to make suggestions within a certain area of the knowledge
representation, for example, the part dealing with "Artificial Intelligence"
and all topics which connect to it.
A third means of focusing the saturation process is to require THOUGHTSTICKER
to include certain topic(s) in any new proposal, or, conversely, to avoid
using certain combinations of topics. The former is equivalent to specifying
a theme around which new entailments are to revolve, for example, all new
suggestions are to contain "Artificial Intelligence." The latter
is equivalent to specifying that "Artificial Intelligence" and
"windows" can be eliminated as a possible combination because
it is incongruous, or simply because the author has nothing to say about
them together.
Saturation derives its power from coherence and contradiction checking.
It is conjectured within CT that the saturation process is constantly at
work in the processes of intelligence, connecting and re-connecting concepts
as they are generated or integrated from outside information. It is this
process which is the foundation of agreement. Without the interrelation
of pre-formed concepts with the influence of new concepts, knowledge would
be trapped within its own capsules, and perforce could not evolve or even
come to exist. The saturation operation of THOUGHTSTICKER mimics this process
of mind in a crude f |
