John Sowa position paper
Tom Gruber <Gruber@sumex-aim.stanford.edu>
Full-Name: Tom Gruber
Message-id: <2878500210-6083771@KSL-Mac-69>
Date: Wed, 20 Mar 91 15:23:30 PST
From: Tom Gruber <Gruber@sumex-aim.stanford.edu>
To: Shared KB working group <srkb@isi.edu>
Subject: John Sowa position paper
\documentstyle{article}
\author{John F. Sowa\\
IBM Systems Research Institute\\
500 Columbus Avenue\\
Thornwood, NY 10594\\
Email: sowa@ibm.com}
\title{Multi-Domain Knowledge Representation:\\
Position Statement}
\begin{document}
\maketitle
\begin{abstract}
AI has been most successful on small domains: the
microworlds of early AI demos; the highly specialized expert systems for
commercial applications; and the machine translation systems like METEO, which
require no human editing, but are restricted to the very narrow topic
of weather reports. Such knowledge bases can be shared and reused, but
only for other projects that are similarly restricted. The position
taken in this paper is that such compartmentalization is inevitable:
all deep knowledge is domain dependent. Only superficial, syntactic
knowledge carries over from one domain to another. A serious question
to consider is whether such superficial knowledge can provide a
framework in which the deeper domain-dependent knowledge can be shared.
The answer given in this paper is maybe: some things can be shared,
but the research needed to support a significant amount of sharing of
knowledge representations across multiple domains is still in a
primitive stage of development.
\end{abstract}
\section*{Examples of Domain-Independent Knowledge}
Many different projects have surface similarities that seem to suggest
that shared knowledge representations are possible. Expert systems
designed to assist automobile drivers, airplane pilots, ship captains,
and locomotive engineers, for example, would seem to have a lot in
common. All of them must deal with time, speed, and distance as well
as fuel consumption, equipment condition, and passenger safety.
Programming languages also have a great deal in common, as the
following assignment statements seem to indicate:
\begin{verbatim}
APL: X <- A + B
FORTRAN: X = A + B
PL/I: X = A + B;
Pascal: X := A + B;
\end{verbatim}
Yet these surface commonalities mask serious differences in detail. A
deeper analysis indicates that the similarities are more syntactic than
semantic: the concepts required for each domain are so tightly bound
to that domain that they cannot be mapped from one to the other.
Generalizations that cover multiple domains have so little detail that
it is not clear whether they can contribute anything significant to the
development of a new knowledge base in any of the more detailed domains.
First consider the possibility of common knowledge bases for
automobiles, airplanes, ships, and trains. A major difference between
these domains is the number of degrees of freedom in the motion. A
train's motion is purely one dimensional because of the rigid tracks.
At a gross level, a car's motion is also one dimensional, but at a
detailed level, the driver must maneuver in two dimensions to keep the
car in lane and avoid other cars and obstacles. A ship's motion is also
two dimensional, but its greater inertia causes a change in course to
take minutes instead of the split-second changes that are possible with
a car. An airplane's motion is three dimensional, but changes in
attitude introduce three more degrees of freedom. Besides differences
in motion, there are different kinds of signals to consider and
different ways of planning a course and following it. As a result, a
driver, a pilot, a captain, and an engineer have totally different ways
of thinking and reacting. A person who is both a driver and a pilot
would have two independent modes of thought with little or nothing in
common. Expert systems designed for each of these domains would also
have few common concepts and practically no common rules.
For the programming languages, the similarities in syntax mask major
differences in semantics. If A, B, and X were all integers or all
floating-point numbers, the results would be the same for each of the
languages. But differences arise when the data types are different.
FORTRAN and PL/I allow type conversions to or from integer and
floating-point, but Pascal only does automatic conversion from integer
to floating and would print an error message if A+B happened to be
floating-point and X were integer. APL also does automatic conversions
in evaluating A+B; but in doing the assignment, it could change the type
of X instead of converting the result of A+B to X's previous type. PL/I
does many other kinds of automatic conversions and would even convert
character strings to and from numbers. APL and PL/I both allow A, B,
and X to be arrays as well as simple scalars; but PL/I places more
restrictions on the dimensions of the arrays, while APL has fewer
restrictions and APL2 has even less. Because of these differences,
terms like {\em addition} or {\em assignment} statement can be given a precise
definition only for a single programming language. In some cases, the
language standards are so loose that the definition may change with
every compiler or even every modification of a compiler. An ontology
might include ADDITION as a concept type, but it would also require
subtypes APL-ADDITION, FORTRAN-ADDITION, and so on for every programming
language and dialect.
