B. Chandrasekaran and John R.

Josephson, Ohio State University

V. Richard Benjamins, University of Amsterdam

EORIES IN AI FALL INTO TWO

broad categories: mechanism theories and
THIS
SURVEYPROVIDES A CONCEPTUAL INTRODUCTION
content theories. Ontologies are content the- TO ONTOLOGlES AND THEIR ROLE IN LVFORMATlON
ories about the sorts of objects, properties of
SYSTEMS AND Al. THE AUTHORS ALSO DlSCUSS HOW
objects, and relations between objects that are
possible in a specified domain of knowledge. ONTOLOGIES CLARIFY THE DOMAIN’S STRUCTURE OF
They provide potential terms for describing
our knowledge about the domain.
KNOWLEDGE AND ENABLE KNOWLEDGE SHARlNG.
In this article, we survey the recent devel-
opment of the field of ontologies in AI. We
point to the somewhat different roles ontolo-
gies play in information systems, natural- mean one of two related things. First of all, requires careful analysis of the kinds of objects
language understanding, and knowledge- ontology is a representationvocabulary, often and relations that can exist in the domain.
based systems. Most research on ontologies specializedto some domain or subject matter. In its second sense, the term ontology is
focuses on what one might characterize as More precisely, it is not the vocabulary as such sometimes used to refer to a body of knowl-
domain factual knowledge, because knowl- that qualifies as an ontology, but the concep- edge describing some domain, typically a
ede of that type is particularly useful in nat- tualizations that the terms in the vocabulary commonsense knowledge domain, using a
ural-language understanding. There is an- are intended to capture. Thus, translating the representation vocabulary. For example,
other class of ontologies that are important terms in an ontology from one language to CYC’ often refers to its knowledge repre-
in KBS-one that helps in sharing know- another, for example from English to French, sentation of some area of knowledge as its
eldge about reasoning strategies or problem- does not change the ontology conceptually.In ontology.
solving methods. In a follow-up article, we engineering design, you might discuss the In other words, the representation vocab-
will focus on method ontologies. ontology of an electronic-devices domain, ulary provides a set of terms with which to
which might include vocabulary that describes describe the facts in some domain, while the
conceptualelements-transistors, operational body of knowledge using that vocabulary is
Ontology as vocabulary amplifiers, and voltages-and the relations a collection of facts about a domain. How-
between these elements-operational ampli- ever, this distinction is not as clear as it might
In philosophy, ontology is the study of the fiers are a type-of electronic device, and tran- first appear. In the electronic-device exam-
kinds of things that exist. It is often said that sistors are a component-ofoperational ampli- ple, that transistor is a component-of opera-
ontologies “carve the world at its joints.” In fiers. Identifying such vocabulary-and the tional amplifier or that the latter is a type-of
AI, the term ontology has largely come to underlying conceptualizations-generally electronic device is just as much a fact about

