Professional Documents
Culture Documents
by
Alexander Maedche
University of Karlsruhe, Germany
Maedche, Alexander D.
Ontology learning for the semantic Web / by Alexander D. Maedche.
p. cm.
Includes bibliographical references and index.
ISBN 978-1-4613-5307-2 ISBN 978-1-4615-0925-7 (eBook)
DOI 10.1007/978-1-4615-0925-7
1. Web site development. 2. Metadata. 3. Ontology. 4. Artificial intelligence. 1. Title.
List of Figures ix
List of Tables xiii
Preface xv
Acknowledgements xviii
Foreword by R. Studer XIX
Part I Fundamentals
1. INTRODUCTION 3
1 Motivation & Problem Description 3
2 Research Questions 4
3 Reader's Guide 6
3 Conclusion 144
3.1 Multi-Strategy Learning 145
3.2 Taxonomic vs. Non-Taxonomic Relations 145
3.3 A Note on Learning Axioms - AO 146
The web in its' current form is an impressive success with a growing number
of users and information sources. However, the growing complexity of the web
is not reflected in the current state of Web technology. The heavy burden of
accessing, extracting, interpretating and maintaining is left to the human user.
Tim Bemers-Lee, the inventor of the WWW, coined the vision of a Semantic
Web in which background knowledge on the meaning Web resources is stored
through the use of machine-processable (meta-)data. The Semantic Web should
bring structure to the content of Web pages, being an extension of the current
Web, in which information is given a well-defined meaning. Thus, the Semantic
Web will be able to support automated services based on these descriptions of
semantics. These descriptions are seen as a key factor to finding a way out of
the growing problems of traversing the expanding web space, where most web
resources can currently only be found through syntactic matches (e.g., keyword
search).
Ontologies have shown to be the right answer to these structuring and mod-
eling problems by providing a formal conceptualization of a particular domain
that is shared by a group of people. Thus, in the context of the Semantic Web,
ontologies describe domain theories for the explicit representation of the seman-
tics of the data. The Semantic Web relies heavily on these formal ontologies that
structure underlying data enabling comprehensive and transportable machine
understanding. Though ontology engineering tools have matured over the last
decade, the manual building of ontologies still remains a tedious, cumbersome
task which can easily result in a knowledge acquisition bottleneck. The suc-
cess of the Semantic Web strongly depends on the proliferation of ontologies,
which requires that the engineering of ontologies be completed quickly and
easily. When using ontologies as a basis for Semantic Web applications, one
has to face exactly this issue and in particular questions about development
time, difficulty, confidence and the maintenance of ontologies. Thus, what one
ends up with is similar to what knowledge engineers have dealt with over the
xvi ONTOLOGY LEARNING FOR THE SEMANTIC WEB
aspect of applying ontology learning techniques deals with the question of how
to measure the quality of the application of these techniques. Therefore, the
third part of this book introduces a new approach and measures for evaluating
ontology learning based on the well-known idea of having gold standards as
evaluation references. The fourth part of this book provides a detailed overview
of existing work that emphasizes topics of interest with similarities to the task
of ontology learning. It analyzes a multitude of disciplines (ranging from in-
formation retrieval, information extraction and machine learning to databases).
The book concludes with a summary of contributions and insights gained. Fi-
nally, a vision of the future and a discussion of future challenges in regards to
the Semantic Web is delineated.
ALEXANDER MAEDCHE
Acknowledgements
Writing a book is a complex project in that many people are involved. I thank
all people supporting me in my research and especially in writing this book. I
appreciate very much the important roles that my colleagues Michael Erdmann,
Siegfried Handschuh, Andreas Hotho, Gerd Stumme, Nenad Stojanovic, Ljil-
jana Stojanovic, York Sure, and Raphael Volz played. I thank all my students
that supported me in my work by doing implementation and evaluation work.
Very special thanks to Raphael Volz, now one of my colleagues, who did heavy
implementation work in his master thesis. Stefan Decker, the Semantic Web
initiator at our research group in Karlsruhe, always and at any time was open
for useful comments. Special thanks to Steffen Staab for giving me the first
ideas on Ontology Learning for the Semantic Web. He always was open for
crazy discussions producing new ideas. I thank Rudi Studer, my advisor and
leader of the research group. He supported me in making great experiences
during my time at Karlsruhe. His way of leading me and the overall research
group created a prolific research environment. Thanks to Jdrg-Uwe Kietz that
provided useful input and comments to my work on ontology learning. Without
all of them them, this work would not have been possible.
