Building Ontologies
Mike Uschold <mfu@aiai.edinburgh.ac.uk>
From: Mike Uschold <mfu@aiai.edinburgh.ac.uk>
Date: Wed, 23 Nov 94 14:08:49 GMT
Message-id: <5896.9411231408@subnode.aiai.ed.ac.uk>
To: ontolingua@HPP.Stanford.EDU
Subject: Building Ontologies
Cc: gbibm3ws@ibmmail.com
Everyone,
The following is a short position statement by a colleague of mine relevant to
how one goes about building ontologies. He is interested in any feedback
you may have. In a subsequent message, I will describe our method and
experiences in building an `enterprise' ontology.
Mike Uschold
======================================================================
On Categorization in Modelling
Levels of granularity or generalization
Martin King
IBM United Kingdom Limited
Introduction
This note examines the choice of categories for definition in
the process of modelling. It assumes that there is some
universe of discourse, typically an identified part of the
real world, and that the process of modelling is to create a
definition which represents those aspects of the universe of
discourse that are of interest to the modeller. Typically such
a modelling process will analyse the universe of discourse and
name and define categories of objects found in it.
This note seeks to apply general ideas on categorization to
the choice of categories in modelling, and in particular ideas
on the generality, or conversely, granularity of
categorization. A prime source of these ideas is work on
cognitive psychology and linguistics from the book "Women,
Fire, and Dangerous Things" by George Lakoff LLakoff 1987]. It
also draws on experience of modelling using the IBM BSDM
method LIBM 1992].
Generality and Categorization
The Range of Choice
A universe of discourse will typically be seen informally as a
collection of objects. As well as individuals, there are
likely to be categorizations or classes already recognized
explicitly or implicitly before the process of modelling
begins. As already mentioned, it is assumed that the process
of modelling will include a large element of choice and
definition of categorizations. These may be based on pre-
existing categorizations found in the UoD or they may be
created for the purpose of the model.
A priori, a class created for the model could be as large as
one general class having every object in the UoD as a member;
at the other extreme there could be a specific class for each
single object in the UoD. Whether or not these extremes would
be used in a model they are the extremes of a spectrum from
the very general categorization to the very specific. In
practice any model is likely to include definitions of several
classes from varying points in the spectrum.
The issue addressed in this note is whether there is any
general guidance on the range within this spectrum from which
useful categorizations for modelling are most likely to be
found.
A Theory of Categorization
Categorization is the subject of the book by Lakoff already
mentioned LLakoff 1987]. He presents categorization as
fundamental to human cognition, a view that few will argue
with. Less classically, he holds that categorization is very
much more complex than the view of hierarchies of classes with
clear common properties and membership criteria being the
norm. The theory he presents is based on a view developed from
the later philosophy of Ludwig Wittgenstein through to
research of Eleanor Rosch and associates.
Lakoff summarizes the work in eleven themes, two of which are
most relevant to the present subject and are described as
follows LLakoff 1987 p 13]:
Basic-level categorization: The idea that categories are
not merely organized in a hierarchy from the most general
to the most specific, but are organized so that the
categories that are cognitively basic are "in the middle"
of a general-to-specific hierarchy. Genralization
proceeds "upward" from the basic level and specialization
proceeds "downward".
Basic-level primacy: The idea that basic-level categories
are functionally and epistemologically primary with
respect to the following factors: gestalt perception,
image formation, motor movement, knowledge organization
Lmy italics], ease of cognitive processing (learning,
recognition, memory, etc.), and ease of linguistic
expression.
A third theme is also relevant to the usability of chosen
categories LLakoff 1987 p 12]:
Functional embodiment: The idea that certain concepts are
not merely understood intellectually; rather, they are
used automatically, unconsciously, and without noticeable
effort as part of normal functioning. Concepts used in
this way have a different, and more important,
psychological status than those that are only thought
about consciously.
Lakoff gives some examples to illustrate the basic level
LLakoff 1987 p 46]:
Superordinate ANIMAL FURNITURE
Basic Level DOG CHAIR
Subordinate RETRIEVER ROCKER
He summarizes the properties of basic-level categories as
follows LLakoff 1987 p 47]:
Perception: Overall perceived shape; single mental image;
fast identification.
Function: General motor program Lreferring to physical
interaction with a category member].
Communication: Shortest, most commonly used and
contextually neutral words, first learned by children and
first to enter the lexicon.
Knowledge Organization: Most attributes of category
members are stored at this level.
Lakoff supports the last point by citing experimental evidence
published by Tversky and Hemenway (1984).
Categorization in Information Systems Modelling
Modelling has been increasingly widely used as information
systems become more pervasive and more complex. Many different
methods are practised. Widely used methods include, explicitly
or implicitly, a strong element of categorization.