20 1094-7167/99/$10.000 1999 IEEE IEEE INTELLIGENT SYSTEMS

its domain as a CYC fact about some aspect edge Level versus Symbol Level)5 all grap- terms such as transistors and diodes; general
of space, time, or numbers. The distinction ple in their own ways with characterizing terms such as functions, causal processes,
is that the former emphasizes the use of content. Ontologies are quintessentially con- and modes; and terms that describe behavior
ontology as a set of terms for representing tent theories, because their main contribution such as voltage. The ontology captures the
specific facts in an instance of the domain, is to identify specific classes of objects and intrinsic conceptual structure of the domain.
while the latter emphasizes the view of ontol- relations that exist in some domain. Of In order to build a knowledge representation
ogy as a general set of facts to be shared. course, content theories need a representa- language based on the analysis, we need to
There continues to be inconsistencies in tion language. Thus far, predicate calculus- associate terms with the concepts and rela-
the usage of the term ontology. At times, the- like formalisms, augmented with type-of tions in the ontology and devise a syntax for
orists use the singular term to refer to a spe- relations (that can be used to induce class encoding knowledge in terms of the concepts
cific set of terms meant to describe the entity hierarchies), have been most often used to and relations. We can share this knowledge
and relation-types in some domain. Thus, we describe the ontologies themselves. representation language with others who
might speak of an ontology for “liquids” or have similar needs for knowledge represen-
for “parts and wholes.” Here, the singular tation in that domain, thereby eliminating the
term stands for the entire set of concepts and Why are ontologies need for replicating the knowledge-analysis
terms needed to speak about phenomena important? process. Shared ontologies can thus form the
involving liquids and parts and wholes. basis for domain-specific knowledge-repre-
When different theorists make different pro- sentation languages. In contrast to the previ-
posals for an ontology or when we speak about Ontological analysis clarifies the structure ous generation of knowledge-representation
ontology proposals for different domains of of knowledge. Given a domain, its ontology languages (such as KL-One), these lan-
knowledge,we would then use the plural term forms the heart of any system of knowledge guages are content-rich; they have a large
ontologies to refer to them collectively.In AI representation for that domain. Without number of terms that embody a complex con-
and information-systems literature, however, ontologies, or the conceptualizations that tent theory of the domain.
there seems to be inconsistency: sometimes we underlie knowledge, there cannot be a vocab- Shared ontologies let us build specific
see references to “ontology of domain” and ulary for representing knowledge. Thus, the knowledge bases that describe specific situ-
other times to “ontologies of domain,” both first step in devising an effective knowledge- ations. For example, different electronic-
referring to the set of conceptualizations for representation system, and vocabulary, is to devices manufacturers can use a common
the domain. The former is more consistent with perform an effective ontological analysis of vocabulary and syntax to build catalogs that
the original (and current) usage in philosophy. the field, or domain. Weak analyses lead to describe their products. Then the manufac-
incoherent knowledge bases. turers could share the catalogs and use them
An example of why performing good in automated design systems. This kind of
Ontology as content theory analysis is necessary comes from the field of sharing vastly increases the potential for
databases6Consider a domain having sev- knowledge reuse.
The current interest in ontologies is the lat- eral classes of people (for example, students,
est version of AI’s alternation of focus be- professors, employees, females, and males).
tween content theories and mechanism the- This study first examined the way this data- Describing the world
ories. Sometimes, the AI community gets base would be commonly organized: stu-
excited by some mechanism such as rule sys- dents, employees, professors, males, and We can use the terms provided by the
tems, frame languages, neural nets, fuzzy female would be represented as types-of the domain ontology to assert specific proposi-
logic, constraint propagation, or unification. class humans. However, some of the prob- tions about a domain or a situation in a
The mechanisms are proposed as the secret lems that exist with this ontology are that stu- domain. For example, in the electronic-
of making intelligent machines. At other dents can also be employees at times and can device domain, we can represent a fact about
times, we realize that, however wonderful the also stop being students. Further analysis a specific circuit: circuit 35 has transistor 22
mechanism, it cannot do much without a showed that the terms students and employee as a component, where circuit 35 is an
good content theory of the domain on which do not describe categories of humans, but are instance of the concept circuit and transistor
it is to work. Moreover, we often recognize roles that humans can play, while terms such 22 is an instance of the concept transistor.
that once a good content theory is available, asfemales and males more appropriately rep- Once we have the basis for representing
many different mechanisms might be used resent subcategories of humans. Therefore, propositions, we can also represent knowl-
equally well to implement effective systems, clarifying the terminology enables the ontol- edge involving propositional attitudes (such
all using essentially the same content.2 ogy to work for coherent and cohesive rea- as hypothesize, believe, expect, hope, desire,
AI researchers have made several attempts soning purposes. and fear). Propositional attitude terms take
to characterize the essence of what it means Second, ontologies enable knowledge propositions as arguments. Continuing with
to have a content theory. McCarthy and sharing. Suppose we perform an analysis and the electronic-device domain, we can assert,
Hayes’ theory (epistemic versus heuristic dis- amve at a satisfactory set of conceptualiza- for example: the diagnostician hypothesizes
t i n ~ t i o n )Marr’s
,~ three-level theory (infor- tions, and their representativeterms, for some or believes that part 2 is broken, or the
mation processing, strategy level, algorithms area of knowledge-for example, the elec- designer expects or desires that the power
and data structures level, and physical mech- tronic-devices domain. The resulting ontol- plant has an output of 20 megawatts. Thus,
anisms level): and Newell’s theory (Knowl- ogy would likely include domain-specific an ontology can represent beliefs, goals,