I thank may parents that financed and supported my long stay at the university.
Mostly, however, I must thank my friend and wife, Ellen, who always accepted
when I was saying that there will come better times with less work. Thank to
all of you for being there.
Alexander Maedche
Karlsruhe, Germany
Foreword
The success of the Web today can be explained to a large extent by its sim-
plicity, i.e. the low level technical know-how that is needed to put information
into the Web and to access Web information by browsing and keyword-based
search. However, the volume of information that is nowadays available on the
Web makes the limits of the current Web drastically obvious for its users: find-
ing relevant information among millions of Web pages becomes more and more
a heavy burden, and more than once it becomes impossible.
The development of the Semantic Web is a promising path towards trans-
forming the Web into a semantically grounded information space that makes
information accessible in a semantic way. It is a common understanding that
machine-processable metadata that come with a semantic foundation as pro-
vided by ontologies, establish the technological basis for such a semantic pro-
cessing of Web information.
All experience in practical settings shows that the engineering of ontologies
is a crucial bottleneck when setting up Semantic Web applications. Further-
more, in fast changing market environments outdated ontologies mean outdated
applications. As a consequence, the systematic management of the evolution
of ontologies is a bottleneck as well.
Rather recently, these challenges gave rise to a new research area: "Ontol-
ogy Learning". Ontology Learning aims at developing methods and tools that
reduce the manual effort for engineering and managing ontologies. Ontology
Learning is an inherently interdisciplinary area bringing together methods from
ontology engineering, knowledge representation, machine learning, computa-
tionallinguistics and information extraction. Nowadays, there is no chance to
fully automate these learning processes. Therefore all approaches assume some
cooperation between humans and machines, i.e. they provide semi-automatic
means for ontology engineering and evolution.
This book describes a comprehensive framework for Ontology Learning.
This framework addresses for the first time the specific aspects of Ontology
xxii ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Learning that arise in the context of the Semantic Web, e.g. the heterogeneity
of the Web sources and the layered representation of Web-based ontologies.
Ontology Learning relies on a tight integration of shallow linguistic process-
ing with ontology representation. Therefore, the Ontology Learning framework
defines a new notion of ontology that establishes precisely defined links between
a linguistic layer, an ontology, and an associated knowledge base that populates
the ontology. This integration paves the way for transforming lexical entries
and linguistic asssociations into conceptual entries of the ontology and related
conceptual relations.
The framework exploits a process-oriented view for Ontology Learning that
distinguishes between the phases Import, Extract, Prune, and Refine. Thus,
Ontology Learning is decomposed into subtasks that address specific aspects
and can therefore solved with methods that are tailored to these subtask-specific
challenges. Given the heterogeneity of the sources that are available in the Web
context as well as the diversity of the different ontology learning tasks it is obvi-
ous that no single learning approach can meet all these different requirements.
Therefore, the framework defines a system architecture that supports multi-
strategy learning, i.e. the results of different learning methods are combined
in order to achieve sufficiently good learning results. Thus, the framework is
open for adding new learning algorithms that may improve the learning results.
The description of the framework elaborates different learning subtasks, espe-
cially the import of ontologies (including ontology integration), the extraction
of ontologies from semi-structured sources, the learning of non-taxonomic rela-
tions, and the pruning of ontologies. As such, a broad collection of techniques
is integrated into the Ontology Learning framework. A considerable part of
the framework have been implemented in the ontology engineering framework
OntoEdit and the learning environment Text-To-Onto.
When learning ontologies an immediate question arises: what is the qual-
ity of the learning results. This is a rather tough problem since there do not
exist obvious quality standards. The ontology learning framework addresses
this problem by introducing a collection of measures for comparing ontologies
to each other. First evaluations indicate that the manual engineering and the
learning of ontologies supplement each other in a nice way and thus open the
way for further elaborating of how to arrange the cooperation between human
and machine for ontology learning.
The ontology learning framework as described in this book is a promis-
ing step in further developing the field of ontology learning. By identifying
clearly defined subtasks, further learning methods may be developed that en-
hance the learning results for respective subtasks. The framework is part of
the development and implementation of the Karlsruhe Ontology and Semantic
Web infrastructure that provides an overall architecture for managing and ap-
plying ontologies in the context of the Semantic Web. Thus ontology learning
xxiii
is tightly integrated with other aspects of the Semantic Web, like e.g semi-
automatic generation of metadata, the alignment of ontologies or inferring new
facts from given metadata and ontologies.