Some methods work from the general to the specific, "top-
down". For example, the processes in a business will be
analysed starting by agreeing a break-down of the whole
business into typically 7 plus or minus 2 major processes. The
number corresponds with the size of the human short-term
memory, but perhaps more directly with the number of labelled
boxes that can fit easily on a page of A4 with a typical
diagramming convention. As the next step each of these major
processes is broken down to a similar number of lower level
processes. The breakdown may be repeated further.
Such methods are widely used, but have been found to have
various problems. It is very difficult to give any measurable
guidelines on when to stop the breakdown. While breaking down
to the next level provides more detail, this does not
necessarily mean more precision. The result often results in
the same lower-level process being duplicated in many
different parts of the hierarchy. The result is often very
sensitive to changes in the hierarchical organization of the
business. On the other hand, such methods often have quick
appeal to senior business management who find it relatively
easy to define the first level of breakdown. Is this in
conflict with the basic-level theme of Lakoff? I suggest that
the mechanism at work here is that over time any large
organization will manage its processes by allocating
responsibility across departments. The subdivision of
processes and naming of the departments will be related, and
management thinking will be strongly influenced by this.
However it is precisely this subdivision of processes and
renaming which is subject to change and leads to an unstable
model.
Some attempts have been made to build models by analysing
large numbers of very specific categories, "bottom-up".
Generally speaking, these have not been very successful. It is
easy to spend very large amounts of time and effort with
little practical result.
A further family of methods attempt to base the analysis
starting at an intermediate point in the spectrum. An example
of these is IBM BSDM. It looks for terms widely used in the
business and tries to identify well-defined concepts to relate
to these terms according to a comprehensive set of criteria.
Though not expressed in the same words, many of these criteria
correspond closely with "gestalt" perception and the other
properties of the basic-level concepts cited from Lakoff
above.
Categorization and Ontology
The relatively new field using the label ontology is also
concerned with modelling. It (inescapably?) relies on
categorization. Much of the published work in the field
appears to have a strong emphasis is on trying to define a
relatively small number of very general concepts. Typical
examples are the terms AGENT or BEHAVIOUR. Ontology work is
typically aimed at being general and being shared. Does trying
to pick a small number of concepts to cover a wide range of
discourse lead to choosing concepts which in Lakoff's terms
are superordinate or above the basic level? Will models
primarily based on superordinate terms be useful in themselves
or will they need basic level concepts? Will models primarily
based on superordinate concepts be a useful intermediate step
in producing models with basic level concepts or is there a
better strategy for building these? Not all these questions
are likely to be answered quickly and easily.
Taking AGENT as an example may start to answer the first
question. It is defined in an ontology under development as:
AGENT is an entity that is able to perform, or
participates in the performance of, one or more
BEHAVIOURs. It is the supplier of a force behind a
BEHAVIOUR.
What perception does this definition give to anyone other than
an ontology specialist? An overall perceived shape? A single
mental image? Fast identification?.
Does it indicate whether any physical interaction is possible,
and what sort?
Is the word short (yes), commonly used (maybe, but with this
meaning?), learned by children?
What attributes are obviously associated with this concept?
Is the definition is intended to mean a higher level
categorization which includes the lower level categorizations
of person, of a machine that can do things instead of a
person, and possibly of a group of persons? If so, contrast
AGENT as defined above with a plausible definition of PERSON.
The answers to the above questions are likely to be
dramatically different.
Conclusions
The idea of trying to build a general model that can be
applied to a wide range of specific situations is attractive.
If general reasoning can be developed once and re-used this
could save effort in each situation. The smaller the number of
general concepts required, the smaller may be the initial
effort. If however trying to limit the number of concepts
leads to trying to define them at a level that is contrary to
what is natural to human cognition, the effort to produce any
result and limitations in it may be counter-productive.
I suggest that workers in the field of ontology should review
what appears to be a fundamental strategy in current work and
check for evidence that it can produce useful results. If not,
the objective of trying to produce ontologies to permit
interchange of information and knowledge between different
users may be better served by a strategy based on basic level
concepts as defined by Lakoff.
References
Lakoff 1987 Lakoff, George, "Women, Fire, and Dangerous
Things", University of Chicago Press, 1987, (Paperback
1990 ISBN 0-226-46804-6)
IBM 1992 "Introducing BSDM", GE19-5387, International
Business Machines Corporation, 1992.
AD Consulting Services Internal: 7-453090
Rosanne House (RH2B) PUBLIC: +44-0707-363090
Welwyn Garden City Internet: gbibm3ws@ibmmail.com
GB - AL8 6TZ FAX: +44-0707-338732