JANUARY /FEBRUARY 1999 21

 ~~

On the one hand there are entities, such as processes and events, which have temporal
parts.... On the other hand there are entities, such as material objects, which are always pre-
sent in their entirety at any time at which they exist at all. The categorical distinctionbetween
entities which do, and entities which do not have temporal parts is grounded in common
sense.Yet various philosophers have been inclined to oppose it. Some ... have defended an entities exist. The modeler of the domains
ontology consisting exclusively of things with no temporal parts. Whiteheadians have favored makes these commitments.As we move from
ontologies including only temporally extended processes. Quine has endorsed a four-dimen- an ontology’s top to lower taxonomic levels,
sional ontology in which the distinctionbetween objects and processes vanishes and every
commitments specific to domains and phe-
entity comprises simply the content of some arbitrarilydemarcated portion of space-time.
One further option, embraced by philosophers such as David Lewis, accepts the opposition nomena appear. For modeling objects on
between objects and processes, while still finding a way to allow that all entities have both earth, we can make certain commitments.For
spatial and temporal parts. example, animals, minerals, and plants are
subcategoriesof objects; has-life(x)and con-
tains-carbon(x)are object properties; and
Figure 1. Call for papers for a special issue on temporal parts for The Monist, An Infernofiono/(luorfer/yJourno/of
Generol Phi/osophico/Inquiry. This quote suggests that ontology has always been an issue of deep concern in philoso. can-eat(x, y ) is a possible relation between
phy and that the issues continue to occupy tontemporary philosophers. any two objects. These commitments are spe-
cific to objects and phenomena in this do-
main. Further, the commitments are not arbi-
hypotheses, and predictions about a domain, ogy in philosophy. For example,many ontolo- trary. For them to be useful, they should
in addition to simple facts. The ontology also gies have thing or entity as their root class. reflect some underlying reality.
plays a role in describing such things as plans However, Figure 2 illustrates that thing and There is no sharp division between do-
and activities, because these also require entity start to diverge at the next level. main-independent and domain-specific on-
specification of world objects and relations. For example, CYC’s thing has the subcat- tologies for representing knowledge. For
Propositional attitude terms are also part of egories individual object, intangible,and rep- example, the terms object, physical object,
a larger ontology of the world, useful espe- resented thing; the Generalized Upper device, engine, and diesel engine all describe
cially in describing the activities and prop- Model’s’ (GUM) um-thing has the subcate- objects, but in an order of increasing domain
erties of the special class of objects in the gories configuration,element, and sequence; specificity. Similarly, terms for relations
world called “intensional entities”-for Wordnet’s* thing has the subcategories liv- between objects can span a range of speci-
example, agents such as humans who have ing thing and nonliving thing, and Sowa’s ficity, such as connected, electrically-con-
mental states. root T has the subcategories concrete, pro- nected, and soldered-to.
Constructing ontologies is an ongoing cess, object, and abstract. (Natalya Fridman
research enterprise. Ontologies range in Noy’s and Carol Hafner’s article discusses Subtypes of concepts. Ontologies generally
abstraction, from very general terms that these differences more fully?) Some of these appear as a taxonomic tree of conceptual-
form the foundation for knowledge repre- differences arise because not all of these izations, from very general and domain-
sentation in all domains, to terms that are ontologies are intended to be general-pur- independent at the top levels to increasingly
restricted to specificknowledge domains. For pose tools, or even explicitly to be ontolo- domain-specific further down in the hierar-
example, space, time,parts, and subparts are gies. Another reason for the differences is chy. We mentioned earlier that different
terms that apply to all domains; malfunction that, in principle, there are many different ontologiespropose different subtypes of even
applies to engineeringor biological domains; taxonomies. very general concepts. This is because, as a
and hepatitis applies only to medicine. Although differences exist within ontolo- rule, different sets of subcategorieswill result
Even in cases where a task might seem to gies, general agreement exists between on- from different criteria for categorization. Two,
be quite domain-specific, knowledge repre- tologies on many issues: among many, alternate subcategorizationsof
sentation might call for an ontology that des- the general concept object are physical and
cribes knowledge at higher levels of gener- * There are objects in the world. abstract, and living and non-living. In some
ality. For example, solving problems in the Objects haveproperties or attributes that cultures and languages, words for objects
domain of turbines might require knowledge can take values. have gender, thus creating another top-level
expressed using domain-general terms such Objects can exist in various relations with classification along the gender axis. We can
as flows and causality. Such general-level each other. easily think of additional subcategorizations
descriptive terms are called the upper ontol- Properties and relations can change over based on other criteria. The existence of alter-
ogy or top-level ontology. There are many time. nate categorizationsonly becomes more acute
open research issues about the correct ways There are events that occur at different as we begin to model specific domains of
to analyze knowledge at the upper level. To time instants. knowledge. For example, we can subcatego-
provide some idea of the issues involved, There areprocesses in which objects par- rize causal process into continuous and dis-
Figure 1 excerpts a quote from a recent call ticipate and that occur over time. crete causal processes along the dimension
for papers. The world and its objects can be in dif- of how time is represented, and into mechan-
Today, ontology has grown beyond philos- ferent states. ical, chemical, biological, cognitive, and
ophy and now has many connectionsto infor- Events can cause other events or states as social processes along the dimension of the
mation technology. Thus, research on ontol- effects. kinds of objects and relations involved in the
ogy in AI and information systems has had to Objects can have parts. description.
produce pragmatically useful proposals for In principle, the number of classification
top-level ontology. The organizationof a top- The representational repertoire of objects, criteria and distinct subtypes is unlimited,
level ontology contains a number of problems, relations, states, events, and processes does because the number of possible dimensions
similar to the problems that surround ontol- not say anything about which classes of these along which to develop subcategories can-

22 IEEE INTELLIGENT SYSTEMS

 CYC Wordnet

not be exhaustively specified.Often, this fact

is not obvious in general-purposeontologies, Individual object Intangible Represented Living Nonliving
because the top levels of such ontologies
commit to the,most commonly useful sub- GUM Sowa’s
Um- Thing

