Jintae Lee and Tom Malone position paper
Tom Gruber <Gruber@sumex-aim.stanford.edu>
Full-Name: Tom Gruber
Message-id: <2878500078-6075869@KSL-Mac-69>
Date: Wed, 20 Mar 91 15:21:18 PST
From: Tom Gruber <Gruber@sumex-aim.stanford.edu>
To: Shared KB working group <srkb@isi.edu>
Subject: Jintae Lee and Tom Malone position paper
\documentstyle{article}
\title{The Position Statement for the Shared KB workshop}
\author{Jintae Lee and Thomas W. Malone\\
MIT Center for Coordination Science}
\begin{document}
\maketitle
In our research, we are concerned with how to represent, manipulate,
and share knowledge that can be used by both people and their
computational agents. This implies two key characteristics of the
KB's we use: They are semiformal and end-user manipulable.
\section{Semiformal KBs}
By a semiformal KB, we mean a knowledge base with the following three
properties: (1) it represents certain information in formally
specified ways that facilitate automatic processing of this
information; (2) it represents and make it easy for humans to process
the same or other information in ways that are not formally specified;
and (3) it allows the boundary between formal processing by computers
and informal processing by people to be easily changed.
For example, the Object Lens system [Lai et al. 88] represents
information using a relatively straightforward form of object
inheritance hierarchy, with slots for each type of object and links
between objects. However, unlike most traditional knowledge bases, we
do not place any necessary restrictions on the contents of slots.
Except when explicitly specified otherwise, for instance, a slot can
contain an arbitrary mixture of uninterpreted text and links to other
objects (including, in principle other uninterpreted media such as
voice or video).
Using this system, people can represent certain information in formal
ways that agents can process. For instance, agents can use
information in the "From" field of a "Message" to sort mail. They can
also use links in the "Supervisor" field of "Employee" objects to do
more sophisticated reasoning about objects (such as finding all people
who report to vice-presidents or all messages from people in a certain
group) and to construct useful displays (such as organization charts).
At the same time, however, people can enter much information into the
knowledge base that is not formally represented at all. For instance,
messages can have text fields, and "Task" objects can have task
descriptions. Even in cases where the system might expect to find
something formal (such as a link to a room for a meeting), users can
enter informal information (such as "We don't know yet.") Thus, the
system can be useful to people, even when they don't have the time,
inclination, or understanding to represent things formally, while
still allowing them to take advantage of as much formal structure as
they care to create.
The Sibyl system [Lee 90], which is built on top of Object Lens,
illustrates the usefulness of semiformal KBs for capturing and reusing
decision rationales. In this system, people represent some knowledge
formally, including the relationships between the "Goals,"
"Alternatives," and "Arguments" for a decision problem. The
descriptions of each of these goals, alternatives, and arguments,
however, can be a mixture of free text and formal objects representing
domain knowledge. Thus, the system can use some structure to display
useful summaries and retrieve previous relevant decisions, without
requiring people to completely encode everything they want to say in
an argument.
\section{End-user manipulable KBs}
By an end-user manipulable KB, we mean one which people with no
computer programming experience can easily see, augment, and
manipulate. For instance, we have taken great pains in the Object
Lens system to make it very easy for end users to define new object
types, add fields to them, and to specify display formats for
collections of objects. Programmers can, of course, do all these
things in any knowledge base, but we have tried to make these things
very easy for unsophisticated users by using a combination of simple
graphical templates and menu choices. For example, users can specify
whether to display a collection of objects in a table, a tree, a
calendar or a matrix by a set of simple menu selections.
\section{Developing knowledge bases as a byproduct of doing other
things}
As to the question, "How should reusable knowledge bases be
developed," one way of breaking the knowledge acquisition bottleneck
is to acquire knowledge as a part of performing the tasks in which the
knowledge will be useful. Elsewhere, we referred to this as
task-embedded knowledge acquisition [Lee 89]. In this mode of
knowledge acquisition, a knowledge base then is a by-product rather
than something that has to be created with extra efforts. For
example, in SIBYL, decision rationales are captured as a part of
making decisions, which is then used by others for similar decisions
in the future. Of course, this mode of knowledge acquisition requires
that the knowledge base be manipulable directly by the end users, and
that it not require more formalization than the user is willing to
provide.
If we have a KB that cumulates each time we perform a task, then we
can sometimes let common ontologies grow out of this KB. For example,
each time we use SIBYL to design something, say a window manager, many
aspects of the design get represented and cumulated. They include
the alternatives, requirements, and arguments considered. As more
and more cases get represented, people can use objects from past
designs and specialize or generalize them. This results in an
ontology for a given type of task that is grounded in actual practice.
Such an ontology is often not good enough to be used as a "standard";
we may need to check and readjust its categories for proper
abstraction, consistency, mutual exclusiveness, etc. However, it at
least provides a useful basis for building a solid
ontology.
\section{Taxonomizing various translation schemes}
Now as to the last question, "What are the critical issues that need
to get resolved," one of the issues that we feel is important in
sharing knowledge is when to try imposing a standard and when to try
providing translations. Elsewhere [Lee
\& Malone 90], we approached this problem by taxonomizing and
analyzing the space of all possible
translation schemes. For instance, one obvious scheme is for everyone
to use the same language, in
which case no translation is necessary. Another obvious scheme is for
there to be no common language
and a separate translator for each pair of languages. (This, of
course, requires n squared translators
for n languages.) A slightly more sophisticated approach is to have a
single common language into and
out of which each language is translated (thus requiring 2n
translators for n languages). One of the
intriguing examples we identified was a translation scheme that
exploited inheritance hierarchies. If,
for example, I send you an instance of type "Student" when you don't
know about students, your system
might still be able to automatically translate this into the nearest
common ancestor we both share
(e.g., "Person"). This translation could, for instance, preserve all
the fields shared with the common
ancestor and put the additional information into an uninterpreted
"comments" field. Our analysis in
this paper includes a proposal for a composite scheme (called
"Partially Shared Views") which combines
the best features of all the translation schemes we considered.
This study admittedly only scratches the surface of the problem. As
applied to the problem of shared reusable KBs, or shared ontologies,
the relevant questions are: Do we want a canonical set of primitives?
When do we want to allow them to be customized? Is translation among
the customized or specialized primitives feasible, desirable? What
kinds of translation mechanism are possible and what are the
dimensions along which tradeoffs occur?
If we proceed to work on a single shared ontology, without considering
these broader issues, then we might make the mistake of having a
technology that solves no real problems.
\section*{References}
\begin{description}
\item
Lai, K.-Y., T. Malone \& K.-C. Yu (1988). Object Lens: A
"Spreadsheet" for Cooperative Work. ACM
Transactions on Office Information Systems 6(4) 332-353
\item
Lee, J. (1989). Task Embedded Knowledge Acquisition through a
Task-Specific Language. Proceedings of IJCAI Workshop on Knowledge
Acquisition Detroit, MI.
\item
Lee, J. \& T. Malone (1990). Partially Shared Views: A Scheme for
Communicating among Groups that Use
Different Type Hierarchies. Transactions on Information Systems 8(1)
1-26
\item
Lee, J. (1990). SIBYL: A Qualitative Decision Management System. in
Winston, P. H. \& S. Shellard
(Eds.) Artificial Intelligence at MIT: Expanding Frontiers. vol. 1.
MIT Press: Cambridge, MA
\end{description}
\end{document}