KNOWLEDGE ACQUISITION WITH OM
A Heuristic Solution
Adolfo Guzman-Arenas and Alma-Delia Cuevas-Rasgado
Centro de Investigación en Computación Av. Juan de Dios Batiz, s/n, Zacatenco, 07738 México City
México and Instituto Tecnológico de Oaxaca Av. Ing. Victor Bravo Ahuja 125 and Calzada Tecnológico
68030 Oaxaca city, México
Keywords: Knowledge acquisition, knowledge models, ontology merging, ontology fusion, knowledge representation.
Abstract: Knowledge scattered through the Web inside unstructured documents (text documents) can not be easily
interpreted by computers. To do so, knowledge contained from them must be extracted by a parser or a
person and poured into a suitable data structure, the best form to do this, are with ontologies. For an
appropriate merging of these “individual” ontologies, we consider repetitions, redundancies, synonyms,
meronyms, different level of details, different viewpoints of the concepts involved, and contradictions, a
large and useful ontology could be constructed. This paper presents OM algorithm, an automatic ontology
merger that achieves the fusion of two ontologies without human intervention. Through repeated application
of OM, we can get a growing ontology of a knowledge topic given. Using OM we hope to achieve
automatic knowledge acquisition. There are two missing tasks: the conversion of a given text to its
corresponding ontology (by a combination of syntactic and semantic analysis) is not yet automatically done;
and the exploitation of the large resulting ontology is still under development.
1 INTRODUCTION
These days computers are not anymore isolated
devices but they are important entry points in the
world-wide network that interchanges knowledge
and carries out business transactions. Nowadays,
using Internet to get data, information and
knowledge interchange is a business and an
academic need. Despite the facilities to access
Internet, people face the problem of heterogeneous
sources, because there are no suitable standards in
knowledge representation. This paper is designed for
this necesity of businesses and academia.
Many answers that people require involve
accessing several sources in the Internet, which are
later manually merged in a “reasonable” way.
Merging the information is an important task. Many
languages and tools (DAML+OIL (URL 15), RDF
(URL 16) and OWL (URL 14) have been developed
to describe and process Internet content but,
unfortunately, they don’t have enough
expressiveness to detail knowledge representation.
Given a document written in a natural language,
it is required that the computer deciphers the
information in it and converts it to a suitable
notation (its knowledge base) that preserves relevant
knowledge. This knowledge base can be an
ontology. To describe a knowledge domain, an
ontology represents the knowledge through nodes
that are joined through relations. Current works that
merge ontologies (Prompt (Noy, et al, 2000),
Chimaera (McGuinness, et al, 2000), OntoMerge
(Dou et al, 2002), FCA-Merge (Stumme et al,
2002) and If-Map (URL 1)) rely on the user to solve
the most important problems found in the process:
inconsistencies and adequate knowledge extraction.
In our fusion also these inconsistencies appear buy
they are solved by OM. OM, the merging algorithm
that we will explain, is totally automatic. This
algorithm solves by itself the inconsistencies found
in the process. In some cases OM applies the
confusion algorithm (Levachkine, S., and Guzman,
A. 2004) at the moment other solutions are studying.
Two important contributions herein presented to
obtain better advantage of the Web resources are:
A new notation to represent knowledge using
ontologies, called OM (Ontology Merging)
Notation, and
An automatic algorithm to merge ontologies,
called OM Algorithm.
This paper it is a summary of (Cuevas, 2006), on
it we describes the second contribution. The first
356
Guzman-Arenas A. and Cuevas-Rasgado A. (2008).
KNOWLEDGE ACQUISITION WITH OM - A Heuristic Solution.
In Proceedings of the Tenth International Conference on Enterprise Information Systems - AIDSS, pages 356-363
DOI: 10.5220/0001717703560363
Copyright
c
SciTePress
contribution, the OM notation it is in (Cuevas,
2006).
OM fuses two ontologies (this is our main
contribution), we are not doing any of: ontology
comparison, this has been done by COM (see below)
and others; ontology alignment, as Prompt (Noy &
Musen 2000); building a gigantic unique ontology,
this may or may not be done (see Discussion); an
ontology server, like Protegé (Noy & Musen 2000).
We neither pretend that our notation to be superior
to others (RDF (URL 16), say).
Some real examples appear where the texts
(unloaded by Internet) have became manually to
ontologies, OM have been applied and the result of
the fusion has been verified manually. We have a
work in the future; to make parser that turns of text
to ontologies and a deductor of intelligent questions
to verify the result of the fusion.