&
types. However, domain-specific ontologies
can contain categorizationsalong dimensions
that are usually outside the general ontology. Concrete Process Object Abstract
Configuration Element Sequence
Task dependenceof ontologies.How task-
dependent are ontologies? Presumably, the ~~~~

kinds of things that actually exist do not Figure 2. Illustration of how ontologies differ in their analyses of the most general concepts.
depend on our goals. In that sense, ontologies
are not task-dependent. On the other hand,
what aspects of reality are chosen for encod- Richard E. Fikes describe KIF (Knowledge erties, and relations in that world. Data struc-
ing in an ontology does depend on the task. InterchangeFormat), an enabling technology tures and procedures implicitly or explicitly
For example, in the domain of fruits, we that facilitates expressing domain factual make commitments to a domain ontology. It
would focus on particular aspects of reality if knowledge using a formalism based on aug- is common to ask whether a payroll system
we were developing the ontology for the mented predicate calculus.‘ Robert Neches “knows” about the new tax law, or whether a
selection of pesticides; we would focus on and his colleagues describe a knowledge-shar- database system “knows” about employee
other aspects of reality if we were develop- ing initiative,12while Thomas R. Gruber has salaries. Information-retrievalsystems, digi-
ing an ontology to help chefs select fruits for proposed a language called Ontolingua to help tal libraries, integration of heterogeneous
cooking. In ontologies for engineering appli- construct portable ontologies.l 3 In Europe, the information sources, and Internet search
cations, categorizing causal processes into CommonKADS project has taken a similar engines need domain ontologies to organize
those that do, and that do not, produce dan- approach to modeling domain kn~wledge.’~ information and direct the search processes.
gerous side effects might be useful. Design These languages use varieties of predicate For example, a search engine has categories
engineers and safety analysts might find this calculus as the basic formalism. Predicate and subcategories that help organize the
a very useful categorization, though it is calculus facilitates the representation of search. The search-engine community com-
unlikely to be part of a general-purposeontol- objects, properties, and relations. Variations monly refers to these categories and subcate-
ogy’s view of the causal process concept. such as situational calculus introduce time gories as ontologies.
Practically speaking, an ontology is un- so as to represent states, events, and pro- Object-oriented design of software sys-
likely to cover all possible potential uses. In cesses. If we extend the idea of knowledge tems similarly depends on an appropriate
that sense, both an ontology for a domain and to include images and other sense modali- domain ontology. Objects, their attributes,
a knowledgebase written using that ontology ties, we might need radically different kinds and their procedures more or less mirror
are likely to be more appropriate for certain of representation. For now, predicate calcu- aspects of the domain that are relevant to the
uses than others and unlikely to be sharable lus provides a good starting point for ontol- application. Object systems representing a
across wide& divergent tasks. This is, by now, ogy-sharing technologies. useful analysis of a domain can often he
a truism in KBS research and is the basic Using a logical notation for writing and reused for a different application program.
insight that led to the current focus on the rela- sharing ontologies does not imply any com- Object systems and ontologies emphasize
tionship between tasks and knowledge types. mitment to implementing a related knowl- different aspects, but we anticipate that over
Presuppositionsor requirementscan be asso- edge system or a related logic. We are simply time convergence between these technolo-
ciated with problem-solvingmethods for dif- taking a knowledge-level5stance in describ- gies will increase. As information systems
ferent tasks so that they can capture explicitly ing the knowledge system, whatever the model large knowledge domains, domain
the way in which ontologies are task-depen- means of implementation. In this view, we ontologies will become as important in gen-
dent. For example, a method might have a pre- can ask of any intelligent system, even one eral software systems as in many areas ofAI.
supposition (or assumptionlo) stating that it implemented as a neural network, “What In AI, while knowledge representationper-
works correctly only if the ontology allows does the system know?” vades the entire field, two application areas
modeling causal processes discretely. There- in particular have depended on a rich body of
fore, assumptionsare a key factor in practical knowledge. One of them is natural-language
sharing of ontologies. Use of ontologies understanding.Ontologies are useful in NLU
in two ways. First, domain knowledge often
In AI, knowledge in computer systems is plays a crucial role in disambiguation.A well-
Technology for ontology thought of as something that is explicitlyrep- designed domain ontology provides the basis
sharing resented and operated on by inference pro- for domain knowledge representation. In
cesses. However, that is an overly narrow addition, ontology of a domain helps identify
view. All information systems tr&c in knowl- the semantic categories that are involved in
There have been several recent attempts to edge. Any software that does anything useful understanding discourse in that domain. For
create engineeringframeworksfor construct- cannot be written without a commitment to a this use, the ontology plays the role of a con-
ing ontologies. Michael R. Genesereth and model of the relevant world-to entities, prop- cept dictionary.In general, for NLU, we need

JANUARY/FEBRUARY 1999 23
 Related work for Knowledge Sharing,” Int’l J. Human-Computer Studies, Vol. 43,
Nos. 516, Nov.-Dec. 1995, pp. 907-928.
The field of ontology attracts an interdisciplinary mix of researchers, R. Studer, V.R. Benjamins, and D. Fensel, “Knowledge Engineering,
both from academia and industry. Here we give a selection of references Principles, and Methods,” Data and Knowledge Eng., Vol. 25, Mar.
that describe related ontology work. Because the literature is vast, a com- 1998, pp. 161-197.
plete list is impossible. For an extensive collection of (alphabetically M. Uschold and M. Gruninger, “Ontologies: Principles, Methods,
ordered) links to ontological work, including proceedings and events, see and Applications,” Knowledge Eng. Rev.. Vol. 11, No. 2, Mar. 1996,
http://www.cs. utexas. edu/u.sers/mj&b/related.html. pp. 93-155.

