Knowledge Graph Considered Harmful for

Ontology

Seiji Koide1

Ontolonomy, LLC., Minami-ku Yokohama 232-0066, Japan,

[email protected],
WWW home page: http://ontolonomy.co.jp/

Abstract. The time of knowledge graph has come. Linked Data is now
rephrased to knowledge graph, and no one has doubt on the future of it.
However, will the time of ontology come in the next step? There exists
a long history of ontology ever since the ancient Greece, and now the
terminology has become popular with the advent of OWL. Nevertheless,
the ontology does not seem to happen so as Linked Data did. There is
a serious gap on the semantic representation between the ontology and
the knowledge graph, and the gap originates semantic networks. In this
paper, we pursue the history of knowledge representation, point out the
serious semantic gap contained in knowledge graph. We propose an al-
ternative representation language for ontological knowledge in Semantic
Webs.

Keywords: New KM, RDF, OWL, knowledge graph, knowledge repre-

sentation, Frame-based, Case-based

1 Introduction

The purpose of this essay is to cause a stir in the community of Semantics

Webs just as Dijkstra did in software engineering.1 It seems that the time of
Linked Open Data (LOD) has come. Ones have become to rephrase “Linked
Data” to “knowledge graph”, since knowledge graph has been used as technical
terminology in the domain of semantic search by Google, IBM, etc.. Because the
both is roughly the same technology from the technical viewpoint. On the other
hand, the term of “ontology” exists ever since the age of ancient Greece, and
the engineering of ontologies has been pursued since 1970s as a part of computer
science and Artificial Intelligence, then it has become popular today with OWL.
However, we have yet a serious gap between the representation of knowledge
graph and ontology. More precisely, we have theoretical, practical, and socio-
technological difficulties such as how to understand the subsumption concept
described in RDF Semantics [2] with respect to the ambiguous IS-A concept [3,
4], how to represent structural link [3, 5] in the form of knowledge graph, and
how to discriminate the notion of class and individual in human mind.
1
See, https://en.wikipedia.org/wiki/Considered_harmful
 The following is a fallible example which novices easily fall into erros.2

Wine a owl:Class ;
rdfs:subClassOf food:PotableLiquid ;
*madeFromGrape WineGrape .

Instead, it should be exactly coded in turtle as follows.

Wine a owl:Class ;
rdfs:subClassOf food:PotableLiquid ,
[ a owl:Restriction ;
owl:onProperty madeFromGrape ;
owl:allValuesFrom WineGrape ] .

The cause of this mistake is three fold; firstly, the abstract syntax of RDF
(namely RDF graph) does not fit to the description of ontologies; secondly, it
it hard to understand the concept of subsumption and property inheritance;
thirdly, most people are sluggish to study hard, some one tends to take an easier
way even if it is impossible to reach the final goal.
Contrary to the error sentence mentioned above, the following code is com-
pletely correct with respect to individuals.

ElyseZinfundel a Zinfundel ;
hasMaker Elyse ;
hasSugar Dry.

In this case, Zinfundel is a class, but ElyseZinfundel, Elyse, and Dry are
individuals. There is no inheritance on properties. Actually, at the error case
shown above, it should be coded with the meanings that every wine (as instance)
is made from some wine-grape (as instance), and it must be described at the class
level on wine. Otherwise, it is forced to describe the property values on every
instances. It should be here noticed that Wine and WineGrape are classes.
Such a kind of problems was involved in the beginning of semantic networks.
Woods [3] pointed out the semantical ambiguity of network links and introduced
the distinction of assertional link and structural link. At the final step of KL-ONE
family, Clark developed a knowledge representation language, KM (Knowledge
Machine) [9, 10]. According to the style in KM-like representation, the error
sentence mentioned above may be paraphrased into the followings.

Wine is a owl:Class and rdfs:subClassOf food:PotableLiquid .

every Wine is madeFromGrape a WineGrape .