Ontology learning is a rather young, yet very promising research field. The
transfer of its research results into scalable products will be an important step
towards making the Semantic Web happen.
FUNDAMENTALS
Chapter 1
INTRODUCTION
Semantic Web - a web of data that can be processed directly or indirectly by machines.
-(Bemers-Lee, 1999)
of ontologies still remains a tedious, cumbersome task which can easily result in
a knowledge acquisition bottleneck. The success of the Semantic Web strongly
depends on the proliferation of ontologies, which requires that the engineer-
ing of ontologies be completed quickly and easily. When using ontologies as
a basis for Semantic Web applications, one has to face exactly this issue and
in particular questions about development time, difficulty, confidence and the
maintenance of ontologies. Thus, what one ends up with is similar to what
knowledge engineers have dealt with over the last two decades when elabo-
rating methodologies for knowledge acquisition or workbenches for defining
knowledge bases. A method which has proven to be extremely beneficial for
the knowledge acquisition task is the integration of knowledge acquisition with
machine learning techniques (see the seminal work described in (Skuce et aI.,
1985; Reimer, 1990; Szpakowicz, 1990; Buntime and Stirling, 1991; Morik
et aI., 1993a; Nedellec and Causse, 1992; Webb, 1996».
2. Research Questions
The extraction and maintenance of ontologies for the Semantic Web opens
a bundle of research questions. In the following the most important ones are
sketched that will be approached in the work described in this book. It has been
mentioned that the idea of using machine learning for knowledge acquisition
is not a new one. However, the question of how to combine machine learning
and knowledge acquisition is still unsolved and difficult to answer in general.
Thus, the following question has to be asked:
Introduction 5
• How to support cooperation between the ontology engineer and the machine
learning algorithm? How can the complexity of the ontology engineering
task be reduced by using machine learning support?
Thus, the question of how to integrate the ontology engineer in the process of
applying ontology learning to achieve a tight integration between human beings
and machines (or the specific algorithm in general) is investigated and possible
solutions are described in this book.
• How to apply existing algorithms for ontology learning? Where are adap-
tations required? (e.g. for the usage of background knowledge)
Evaluating ontology engineering or more general knowledge engineering is
not this well researched. It lacks generic methods and concrete evaluation and
comparison measures (compared to recalllprecision in information retrieval)
that may be applied for evaluating ontology learning:
• How can the ontology learning results be evaluated? How can two given
ontologies be formally compared?
3. Reader's Guide
Every chapter is preceded with a brief introductory paragraph which explains
how the work presented in the section fits in the overall structure of the book 2 .
The book is divided into four main parts, Fundamentals (I), Ontology Learning
for the Semantic Web (II), Implementation & Evaluation (III) and Related Work
& Outlook (IV). These four main parts are organized as follows:
Part I - Fundamentals.
Introduction
I Fundamentals
Ontology - Overview
'---y-------' and Definition
evaluation results are compared with the human modeling results indicating
that ontology learning compares quite well to human modeling, in the sense
that they complement each other.
• Chapter 9 deals with the difficult task of giving an overview of related work
on ontology learning. Until now ontology learning has not existed. How-
ever, much work in a number of disciplines, like computational linguistics,
information retrieval, machine learning, databases, software engineering has
researched and applied techniques for solving part of the overall problem of
ontology learning for the Semantic Web. This chapter gives an overview of
related work from a number of different communities.
main contributions of the work described in this book and lists a number of
insights gained doing this research. Additionally, unsolved questions and
further research issues are defined.
Figure 1.1 gives a graphical overview on how to read the book. Readers that
are acquainted with ontology engineering for the Semantic Web may skip chap-
ter 3 and directly go to chapter 4 that introduces the overall ontology learning
framework. Readers only interested in the implemented tool environment for
ontology learning may skip the technical chapters 5 and 6 on data import & pro-
cessing and ontology learning algorithms. Chapter 8 contains a detailed study
on ontology evaluation and comparison. Readers interested in related work and
getting an overview on ontology learning should read chapter 9. Chapter 10
concludes and lists issues that have been left open in this book.