Special issues on ontology Natural-languageontology

N. Guarino and R. Poli, “The Role of Ontology in the Information J.A. Bateman, B. Magini, and F. Rinaldi, “The Generalized Upper
Technology,” Int’l J. Human-Computer Studies, Vol. 43, Nos. 5/6, Model,” Working Papers 1994 European Con$ Artificial Intelligence
Nov.-Dec. 1995, pp. 623-965. (ECAI ‘94)Workshop on Implemented Ontologies, 1994, pp. 34-45;
G.Van Heijst, A.T. Schreiber, and B.J. Wielinga, “Using Explicit http://ww w.darmstadt.gmd.de/publish/komet/papers/ecai94.p~.
Ontologies in KBS Development,” Int’l J. Human-Computer Studies, K. Knight and S. Luk, ‘‘Building a Large-scale Knowledge Base for
Vol. 46, Nos. 213, Feb.-Mar. 1997. pp. 183-292. Machine Translation,” Proc. AAAI ’94, AAAI Press, Menlo Park,
M. Uschold and A. Tate, “Putting Ontologies to Use,” Knowledge Calif. 1994.
Eng. Rev.,Vol. 13, No. 1, Mar. 1998, pp. 1-3. G.A. Miller, “Wordnet: An Online Lexical Database,” Int’l J.
Lexicography, Vol. 3 , No. 4, 1990, pp. 235-312.
P.E. Van de Vet, P.H. Speel, and N.J.I. Mars, “The Plinius Ontology of
Ontology development Ceramic Materials,” Working Papers 1994 European Con$ Artificial
Intelligence (ECAI ’94)Workshop on Implemented Ontologies,
J. Benjamin et al., “Ontology Construction for Technical Domains,” ECCAI, Amsterdam, 1994, pp. 187-206.
Proc. EKAW ’96: European Knowledge Acquisition Workshop, k c -
ture Notes in Artificial Intelligence No. 1076, Springer-Verlag.
Berlin, 1996, pp. 98-1 14.
Ontologies and informationsources
W.N.Borst and J.M. Akkermans, “Engineering Ontologies,” Int’l J.
Human-Computer Studies, Vol. 46, Nos. 213, Feb.-Mar. 1997, pp. Y. Arens et al., “Retrieving and Integrating Data from Multiple Infor-
365406. mation Sources,” Int‘l J. Intelligent and Cooperutive Information
A. Farquhar, R.Fikes, and J. Rice, “The Ontolingua Server: A Tool
for Collaborative Ontology Construction,” Int’l J. Human-Computer - Systems, Vol. 2 , No. 2, 1993, pp. 127-158.
S . Chawathe, H.Garcia-Molina, and J. Widom, “Flexible Constraint
Studies, Vol. 46, No. 6, June 1997, pp. 707-728. Management for Autonomous Distributed Databases,” IEEE Data
A. Gomez-Perez, A. Fernandez, and M.D. Vicente, “Towards a Eng. Bulletin, Vol. 17, No. 2, 1994, pp. 23-27.
Method to Conceptualize Domain Ontologies,” Working Notes 1996 S. Decker et al., “Ontobroker: Ontology-Based Access to Distributed
European Con$ Artificial Intelligence (ECAI ’96)Workshop on Onto- and Semi-Structured Information,” Semantic Issues in Multimedia

- logical Eng., ECCAI, Budapest, Hungary, 1996, pp. 41-52.

T.R. Gruber, “Towards Principles for the Design of Ontologies Used
Systems, R. Meersman et al., eds., Kluwer Academic Publishers,
Boston, 1999.