Note that these sentences are not English, rather an artificial knowledge repre-
sentation language. It is a sort of syntax sugar of the turtle syntax and preserves
2
This example is taken from Wine Ontology [6], and where an asterisk is attached to
the head of wrong lines.
RDF and OWL semantics. We can translate KM-like sentences into turtle sen-
tences, just as turtle sentences can be translated into RDF/XML format without
the loss of information.
It is conceived that Description Logics inherited the heritage of semantic
networks and frame systems [4]. In fact, many features of KL-ONE family came
into OWL. In this paper, we pursue the history of knowledge representation,
specifically focusing on the development of class notion and subsumption con-
cept, then point out the problem of knowledge graph representation for OWL.
Finally, we propose a new representation language, the New KM, a successor of
KM by Clark and SWCLOS (Semantic Web processor on top of Common Lisp
Object System) [11, 12].
The rest of this paper is organized as follows. In section 2, we present a
brief history of the knowledge representation on semantic networks and frame
systems, focusing on class-instance notions and the inheritance concept. section
3 presents an idea of New KM, which is a unified successor of KM by Clark and
our own Semantic Web Processor, SWCLOS. Finally, we conclude some remarks
and the future work.

2 Historical Views on Knowledge Representation

2.1 Semantic Networks to Description Logic and OWL
The study of semantic networks was started as an associative link network at
1966 by Quillian’s work [7, 8], for the purpose of making artificial memory for
words and meanings. At the time, such an associative link structure was taken as
a firm base in modeling human memory. However, many problems that were in-
volved in such simple networks were soon exposed by many researchers.3 Human
epistemology is more complex than the association.
Woods [3] analyzed the semantics of semantic networks and introduced the
distinction between structural links, which present propositional statements on
things, and assertional links on assertional relation among things. He addressed
the following network structure.

N12368
SUPERC TELEPHONE
MOD BLACK

The meanings of this sentence may be interpreted in two ways. In OWL, it would
be written distinctively as follows, into a proposition in TBox (left side) or an
assertion in ABox (right side).
N12368 N12368 a TELEPHONE ;
owl:intersectionOf hasColor BLACK .
( TELEPHONE
3
The original by Quillian was not so simple as the successors. The original had devices
of type, token, plane, and notions of class, subclass, and modification. [13]
 [ a owl:Restriction ;
owl:onProperty hasColor ;
owl:hasValue BLACK ] ) .

The former insists the class of black-colored telephone, but the latter asserts
the existence of a telephone whose color is accidentally black. On the idea that
the notation must be specified more precisely, he started the discussion on the
problem involved in semantic networks. However, the issues, in the viewpoints
at the present, were confused and spread out widely in distinct levels, from
semantics to pragmatics, from denotational levels to logical levels.
Up to the mid of 1970s, semantic networks constitute the primer knowl-
edge representation and many attempts revealed that they never lived up to
researcher’s expectations. Brackman [13] demonstrated complex semantic net-
works that contain both conceptual networks and their particulars, and showed
special existences a set of links that allows the specification of a concept as a
set of attribute definitions in conjunction with a structural interrelation between
those attributes, see details in [13].
At 1979, Brackman [14] presented a comprehensive survey on semantic net-
works. He investigated the work by Quillian, Woods, Brackman, Collins, Car-
bonell, Winston, expanding to Fillmore, Simmons, Hendrix, Rumelhart, Schank,
Heidom, Anderson, Shapiro, Cercone, Phillip Hayes, Norman, and Szolovits, then
clarified five levels of characteristics on semantic networks, “implementational,”
“logical,” “epistemological,” “conceptual,” and “linguistic” levels. Due to limita-
tions of space, we summarize the result at Table 1. Note that we can capture that

Table 1. Characteristics Levels of Semantic Networks

Level Primitives description

Implementational atoms, pointers A Link is a pointer and a node is a
destination.
Logical propositions, predicates, Logical primitives with a structured
logical operators index like AND, SUBSET, EXISTS
Epistemological concept types, subpieces, Formal structure of conceptual units
inheritance, and interrelationships as them.
structuring relations Independent of any knowledge edge
Conceptual semantic relations, Language-independent conceptual
primitive objects and actions primitives and case structure.
Linguistic arbitrary concepts, words, Networks whose primitives
expressions are language-specific.