Even the same physical object may be represented in totally different
ways for different purposes. A highway, for example, is one-dimensional
on a map. For an automobile driver, it is two-dimensional. For the
workers building the roadbed, it is three dimensional, but highly
regular. And for the surveyors who are planning a level road through
hilly terrain, it is three dimensional with highly irregular amounts of
cut and fill. Any physical object or system can be represented at an
unlimited number of levels of detail. There is no stopping point that
is natural to the object itself; the stopping point depends entirely on
the purpose for which that object is being used.
\section*{Is Natural Language Domain Independent?}
Natural languages can express knowledge about any topic in any domain.
But that does not make them domain independent. The syntax of language
and the constraints at the level of case frames are largely domain
independent, but the meaning of each word is highly dependent on the
domain. As an example, consider the following four sentences:
\begin{flushleft}
Tom supported the tomato plant with a stick.\\
Tom supported his daughter with \$8,000 per year.\\
Tom supported his father with a decisive argument.\\
Tom supported his partner with a bid of 3 spades.
\end{flushleft}
These sentences all use the verb {\em support} in the same syntactic pattern:
\begin{flushleft}
A person supported NP1 with NP2.
\end{flushleft}
Yet each use of the verb can only be understood with respect to a
particular subject matter or domain of discourse: physical structures,
financial arrangements, intellectual debate, or the game of bridge. For
each of these domains, the concept type SUPPORT would require different
subtypes, such as PHYSICAL-SUPPORT, FINANCIAL-SUPPORT, or INTELLECTUAL-
SUPPORT. Each of those subtypes could be subdivided further: physical
support by being tied to a stick could be distinguished from support by
being propped up from below or being suspended from above; financial
support by an allowance could be distinguished from support by a trust
fund or support by payments at irregular intervals. Each difference in
concept type makes a difference in reasoning and behavior: a child with
a regular allowance enjoys some measure of stability, while a child who
gets irregular payments must be on good behavior, always hoping for
another grant at any moment.
The point of these examples is that vagueness and ambiguity do not
result from the nature of language. Instead, they result from the use
and reuse of the same words in many different domains and applications.
The same kinds of ambiguities that arise with a technical term like
assignment statement also arise with a common verb like {\em support}. The number of different concept types associated with a word is unlimited,
and the totality of meanings may be inconsistent. An interior
decorator, for example, may think of walls as parts of a room, while a
construction contractor may think of them as separators between rooms.
Each view is correct for a certain purpose and point of view, but they
are incompatible with one another. The word senses listed in
dictionaries represent the most common applications, and larger
dictionaries list more of them. But even the largest dictionaries fail
to distinguish such nuances as addition in APL vs. addition in FORTRAN
or support by an allowance vs. support by irregular payments. Although
the different meanings of addition, support, and wall are incompatible,
they still have something in common. It is easier for a person to learn
and use a single word for them than to learn different words that change
with every application. But that implies that the only thing that is
easily shared or reusable is the syntax, not the deeper semantics of the
knowledge base.
\section*{Language Games}
The traditional AI approach to knowledge representation resembles the
early philosophy of Ludwig Wittgenstein, as presented in the {\em Tractatus Logico-Philosophicus}. In his later philosophy, Wittgenstein presented
scathing criticisms of his earlier work -- all of which apply equally
well to the current attempts to build shared, reusable knowledge bases.
Yet his later work is not totally negative; it contains the basis for a
solution. His theory of language games suggests that the way to build
large, flexible intelligent systems is to provide a framework that can
use and reuse the same syntactic tokens in different language games for
different domains. Some of the implications of these ideas for AI were
discussed in the last chapter of a book (Sowa 1984), two recent papers
(Sowa 1990, 1991), and a workshop on large knowledge bases (Silverman and
Murray 1991).
\section*{References}
\begin{description}
\item
Silverman, Barry G., and Arthur J. Murray (1991) "Full-sized
knowledge-based systems research workshop," {\em AI Magazine}, vol. 11, no. 5,
January 1991, pp. 88-94.
\item
Sowa, J. F. (1984) {\em Conceptual Structures: Information Processing in Mind and Machine}, Addison-Wesley, Reading, MA.
\item
Sowa, J. F. (1990) "Crystallizing theories out of knowledge soup," in
{\em Intelligent Systems: State of the Art and Future Directions}, edited by
Zbigniew W. Ras and Maria Zemankova, Ellis Horwood, New York, pp.
456-487.
\item
Sowa, J. F. (1991) "Lexical structures and conceptual structures," in
{\em Semantics in the Lexicon}, edited by James Pustejovsky, to be published by Kluwer Academic Press.
\item
Wittgenstein, Ludwig (1921) {\em Tractatus Logico-Philosophicus}, Routledge and
Kegan Paul, London, 1961.
\item
Wittgenstein, Ludwig (1953) {\em Philosophical Investigations}, Basil
Blackwell, Oxford.
\end{description}
\end{document}