both a general-purpose upper ontology and a edge in addition to domain-specific knowl- enable them to choose among altemative rea-
domain-specific ontology that focuses on the edge. The initial motivation for CYC was to soning paths. Ontology specification in
domain of discourse (such as military com- provide such a body of sharable common- knowledge systems has two dimensions:
munications or business stories). CYC, Word- sense knowledge for knowledge-based sys-
net,8 and SensusI5 are examples of sharable tems. There is a similar need for developing 9 Domain factual knowledge provides
ontologies that have been used for language domain-specific knowledge. Thus, ontology- knowledge about the objective realities in
understanding. based knowledge-base development provides the domain of interest (objects, relations,
Knowledge-based problem solving is the a double advantage. The ontologies them- events, states, causal relations, and so
second area in AI that is a big consumer of selves are sharable. With these ontologies, forth).
knowledge. KBPS systems solve a variety of we can build knowledge bases using the Problem-solving knowledge provides
problems-such as diagnosis, planning, and structure of conceptualizations to encode knowledge about how to achieve various
design-by using a rich body of knowledge. specific pieces of knowledge. The knowledge goals. A piece of this knowledge might be
Currently, KBPS systems employ domain- bases that we develop using these ontologies in the form of a problem-solving method
specific knowledge, which is often sufficient can be shared more reliably, because the for- specifying-in a domain-independent
for constructing knowledge systems that tar- mal ontology that underlies them can help manner-how to accomplish a class of
get specific application areas and tasks. How- clarify the representation’s semantics. goals.
ever, even in specific application areas, Information systems and NLU systems
knowledge systems can fail catastrophically need factual knowledge about their domains Most early research in KBPS mixed fac-
when they are pushed to the edge of the capa- of discourse. The inferences they make are tual and problem-solving knowledge into
bility of the domain-specific knowledge. In usually simple. Problem-solving systems, in highly domain-specificrules, called domain
response to this particular shortcoming, contrast, engage in complex sequences of knowledge. As research progressed, it be-
researchers have proposed that problem- inferences to achieve their goals. Such sys- came clear that there were systematic com-
solving systems need commonsense knowl- tems need t o have reasoning strategies that monalities in reasoning strategies between

24 IEEE INTELLIGENT SYSTEMS

 S. Luke et al., “Ontology-Based Web Agents,” Proc. F i h Int’l Con$ Reuse,” Int’l J. Human-Computer Studies, Vol. 41, No. 3, Sept. 1994,
Autonomous Agents, ACM Press, New York, 1997, pp. 59-66; http:ll pp. 399-424.
www.cs.umd.edu/projects/pluslSHOEl1997. A. Tate, “Roots of SPAR-Shared Planning and Activity Represen-
S.T. Polyak et al., “Applying the Process Interchange Format (PIF) to tation,”Knowledge Eng. Rev., Vol. 13, No. 1, Mar. 1998, pp.
a Supply Chain Process Interoperability Scenario,” Proc. 1998 Euro- 121-128.
pean Con$ Artificial Intelligence (ECAI ’98)Workshop on Applica- Y.A. Tijerino and R. Mizoguchi, “Multis 11: Enabling End-Users
tions of Ontologies and Problem-Solving Methods, ECCAI, Brighton, to Design Problem-Solving Engines via Two-Level Task Ontolo-
England, 1998, pp. 88-96. gies,” Proc. EKAW ‘93: Seventh European Workshop on Know-
G. Wiederhold, “Intelligent Integration of Information,” J. Intelligent ledge Acquisition f o r Knowledge-Based Systems, Lecture Notes in
Information Systems, Vol. 6, Nos. 213, 1996. Artificial Intelligence No. 723, Springer-Verlag, 1993, pp.
G. Wiederhold and M. Genesereth, “The Conceptual Basis for Medi- 340-359.
ation Services,” IEEE Intelligent Systems, Vol. 12, No. 5, Sept./Oct.
1997, pp. 38-47.

Ontology workshops
Ontologies and knowledge management Applications of Ontologies and Problem-Solving Methods, ECAI ’98
(European Conf. AI), http://delicias.dia.ji. upm.es/WORKSHOP/
A. Abecker et al., “Toward a Technology for Organizational Memories,”
ECAI98/index. html
IEEElntelligent Systems, Vol. 13, No. 3, MayIJune 1998, pp. 4 M 8 .
Building, Maintaining, and Using Organizational Memories, ECAI
V.R. Benjamins and D. Fensel, “The Ontological Engineering Initia-
’98, http://www. aijb.uni-karlsruhe. de/WBS/ECA198OM/
tive (KA)2,” Formal Ontology in Information Systems, N. Guarino,
Formal Ontologies in Information Systems (FOIS ’98), hrtp://krzirst.
ed., 10s Press, Amsterdam, 1998, pp. 287-301.
itc. it: 1024/fois98/program. html
M.S. Fox, J. Chionglo, and F. Fadel, “A Common-Sense Model of the
Intelligent Information Integration, ECAI ’98, http://www. tzi.de/
Enterprise,” Proc. Industrial Eng. Research Con$, Inst. for Industrial
grp/i3/ws-ecai98/
Engineers, Norcross, Ga., 1993, pp. 425429.
Sharable and Reusable Components for Knowledge Systems, KAW
Manual of the Toronto Virtual Enterprise, tech. report, Enterprise
’98 (Workshop on Knowledge Acquisition, Modeling, and Manage-
Integration Laboratory, Dept. of Industrial Eng., Univ. of Toronto,
ment), http://ksi.cpsc.ucalgary.cu/KAW/kAW98/KAW98Proc. html
Toronto, 1995.
Ontological Engineering, AAAI Spring Symp. Series, Stanford,
M. Uschold et al., “The Enterprise Ontology,” Knowledge Eng. Rev.,
Calif ., 1997, http://www.aaai.org/SymposidSpring/I 997/sss-97. html
Vol. 13,No. 1,Mar. 1998.
Problem-Solving Methods, IJCAI ’97 (Int’l Joint Conf. AI), htfp://
www.aifl. uni-karlsruhe. de/WBS/dfe/PSM/main. html
Ontological Engineering, ECAI ’96, http://wwwis.cs.utwente.nl:
Task and method ontologies
D. Fensel et al., “Using Ontologies for Defining Tasks, Problem-
Solving Methods, and Their Mappings,” Knowledge Acquisition,
. 8080kbs/EcaiWorkshophomepage.html
Practical Aspects of Ontology Development, AAAI ’96
Sharable and Reusable Ontologies, KAW ’96, http://ksi.cpsc.
Modeling, and Management, E. Plaza and V.R. Benjamins, eds., ucalgary.cdKAW/KAW96/KAW96Proc. html
Springer-Verlag, Berlin, 1997, pp. 113-128. Sharable and Reusable Problem-Solving Methods, KAW ’96, http://
J.H. Gennari et al., “Mapping Domains to Methods in Suppport of ksi. cpsc.ucalgaryca/KAW/KAW96/kAW96Proc. html