RDFS and OWL fall into epistemological level, which was discovered by Brack-
man as a missing level in his comprehensive investigation. Usual ontologies that
are built using RDFS and OWL are at conceptual level. As an instantiation of
such an epistemological level, he advocated more elaborated “Structured Inher-
itance Networks”, in which Role/Filler Description and Structural Description
are derived.
Eventually, Brackman [15] published KL-ONE at 1985, and after that many
KL-ONE-ish systems succeeded. They are, as a whole, called KL-ONE family.
All systems of KL-ONE family, except KM [9, 10] is listed at [16].
The class-instance notion gradually emerged in the development of KL-
ONE family. In the original KL-ONE, the notion of class was the produc-
tion of a classifier, and did not provide any explicit primitives for the class-
instance indication. This situation is carried over Description Logics and OWL.
In CLASSIC [17], the operators for individuals were identical to that for classes,
but the function cl-create-ind was provided to create an individual under
a CLASSIC-description. Moreover, LOOM [18] prepared 17 operators (func-
tions/macros/slots) for instances. Ideas of objects and mixin classes were bor-
rowed from CLOS (Common Lisp Object System) and provided the mixin-
inheritance functionality.
KM [9, 10] is, which is not regarded so, the last system in the line of KL-ONE
family in thought. It has a frame-like syntax as well as CLASSIC and LOOM.
However, it presented an easier way to encode the inheritance attributes for
instances at classes. The following is actual programming code in KM of Buy
event.
(every Buy has
(buyer ((a Agent)))
(object ((a Thing)))
(seller ((a Agent)))
(money ((the cost of (the object of Self))))
(subevent1 ((a Give with
(agent ((the buyer of Self)))
(object ((the money of Self)))
(rcpt ((the seller of Self))))))
(subevent2 ((a Give with
(agent ((the seller of Self)))
(object ((the object of Self)))
(rcpt ((the buyer of Self)))))))
This code axiomatize that two agents appear as buyer and seller, every event of
Buy is accompanied by two distinctive Give events in which one event for a buyer
the money is received by the buyer, in the other event for a seller the object is
received by seller. In the above example, the event money is instantiated as the
cost of the event. KM denotes two fundamental types, instances and classes. A
class has the extension of the individuals, and properties of individuals of a class
are expressed of the form:
(every <class> has
(<slots1> (<expr11> <expr12> ... ))
(<slots2> (<expr21> <expr22> ... ))
...
 )

Thus, “(every ...” form describes properties for individuals of a class to be

inherited in accordance with superclass-subclass relation. On the other hand, the
form without “every” but with “superclasses” attribute for classes denotes
superclasses of a subjective class with other properties for the class per se.

(<class> has
(superclasses (<superclass1> <superclas2> ... ))
(<slots1> (<expr11> <expr12> ... ))
(<slots2> (<expr21> <expr22> ... ))
...
)

This grammar greatly reduces the burden of awkward expression in knowledge

graph. We propose such a grammar for RDF and OWL in Section 3.

2.2 Frames to RDFS

Minsky [19] published the idea of framework of human cognitive mechanisms at

1974.