2 KNOWLEDGE ACQUISITION
The plan to follow is to acquire many “individual”
ontologies distilled from text documents, and then to
fuse them, two at the time, into a larger one. The
conversion of text into ontologies is hard, it is made
by a parser or syntactic analyzer, and will not be
covered in this work. This paper is focused to the
fusion of ontologies (arising from different sources)
amount computers. During this fusion the same
problems (redundancy, repetition, inconsistency…)
arise; the difference is that the machines have no
common sense (Lenat &Guha, 1989) and the
challenge is to make them to understand that
beneficial is the same to generous, and that triangle
represents: a three-sided polygon; a musical
percussion instrument; or a social situation
involving three parts. The computer solution to
fusion should be very close to people’s solution.
This paper explains a process of union of
ontologies in automatic and robust form. Automatic
because (unaided) computer detects and solves the
problems appearing during the union, and robust
because it performs the union in spite of different
organization (taxonomies) and when the sources are
jointly inconsistent.
The fusion is demonstrated by samples taking of
real Web documents and converting them by hand to
ontologies. These are then fed to the computer,
which produces (without human intervention) a third
ontology as result, like in (Kotis, 2006). This result
is hand-compared with the result obtained by a
person. Mistakes are below (section 3.3, Table 1).
2.1 Ontology
Formally, an ontology is a hypergraph (C, R) where
C is a set of concepts, some of which are relations;
and R is a set of restrictions of the form (r c
1
c
2
c
k
) between relation r and concepts c
1
to c
k
. It is said
that the arity of r is k. Check that relations are also
concepts.
An important task when dealing with several
ontologies is to identify most similar concepts. We
wrote COM (Olivares, 2002) that finds this
similarity throught ontologies.
2.2 The Role of Ontologies
An ontology is a data structure where information is
stored as nodes (representing concepts such as
hammer, printer, document, appearing in this
paper in Courier font) and relations
(representing restrictions among nodes, such as cuts,
transcribes or hair color, they appear in this paper in
Arial Narrow font, how in (hammer cuts wood),
(printer transcribes document). Usually, the
information is stored as “high level” and it is known
as knowledge.
Ontologies are useful when arbitrary relations
need to be represented, because it offers more
freedom to represent different types of concepts and
relations.
Currently notations to represent ontologies are
DAML+OIL (Connoly et al, 2001), RDF (URL 16)
and OWL (URL 14). These languages are a notable
accomplishment, but it does not have enough
features:
A relation can not be a concept. For instance, if
color is a relation, it is difficult to relate color to
other concepts (such as shape) by using other
relations.
Partitions (subsets with additional properties)
can not be represented (Gómez P. A., and
Suárez F. 2004).
2.3 Exploitation of Distributed Content
Works exist (McGuinness et al, 2000; Noy &
Musen, 2000 and Dou et al, 2002) that perform the
union of ontologies in a semiautomatic way
(requiring user’s assistance). Others (Kalfoglou &
Schorlemmer, 2002 and Stumme & Maedche, 2002)
require ontologies organized in a formal way, and to
be consistent with each other. In real life, ontologies
come from different sources are not likely to be
similarly organized, nor they are expected to be
KNOWLEDGE ACQUISITION WITH OM - A Heuristic Solution
357
mutually consistent. The automation of fusion needs
to solve these problems.
3 INTEGRATION DESIGN
This section explains the procedure that follows OM
as well as the cases in which it has been applied.
3.1 The OM Algorithm
This algorithm fuses two ontologies (Cuevas, 2006)
A and B into a third ontology C = A B containing
the information in A, plus the information in B not
contained in A, without repetitions (redundancies)
nor contradictions. OM proceeds as follows:
1. Ontology A is copied into C. Thus, initially, C
contains A.
2. Using the algorithm COM (Olivares, 2002) seek
in B each concept c
C
of C called the most similar
concept of C into B. The search starts from the root
concept of C, taking each one of its son of this until
to visit all the concepts of C. There are just two
options:
A. If c
C
has a most similar concept cms B,
then:
i. Relations that are synonyms (section 3.2,
example 2) are enriched.
ii. New relations (inluding Partitions) that
cms has in B, are added to c
C
. Concepts that
which are in the new relations which come
from cms are copied to C (if they are not).
iii. Inconsistencies (section 3.2) between the
relations of c
C
and those of cms are detected.