goals of similar types. These reasoning ods appropriate for different types of prob- References
strategies were also characterized by their lems fueled second-generation research in
need for specific types of domain factual knowledge systems.Is M o s t of the KBPS 1. D.B. Lenat and R.V. Guha, Building Large
knowledge. It soon became clear that strate- community’s work on knowledge represen- Knowledge-Based Systems: Representation
gic knowledge could be abstracted and tation is not well-known t o the general and Inference in the CYC Project, Addison-
reused. knowledge-representation community. In the Wesley, Reading, Mass., 1990.
With few exceptions,I6.l 7 the domain fac- coming years, we expect an increased focus 2. B. Chandrasekaran, “AI, Knowledge, and the
tual knowledge dimension drives the focus o n method ontologies as a sharable knowl- Quest for Smart Systems,” IEEE Expert, Vol.
of most of the AI investigations on ontolo- edge resource. 9, No. 6, Dec. 1994, pp. 2-6.
gies. This is because applications to language
3 . J . McCarthy and P.J. Hayes, “Some Philo-
understanding motivates much of the work
sophical Problems from the Standpoint of
on ontologies. Even CYC, which was origi- Artificial Intelligence,” Machine Intelligence
nally motivated by the need for knowledge Acknowledgments Vol. 4, B. Meltzer and D. Michie, eds., Edin-
systems to have world knowledge, has been burgh University Press, Edinburgh, 1969, pp.
tested m o r e in natural-language than in This article is based on work supported by the 463-502.
knowledge-systems applications. Office of Naval Research under Grant N00014-96-
1-0701.We gratefully acknowledge the support of 4. D. Man; Vision: A Computational Investiga-
ONR and the DARPA RaDEO program. Any opin- tion into the Human Representation and Pro-
ions, findings, and conclusions or recommenda- cessing of Visual Information, W.H. Freeman,
tions expressed in this publication are those of the San Francisco, 1982.
PS R E S E A R C H E R S R E A L I Z E D authors and do not necessarily reflect the views of
ONR. Netherlands Computer Science Research 5. A. Newell, “The Knowledge Level,” Artifi-
that, in addition to factual knowledge, there Foundation supported Richard Benjamins with cial Intelligence, Vol. 18, 1982, pp. 87-127.
is knowledge about how to achieve problem- financial support from the Netherlands Organiza-
solving goals. In fact, this emphasis on meth- tion for Scientific Research ( W O ) . 6. R. Wieringa and W. de Jonge, “Object Iden-

JANUARY/FEBRUARY 1999 25
 . ...__..- - .

Inf elligen t Agents

Guest Editor Jim Hendler of DARPA will present articles discussing development techiques for and the
practical application of intelligent agents, which might be the solution to handling the data and
information explosion brought about by the Internet. Scheduled topics include
Agents for the masses: Is it possible to develop sophisticated agents simple enough
to be practical?
Extempo Systems’ interactive characters, who engage, assist, and entertain people
on the Web
Agentsoft’s efforts at commercializing intelligent agent technology

lnfelligent Systemswill also continue its coverage of the ontologies track, started in this issue, and
the track on vision-based vehicle guidance, which began in the NovembedDecember 1998 issue.

/FEE lntelligenf Systems covers the full range of intelligent system developments for the AI
practitioner, researcher, educator, ond user.