We can think of a frame as a network of nodes and relations. The

top levels of a frame are fixed, and represent things that are always true
about the supposed situation. The lower levels have many terminals –
slots that must be filled by specific instances or data. [. . . ] Collections
of related frames are linked together into frame-systems. The effects of
important actions are mirrored by transformations between the frames
of a system. [Minskey, 1974]

As Minsky mentioned in his paper, the basic idea of frame is not his invention
and his presentation was not complete, but he pointed out several important
notions of frame systems such as sharing terminals, a frame and subframes,
variables, attachments, default assignment, and so on. He talked the image of
frame-based cognition in many scenes, vision, linguistics, memory acquisition,
retrieval of knowledge, and control. Minsky’s prevision had become the source
of many frame systems after that.
KRL and FRL were the first two systems embodied Minsky’s idea. Espe-
cially, Bobrow’s KRL[20] gave some inspiration to KL-ONE family, where the
appearance of network disappeared and frame-like forms, e.g., UNIT, appeared.
The followings are an example of event description described in KRL.

[Event234 UNIT Individual

<SELF (a Give with
object = (a Pen)
giver = (Person2 (which IsHusbandOf Person3)}
recipient = Person1)>]
[Give UNIT Specialization
<SELF (an Event)>
<object (a Thing)>
<giver (a Person)>
<recipient (a Person)>]

[Lawyer UNIT Specialization

<SELF (a Person)>]

[Pen UNIT Basic

<SELF (a PhysicalObject)>]

[Lawyer UNIT Specialization

<SELF (a Person)]

[Person1 UNIT Individual

<SELF (a Person with firstName = "David")>]

[Person2 UNIT Individual

<SELF {(a Person with firstName = "Jonathan")
(which IsHusbandOf Person3)}>]

[Person3 UNIT Individual

<SELF {(a Person with firstName = "Ellen")
(a Lawyer)}>]

KRL allowed to explicitly express both classes and individuals in a uniform

pattern named UNIT. It is obvious that “(a Pen)” denotes an instance of class
Pen. The form “(a <class> with <slot-name> = <slot-value>)” indicates
an instance of <class> has <slot-value> to the <slot-name> in both class
and individual descriptions.
FRL [21, 22] was an implementation for several representation techniques
suggested by Minsky and additional functionalities that are, today, conceived
common in frame systems. Namely, local procedure attachments, IF-NEEDED,
IF-ADDED, and IF-REMOVED were demonstrated in addition to basic func-
tionalities of default value and inheritance along AKO (a kind of). However,
the semantics of the inheritance, which should be turned into the subsumption
from the viewpoint of RDF Semantics [2], is not formalized, whereas it worked
practically.
Schank’s Memory Organization Package (MOP) [23] was a very unique pack-
age for Case-Based Reasoning (CBR). It is possible to deem it a successor of
frame systems, but the most strong impact of MOP on Semantic Webs is that it
has made clear the bunch of slots (a set of pairs of property-name and property-
value) of an instance define the class of the instance. In other words, the inten-
sion (slots) decides the extension (class). Actually, we utilized the algorithm of
slots->mop in MOP onto building Agriculture Activity Ontology (AAO) [24]
with SWCLOS. However, the MOP has no structural difference between classes
and instances and the actual deference of MOP between them is just the flag for
instance/class is true or false.
Common Lisp Object System (CLOS) [25, 26] is the first genuine class-based
object system, in which the instance object structure is completely different
from the class object structure, and we need the definition of a class before the
instance object creation. We can define Wine class in pure CLOS as follows.

(defclass Wine (food:PotableLiquid)

((madeFromGrape :accessor madeFromGrape :type WineGrape))
(:metaclass owl:Class))

Furthermore, if rdfs:label is defined as slot at the metaclass owl:Class

with the meta-object protocol [27], we can add the rdfs:label slot of class Wine
meta-object as follows.

(make-instance ’owl:Class :name ’Wine rdfs:label "Wine")

3 A Brief Overview of New Knowledge Machine

Along with the progression and popularization of Linked Open Data, more com-
prehensible and comprehensive tools are required for new entrants from ordinary
people. It is easy to write down knowledge graphs at instance levels, but it is
difficult to expand the knowledge graph to the ontology. It is obvious that we
need new guides who lead starters in LODs to the ontologies. For the purpose of
that, we are now tackling to develop a new platform of Semantic Webs that cov-
ers from LODs to ontologies. It will be an amalgam of a database for knowledge
graph, gently guiding for novices, straight-forward representation of knowledge,
etc. In this section, we give rough sketches of several parts of the tools.