1. If it is possible, by using confusion
algorithm (Levachkine & Guzman 2007),
to resolve the inconsistency into c
C
.
2. When the inconsistency can not be
solved, OM rejects the contradicting
information in B, and c
C
keeps its original
relation from A.
B. If c
C
does not have a cms B, go to step 3.
3. It takes the next descendant of c
C
into C. Goes
back to step 2 until all the nodes of C are visited
(including the new nodes that are being added by
OM). (Cuevas, 2006) explains how this works.
3.2 Problems that OM Solves
In this section, figures show only relevant parts of
ontologies A, B and the resultant C, because they are
too large to fit.
Example 1: Merging Ontologies with Inconsistent
Knowledge. Differences between A and B could be
the following: different subjects, different names of
concepts or relations; repetitions; reference to the
same facts but with different words; different level
of details (precision, depth of description); different
perspectives (people are partitioned in A into male
and female, but in B they are young or old); and
contradictions. For example A (URL 12) contains:
The Novelist, poet and writer Don Miguel de
Cervantes was born in Alcalá de Henares, Madrid
while B contains: The writer, (URL 13) poet and
romantic Cervantes was born in Madrid, Spain. Both
ontologies duplicate some information (about
Cervante’s place of birth), different expressions
(novelist, poet and writer versus writer, poet and
romantic), different level of details (Don Miguel de
Cervantes versus Cervantes), and contradictions
(Alcalá de Henares, Madrid vs. Madrid, Spain). A
person will have in her mind a consistent
combination of information: Cervantes and Don
Miguel de Cervantes are not the same person, or
perhaps they are the same, they are synonyms. If she
knows them, she may deduce that Don Miguel de
Cervantes is the complete name of Cervantes. We
solve these problems everyday, using previously
acquired knowledge and common sense knowledge
(Lenat & Guha, 1989), which computers lack. Also,
they did not have a gradual and automatical way to
grow their ontology. OM measures the inconsistency
(of two apparently contradicting facts) by asking
conf (Levachkine & Guzman 2007) to determine the
size of the confusion in using Alcalá de Henares in
instead of Madrid and vice versa, or the confusion of
using Don Miguel de Cervantes instead of
Cervantes.
OM does not accept two different concepts for a
birthplace. If A said that Don Miguel de Cervantes
was born in Alcalá de Henares and B says that
Cervantes was born in Madrid, OM chooses Alcalá
de Henares instead of Madrid because it is more
specific place while Madrid that is more general (it
deduces this from a hierarchy of places in Europe).
Small inconsistencies cause C to retain the most
specific value, while if it is large, OM keeps C
unchanged (ignoring the contradicting fact from B).
In case of inconsistency, A prevails. This is also
because we can consider that an agent’s previous
knowledge is A, and that such agent is trying to
learn ontology B. In case of inconsistency, it is
natural for the agent to trust more its previous
knowledge, and to disregard inconsistent knowledge
in B as “not trustworthy” and therefore not acquired
ICEIS 2008 - International Conference on Enterprise Information Systems
358
– the agent refuses to learn knowledge that it finds
inconsistent, if the inconsistency is too large.
Example 2: Joining Partitions, Synonym
Identification, Organization of Subset to
Partition, Identification of Similar Concepts,
Elimination of Redundant Relations and
Addition of New Concepts. This example is
accomplished through eight cases:
a. Relation that are not Copied, if there are
Contradictions. Figure 1 shows relation
neutron without charge, that mean:
neutron does not have charge. Another
ontology containing: neutron with
positive charge that contradict what it is
before (see Figure 1), thus the contradicting
relation will not be copied into the fused result.
Figure 1: The concept neutron has a relation without
charge that OM recognizes as absence of that property in
neutron. Another ontology having relation has
charge, has negative charge linked to neutron,
will be an inconsistent information. When fusing both
ontologies, OM will not copy neutron has negative
charge into the result.
b. Copying New Partitions. building is a partition in
A (indicated in the small circle ) of Monte
Albán, therefore it is added to the resulting
ontology C (Figure 2).
c. Copying new concepts. Concepts Mixtec and
Mexican Republic were not in A, but they
appear in B. Therefore, they were copied by
OM to C (Figure 2).
d. Reorganization of Relations. Relation located in
appears twice but with different values,
therefore they are added to C because it is
possible that the relation to have several values
(Figure 2). In case of single-valued relations,
confusion algorithm (Levachkine & Guzman,
2007) is used.