/E€€ Intelligent Systems

tifiers, Keys and Surrogates: Object Identi- Dec. 1994, pp, 28-37. 43210; [email protected]; http://www.
fiers Revisited,” Theory and Practice of cis.ohio-state.edu/lair/.
Object Systems (TAPOS),Vol. 1, No. 2,1995, 15. K. Knight and S. Luk. “Building a Large-
pp. 101-114. Scale Knowledge Base for Machine Transla- John R. Josephson is a research scientist and the
tion,” Pmc. Am. Assoc. Arti$cial Intelligence, associate director of the Laboratory for AI
7. J.A. Bateman, B. Magini, and F. Rinaldi, ‘The AAAI Press, Menlo Park, Calif., 1994. Research in the Department of Computer and
Generalized Upper Model,” Working Papers Information Science at Ohio State University. His
16. D. Fensel et al., “Using Ontologies for Defin-
1994 European Con5 Artificial Intelligence ing Tasks, Problem-Solving Methods and primary research interests are knowledge-based
(ECAI ‘94)Workshopon Implemented Ontolo- systems, abductive inference, causal reasoning,
Their Mappings,” Knowledge Acquisition,
gies, 1994, pp. 3 4 4 5 ; http://www.darmstadt theory formation, speech recognition, perception,
Modeling and Management, E. Plaza and V.R.
.gmd.de/publishflcomet/papers/ecai94.p~. diagnosis, the logic of investigation, and the foun-
Benjamins, eds., Springer-Verlag, New York,
dations of science. He received his BS and MS in
8. G.A. Miller, “Wordnet: An Online Lexical 1997, pp. 113-128.
mathematics and his PhD in philosophy, all from
Database,” Int’l J. Lexicography, Vol. 3, No. 17. R. Mizoguchi, J. Van Welkenhuysen, and M. Ohio State University. He has worked in several
4,1990, pp. 235-312. Ikeda, “Task Ontology for Reuse of Problem application domains, including medical diagnosis,
Solving Knowledge,” Towards Very Large medical test interpretation, diagnosis of engineered
9. N. Fridman Noy and C.D. Hafner, “The State systems, logistics planning, speech recognition,
of the Art in Ontology Design,” AZMagazine, Knowledge Bases, N.J.I. Mars, ed., 10s Press,
Amsterdam, 1995. molecular biology, design of electromechanical
Vol. 18, No. 3, 1997, pp. 53-74. systems, and interpretation of aerial photographs.
10. D. Fensel and V.R. Benjamins, “The Role of 18. J.M. David, J.P. Krivine, and R. Simmons, He is the coeditor with Susan Josephson ofdbduc-
Assumptions in Knowledge Engineering,” Second Generation Expert Systems, Springer- tive Inference: Computation, Philosophy, Tech-
Int’l J. Intelligent Systems, Vol. 13, No. 7, Verlag, 1993. nology, Cambridge Univ. Press, 1994. Contact him
1998, pp. 715-747. at the Laboratory for AI Research, Ohio State
Univ., Columbus, OH, 43210; [email protected]
1 1. M.R. Genesereth and R.E. Fikes, Knowledge B. Chandrasekaran is professor emeritus, a statcedu; http://www.cis.ohio-state.edullair/.
Interchange Format, Version 0.3, Knowledge senior research scientist, and the director of the
Systems Lab., Stanford Univ., Stanford, Calif. Laboratory for AI Research (LAIR) in the Depart- Richard Benjaminsis a senior researcher and lec-
1992. ment of Computer and Information Science at turer at the Department of Social Science Infor-
Ohio State University. His research focuses on matics at the University of Amsterdam. His
12. R.Neches et al., “Enabling Technology for knowledge-based systems, causal understanding, research interests include knowledge engineering,
Knowledge Sharing,” AZ Magazine, Vol. 12, diagrammatic-reasoning,and cognitive architec- problem-solving methods and ontologies, diagno-
NO. 3, 1991, pp. 36-56. tures. He received his BE from Madras University sis and planning, and AI and the Web. He obtained
13. T.R. Gruber, “A Translation Approach to and his PhD from the University of Pennsylvania, his BS in cognitive psychology and his PhD in arti-
Portable Ontology Specifications,” Knowl- both in electrical engineering. He was Editor-in- ficial intelligence from the University of Amster-
edge Acquisition, Vol.5, 1993, pp. 199-220. Chief of IEEE Expert from 1990 to 1994, and he dam. Contact him at the Dept. of Social Science
serves on the editorial boards of numerous inter- Informatics, Univ. of Amsterdam, Roetersstraat
14. G. Schreiber et al., “CommonKADS: A national journals. He is a fellow of the IEEE, 15, 1018 WB Amsterdam, The Netherlands;
Comprehensive Methodology for KBS AAAI, and ACM. Contact him at the Laboratory [email protected]; http://www.swi.psy.uva.
Development,” ZEEE Expert, Vol. 9, No. 6, for AI Research, Ohio State Univ., Columbus, OH, nl/usr/richard/home.html.

26 IEEE INTELLIGENT SYSTEMS