3.1 Light Weight Database and Indexing in Knowledge Graph

The new platform has an interface to RDF stores, AllegroGraph, Virtuoso, Star-
dog, and so on. In addition, we have developed an internal graph memory as
cache based on DTREE and the unification4 . It demonstrated that Princeton
WordNet [28] was absorbed on memory for the machine of Intel Core i7-4770,
16GB. In this module, we can perform simple information retrievals on one shot
via simple pattern matching that is similar to SPARQL queries.
4
See Paradigms of Artificial Intelligence Programming by Peter Norvig Section 14.8.
3.2 Case-based Hierarchy Adjustment
In addition to domain and range constraints in RDFS, which was already equipped
on SWCLOS, we are going to implement flexible hierarchical structure adjust-
ment. Schank’s MOP provided the functionality of instance adjustment but not
class adjustment. We will expand such adjustment to the class level. The rules
of instantiation will work effectively at the class level.

3.3 New KM Language

We are developing a new KM language for Semantic Webs, which imitates the
behavior of Clark’s KM. Note that RDF semantics and OWL semantics are
strictly preserved in the New KM. Especially, it will allows to represent meta-
level ontologies [29] like:

Species is a owl:Class and rdfs:subClassOf owl:Class .

EndangeredSpecies is rdfs:subClassOf Species .
Eagle is a owl:Class and is a EndangeredSpecies .
Harry is a Eagle .

4 Conclusion and Future Work

This paper introduced a new knowledge representation machine, the New KM,
which is a successor of SWCLOS, Schank’s MOP, and Clark’s KM, and more.
It will ease the difficulties of entering Linked Open Data and proceeding to
ontologies and meta-modeling of ontologies.
This paper pursued the history of knowledge representation from the be-
ginning of semantic networks and frame systems, focusing on the evolution of
class-instance notion and the implementation on representation languages. This
paper rediscovered the five levels of semantic networks. We found that the high-
est level, Linguistic level, in which we can investigate any concepts and words as
networks, is left as unexplored domain. The comprehensive platform based on
RDF and OWL will be indispensable to build extensive ontologies from compre-
hensive web resources like Wikipedia, and others.