e. Synonym Identification. Relation built by in A
(Figure 2) and made by in B are both
synonymous because in the definition of make
by in B (the words that defines it, between
parenthesis) we found the word build. OM fuses
in C the relation built by of A, with both
descriptive phrases build and make by.
f. Identification of Similar Concepts. In the
Figure 3, concept sculpture of a
jaguar in A and throne in the shape
of jaguar in B have the same properties
(Color and its value) therefore, OM fuses them
into a single concept. The same happens with
El Castillo and Pyramid of
Kukulkan because they have the same
properties and children.
g. Removing Redundant Relations. In A,
Chichen Itza is member of pre-
Columbian archaeological site
(Figure 3), which is a member of
archaeological sites. In B, Chichen
Itza is member of archaeological site
(which is parent of pre-Colombian
archaeological site in B), therefore it
is eliminated in C because it is a redundant
relation. In C, pre-Columbian
archaeological site i
s
parent of
Chichen Itza. The same occurs with
Isotope subset of chemical element in
figure 5, where A shows that Isotope is a
subset of chemical element and B shows
that Isotope is a subset of atom. OM check
that Isotope is a subset of atom that is at the
same time a subset of chemical element
and Isotope is a subset of chemical
element (OM erases this last relation because
it is redundant). But not in the Figure 4 where
the relation isotope subset of chemical
element is different to isotope member of
atom and atom subset of chemical
element.
h. Organization of Subset to Partition. In the
building partition in A there are six subsets
(Figure 3): Ballcourt, Palace, Stage,
Market and Bath. OM identifies them in B,
where they appear as subsets of Chichen
Itza. OM copies then into C like a partition
and not as simple subsets. OM prefers the
partition to just subset.
KNOWLEDGE ACQUISITION WITH OM - A Heuristic Solution
359
Figure 2: Ontology A describes Monte Alban. From a different view point, ontology B does the same. Ontology C is the
result of OM fusing them. The relation built by in ontology A and make by in B are identified (case e) as synonyms, hence it
is enriched during the fusion (into C). Only relevant parts of A, B and C are shown.
Figure 3: Ontology A and B describe Chichen Itza, where concepts Chac Mool in ontology A and Chac in B are
identified (case e) as synonyms. A more interesting case is case e, that identifies sculpture of a jaguar in A as a
similar concept (a synonym) to throne in the shape of jaguar in B. Also The Castle in A and Pyramid
of Kukulkan in B are found to be the same. Case f removes redundant relations (marked with an X in the result C). Case
g (see text) upgrades a set of subsets into a partition.
ICEIS 2008 - International Conference on Enterprise Information Systems
360
Figure 4: In ontology A concept isotope is a subset of
Chemical element, but in B isotope is a member of
atom; into C, the resultant ontology OM provides
isotope with both ancestors Chemical element
and atom.
Figure 5: In B, isotope is subset of atom, whereas in A
isotope is subset of chemical element. This last
relation is redundant in C, and OM does not add it to C.
OM fuses carefully, eliminating redundant relations.
3.3 More Applications of OM in Real
Cases taken from the Web
OM has merged ontologies derived from real
documents. The ontologies were obtained manually
from several documents: 100 Years of Loneliness
(URL 9 and 11], Oaxaca (URL 5 and 10), poppy
(URL 2 and 4) and turtles (URL 6 and 7), describing
the same thing. These ontologies were merged
(automatically) by OM, and the product was
manually validated, obtaining good results (Table 1).
Table 1: Performance of OM in some real examples: The
columns “error” give the ratio of (number of wrong
relations) / (total number of relations) and (number of
wrong concepts) / (total number of concepts), respectively.
More real examples in (Cuevas, 2006).
Ontologies Error in the
merging of
relations
Error in the
merging of
concepts
Turtles 0 0
Hammer 0 0
Poppy 0 0
100 Years of
Loneliness
2.7% 5.3%
Oaxaca 0 0.3%
4 DISCUSSION
Is it possible to keep fusing of several ontologies
about the same topic, in order to have a larger
ontology that faithfully represents and join the
knowledge in each of the formant ontologies? OM
say “yes, it is possible.” What are the main
roadblocks? As we perceive them, they are:
a. Exploitation of hypergraphs. Although we
define ontologies as hypergraphs (section 2.1),
the restrictions (r c
1
c
2
c
k
), where r is a
relation, are lists, and consequently, order
matters. For instance, it is not the same (kills;
Cain; Abel; jaw of donkey) that (kills;
Abel; Cain; jaw of donkey). However,
the role of each “argument” or element of the
restriction (such as jaw of donkey) must be
explained –in the example it is the instrument
used in the killing. Restrictions have different
number of arguments, each one with different
roles: consider (born; Abraham Lincoln;
Kentucky; 1809; log cabin). We can
expect a lot of arguments in a fragment of the
text. The role of each argument must be
explained or described in a transparent (not
confuse) form –ideally, we suggest OM notation
explain in (Cuevas, 2006)-, where OM can
understand such explanations, manipulate them
and create new ones. For instance, from a given
argument, it should be able to take two different
explanations (coming from ontologies A and B,
respectively) and fuse them into a third
explanation about such argument, to join into C.
Ways to do all of this should be devised.
b. A good parser. Documents are now transformed
by hand into ontologies, although fusion is
totally automatic, but the work of verify the
fusion is hard because it is also by hand. It has
been found difficult to build a parser that
reliably transforms a natural language document
KNOWLEDGE ACQUISITION WITH OM - A Heuristic Solution
361
into its suitable ontology, due to the ambiguity
of natural language and to the difficulty of
representing relations (verbs, actions, processes)
in a transparent way (see next point). Probably a
good parser will profit from the current
knowledge that OM has stored in the ontology
that was built before, as well as in additional
knowledge sources (point c below).
c. Additional language-dependent knowledge
sources could effort enhance OM. For instance,
WordNet, WordMenu, automatic discovery of
ontologies by analyzing titles of conferences,
university departments (Makagonov, 2007).
d. A query-answerer that queries a large ontology
and makes deductions. (Botello, 2007) works on
this for databases, not for ontologies. He has
obtained no results for real data, yet.
In addition, some caveats are:
e. OM does not have a way to know what is true
and what is false. All it does is to compute
ontology C as the fusion of A and B, in a
consistent form. If A and B say the same lies,
these will go into C.
f. Probably the first ontologies should be carefully
done by hand (even if parser existed), like, first
documents (their ontologies, that is) to be fed to
OM (“the first things OM will learn”) have to
be consistent, clear, and at a “low level.”
g. The formal support behind OM and OM
notation should be clearly adhered to.
5 CONCLUSIONS AND FUTURE
WORK
This paper presents an automatic procedure (the OM
algorithm) to fuse two ontologies about the same
topic, which produces good results.. Thus, it is an
important improvement to the computer-aided
merging editors currently available (section 2.3).
OM is an automatic, robust algorithm that fuses two
ontologies into a third one, which preserves the
knowledge obtained from the sources, solving some
inconsistencies, detecting synonyms and homonyms,
and expunging some redundant relations.
The examples shown, as well as others in
(Cuevas, 2006; Cuevas & Guzman, 2007), provide
evidence that OM does a good job, in spite of very
general or very specific of joining ontologies. This is
because the algorithm takes into account not only
the words in the definition of each concept, but its
semantics [context, synonyms, resemblance (through
conf) to other concepts…] too. In addition, its base
knowledge (some pre-built knowledge, such as
synonyms, external language sources, stop words,
words that change the meaning of a relation, among
others) helps.
OM has not been tried on extensive, “real”
ontologies (for instance, an ontology describing the
complete work “100 Years of Loneliness”), due to
the tedious work to hand-craft such ontology from
the written document. Section 5.b addresses this.
REFERENCES
Botello, A. 2007. Query resolution in heterogeneous data
bases by partial integration. Ph. D. Thesis in progress,
Centro de Investigación en Computación (CIC),
Instituto Politécnico Nacional (IPN), Mexico. In
Spanish.
Cuevas, A. 2006 Merging of ontologies using semantic
properties. Ph. D. thesis. CIC-IPN. In Spanish.
http://148.204.20.100:8080/bibliodigital/ShowObject.j
sp?idobject=34274&idrepositorio=2&type=recipiente
Cuevas, A. and Guzman, A. 2007. A language and
algorithm for automatic merging of ontologies.
Chapter of the book Handbook of Ontologies for
Business Interaction, Peter Rittgen, ed. Idea Group
Inc. Pp. 381-404.
Dou D., McDermott, D., and Qi. P. 2002. Ontology
Translation by Ontology Merging and Automated
Reasoning. Proc. EKAW Workshop on Ontologies for
Multi-Agent Systems, Spain. Pp. 3-18.