References
1. Franz Baader, Diego Calvanese, Deborah McGuinness, Daniele Nardi, Peter Patel-
Schneider (eds.): The Description Logic Handbook, Cambridge (2003)
2. Hayes, Patrick, Brian McBride: RDF Semantics, http://www.w3.org/TR/rdf-mt/
(2004)
3. Woods, William A.: What’s in a Link: Foundations for Semantic Networks, Repre-
sentation and Understanding; Studies in Cognitive Science, D. G. Bobrow, A. M.
Collins (eds.), 35–82, Academic Press (1975)
4. Nardi, Daniele, Ronald J. Brackman: An Introduction to Description Logics, in [1]
1–40.
5. Sattler, Ulrike, Diego Calvanese, Ralf Molitor: Relationships with other Formalisms,
in [1] 137–177.
6. Noy, Natalya F., Deborah L. McGuinness: Ontology Development 101: A Guide
to Creating Your First Ontology, https://protege.stanford.edu/publications/
ontology\_development/ontology101-noy-mcguinness.html, Stanford.
7. Quillian, M. Ross: Semantic Memory, Report AFCRL-66-189, BBN, Cambridge
(1966)
8. Quillian, M. Ross: Word Concepts: A Theory and Simulation of Some Basic Seman-
tics Capabilities, Behavioral Science, 12, 410–430 (1967), reprinted in Readings in
Knowledge Representation (eds.) R. J. Brackman, H. H. Levesque, 98–118, Morgan
Kaufmann (1985)
9. Clark, Peter, Bruce Porter: KM - The Knowledge Machine: Users Manual, Technical
Report, AL Lab, Univ. Texas at Austin (1999)
10. Clark, Peter, Bruce Porter: KM - The Knowledge Machine 2.0: Users Manual,
http://www.cs.utexas.edu/users/mfkb/RKF/km.html
11. Koide, Seiji, Masanori Kawamura: SWCLOS: A Semantic Web Processor on Com-
mon Lisp Object System, ISWC2004, demo, http://iswc2004.semanticweb.org/
demos/32/ (2004)
12. Koide, Seiji, Hideaki Takeda: OWL-Full Reasoning from an Object Oriented Per-
spective, ASWC2006, LNCS-4185, 263–277, Springer (2006)
13. Brackman, Ronald J.: What’s in a Concept: Structural Foundations for Semantic
Networks, Int. H. Man-Machine Studies, 9, 127–152 (1977)
14. Brackman, Ronald J.: On the Epistemological Status of Semantic Networks, Asso-
ciative Networks: Representation and Use of Knowledge by Computers, N. V. Findler
(ed.) 3–50, Academic Press (1979)
15. Brackman, Ronald J., James G. Schmolze: An Overview of the KL-ONE Knowledge
Representation System, Cognitive Science, 9, 171–216 (1985)
16. Woods, William A., James G. Schmolze: The KL-ONE Family, Computers Math.
Applic., 23-2-5, 133–177, Pergamon Press (1992)
17. Resnick, Lori A., Alex Borida, Ronald J. Brackman, Charles L. Isbell, Debo-
rah L. McGuinness, Peter F. Patel-Schneider, Kevin C. Zalondek: CLASSIC De-
scription and Reference Manual, https://cse.buffalo.edu/~shapiro/Courses/
CSE663/classicmanual.ps (1995)
18. Brill, David: Loom Reference Manual Version 2.0, ftp://ftp.isi.edu/pub/loom/
reference2.0-twoside.ps (1993)
19. Minsky, Marvin: A Framework for Representing Knowledge, Mind Design, J.
Haugeland (ed.), pp.95–128, MIT Press (1981) Originally published as AI-Memo
306 (1974), https://dspace.mit.edu/bitstream/handle/1721.1/6089/AIM-306.
pdf?sequence=2
20. Bobrow, Daniel G., Terry Winograd: An Overview of KRL, a Knowledge Repre-
sentation Language, Cognitive Science, 1-1, 3–46 (1977)
21. Roberts, R. Bruce, Ira P. Goldstein: The FRL Primer, AI Memo 408, MTI (1977)
22. Roberts, R. Bruce, Ira P. Goldstein: The FRL Manual, AI Memo 409, MIT (1977)
23. Riesbeck, Christopher K., Roger C. Schank: Inside Case-based Reasoning, Laurense
Eribaunm Associates (1989)
24. Joo, Sungmin, Seiji Koide, Hideaki Takeda, Daisuke Horyu, Akane Takezaki,
Tomokazu Yoshida: Designing of Ontology for Domain Vocabulary on Agriculture
Activity Ontology (AAO) and a Lessons Learned, JIST 2016, 32–46, Springer (2016)
25. X3J13 Document 88-002R: Common Lisp Object System Specification, (1988)
26. Pitmann, Kent: Hyperspec, http://www.lispworks.com/documntation/
Hyperspec/Body/04\_c.htm
27. Kiczales, Gregor, Jim des Rivières, Daniel G. Boblow: The Art of the Meaobject
Protocol, MIT (1991)
28. Koide, Seiji, Hideaki Takeda: RDFization of Japanese Electronic Dictionaries and
LOD, 2nd Workshop on Linked Data in Linguistics (LDL-3024), (2013)
29. Koide, Seiji, Hideaki Takeda: Inquiry into RDF and OWL Semantics, JIST 2016,
25–31, Springer (2016)