Gómez P. A., and Suárez F. M. 2004. Evaluation of
RDF(S) and DAML+OIL Import/Export Services
within Ontology Platforms. MICAI 2004. Pp. 109-
118.
Kalfoglou, Y., and Schorlemmer, M. Information-Flow-
based Ontology Mapping. 2002. Proceedings of the
International Conference on Ontologies, Databases
and Application of Semantics for Large Scale
Information Systems, volume 2519 of Lecture Notes in
Computer Science. Pp. 1132-1151.
Kotis K., Vouros G. and Stergiou, K. 2006. Towards
Automatic of Domain Ontologies: The HCONE-merge
approach. Journal of Web Semantics (JWS), Elsevier,
vol. 4:1, Pp. 60-79.
Lenat, D., and Guha, V. Building Large Knowledge-Based
Systems. (Addison-Wesley 1989).
Levachkine, S., and Guzman, A. 2004 Hierarchy
Measuring Qualitative Variables. Lecture Notes in
Computer Science LNCS 2945. Computational
Linguistics and Intelligent Text Processing, Springer-
Verlag. Pp. 262-274. ISSN 0372-9743.
Makagonov, P. 2007. Automatic formation of ontologies
by analysis of titles of conferences, sessions and
articles.Work in preparation.
McGuinness, D., Fikes, R., Rice, J., and Wilder, S. 2000.
The Chimaera Ontology Environment. Proceedings of
the Seventeenth National Conference on Artificial
ICEIS 2008 - International Conference on Enterprise Information Systems
362
Intelligence and Twelfth Conference on Innovative
Applications of Artificial Intelligence. ISBN: 0-262-
51112-6. Pp. 1123-1124.
Noy, N., and Musen, M. 2000. PROMPT: Algoritm and
Tool for Automated Ontology Merging and
Alignment. In Proc. Of the National Conference on
Artificial Intelligence. Pp. 450-455, Austin, TX, USA.
Olivares, J. 2002. An Interaction Model among Purposeful
Agents, Mixed Ontologies and Unexpected Events. Ph.
D. Thesis, CIC-IPN. México.
www.jesusolivares.com/interaction/publica.
Stumme, G., and Maedche, A. 2002. Ontology Merging
for Federated ontologies on the semantic web. In: E.
Franconi, K. Barker, D. Calvanese (Eds.): Proc. Intl.
Workshop on Foundations of Models for Information
Integration, Viterbo, Italy. LNAI, Springer 2002.
URLs:
1. Accessed 5 march 2008. AKT:
plainmoor.open.ac.uk/ocml/domains/aktive-portal-
ontology/techs.html
2. Accessed 5 march 2008.
es.wikipedia.org/wiki/Amapola
3. Accessed 5 march 2008.
es.wikipedia.org/wiki/Miguel_%C3%81ngel
4. Accessed 5 march 2008.
www.buscajalisco.com/bj/salud/herbolaria.php?id=1
5. Accessed 5 march 2008.
www.elbalero.gob.mx/explora/html/oaxaca/geogra
fia.html
6. Accessed 5 march 2008.
www.damisela.com/zoo/rep/tortugas/index.htm
7. Accessed 5 march 2008.
www.foyel.com/cartillas/37/tortugas_-
_accesorios_para_acuarios_i.html
8. Accessed 5 march 2008.
www.historiadelartemgm.com.ar/biografiamichelangel
obuonarroti.htm
9. Accessed 5 march 2008.
www. monografias.com/trabajos10/ciso/ciso.shtml
10. Accessed 5 march 2008.
www.oaxaca-mio.com/atrac_turisticos/infooaxaca.htm
11. Accessed 5 march 2008.
www.rincondelvago.com/cien-anos-de-
soledad_gabriel-garcia- marquez_22.html
12. Accessed 5 march 2008.
es.wikipedia.org/wiki/Miguel_de_Cervantes
13. Accessed 5 march 2008.
www.biografiasyvidas.com/monografia/cervantes/
14. Accessed 5 march 2008.
http://www.w3.org/TR/2004/REC-owl-ref-20040210/
WWW page
15. Accessed 6 march 2008.
http://www.w3.org/TR/2001/NOTE-daml+oil-
reference-20011218. WWW page
16. Accessed 6 march 2008.
http://www.w3.org/TR/2004/REC-rdf-primer-
20040210/. WWW page.
KNOWLEDGE ACQUISITION WITH OM - A Heuristic Solution
363