10 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Notes
1 The reader may note that "Ontology Learning" has become a hot research
topic in the last two years. The relevancy for research on ontology learn-
ing is obvious. Two successful ontology learning workshops (Staab et aI.,
2000c; Maedche et aI., 200lb) have been organized in 2000 and 2001. Ap-
proximately 10 % of accepted papers at classical knowledge acquisition
conferences (such as EKAW and K-CAP) deal with ontology learning.
2 Relevant publications of the author with respect to the specific chapter are
cited in the introductory paragraph.
Chapter 2
This chapter starts with a motivation why computer science requires the
concept of ontology. Subsequently, the basic origin of ontology in philosophy
and and its relation to the meaning triangle known from semiotics is intro-
duced. Based on this foundations it will be shown what an agreement by an
ontology implies for human-human communication, human-machine commu-
nication and machine-machine communication. The underlying idea of the
meaning triangle will be adopted and related to the current view on ontologies
and their application in computer science leading to a layered ontology struc-
ture. A formal introduction of the layered ontology structure describing its core
elements and their interaction will be given in Definition 2.1. In addition to the
layered ontology structure a knowledge base structure that may be instantiated
on top of the ontology will be defined. These definitions will be used throughout
the book.
Subsequently, we will shortly sketch different classification schemes for on-
tologies. Thus, one has to stress, that for ontologies being cost-effectively
deployed, one requires a clear understanding of the various ways ontologies are
being used and developed (Jasper and Uschold, 1999). Finally before we con-
clude this chapter, several application areas and applications that heavily build
on the conceptual background knowledge defined through domain-specific on-
tologies will be introduced.
The need for ontologies in computer science. In general humans use their
language to communicate and to create models of the world. Naturallanguages
are not suitable for building models in computer science, because they are too
ambigouous. Therefore, so calledformallanguages are used to specify models
of the world. One may consider mathematics as such a formal language. Frege
(1848-1925) researched the formal foundations of mathematics as a formal
language. In his work he tried to separate between language and the propositions
that are applied for human reasoning. In his Begriffsschrift (Frege, 1922) he
described one of the best known formal language, namely first order logic
(FOL). Citing Wittgenstein, it can be said that
In practice, language is always more or less vague, so what we assert is never quite precise.
Thus, logic has two problems to deal with in regard to symbolism: (I) the conditions
for sense rather than nonsense in combination of symbols and (2) the conditions for
uniqueness of meaning or reference in symbols or combinations of symbols." A logically
perfect language has rules of syntax which prevent nonsense, and has single symbols
which always have a definite and unique meaning.
-(Wittgenstein, 1922)
Looking at the citation above one can see that the first point is well researched
in computer science: Formal languages are formulated as term substitution
systems based on the work of Thue and Chomsky (Chomsky, 1965). A given
Ontology - Definition & Overview 13
finite set of signs (alphabet) and a finite set of production rules produces an
infite set of expressions or sentences that define the language. The second
point described in Wittgenstein's citation is not as well researched. Producing
a syntactic correct language does not mean that one has captured the meaning
and sense of the sentences of a given language.
In the following the claim that ontologies are means to bridge the "semantic
gap" existing between the actual syntactic representation of information and
its conceptualization is made. Sharing or reusing knowledge across systems
becomes difficult, as different systems use different terms for describing infor-
mation. What an ontology is from its origin point of view will be introduced.
Based on this introduction the original view with its current use in computer
science will be combined. This combination will lead to a formal definition of
an ontology that will be the basis for the overall book.
Ontology - Its origin. It has been already introduced that research in "On-
tology" has its origin in philosophy. It is a philosopical discipline, a branch
of philosophy that deals with the nature and the organisation of being. How
thoughts, words and things relate to one another has been a recurrent topic in
philosophy and language as early as Plato to the modem era (Campbell et aI.,
1998). Plato dealt with the question of the proper naming of things. In his opin-
ion, the use of names in an "optimal world" would be to ensure that a particular
expression will make everybody think of one and only one thing. However,
he was doubtful that perfect names could ever be given, because things are
continually changing.
Aristotles' work went beyond the question of names and was interested in
definitions. His notion of definition was not simply the meaning of a word. A
definition was meant to clearly explain what a thing is by being a statement of
the "essence" of the entity. Therefore, he believed that to say what something
is always requires to say why something is. An Aristotelian definition is given
by specifying the genus and differentia of individuals, and then using logical
arguments to categorize those individuals based upon their definitions. By
identifying common definitional properties of similar individuals, the definition
explains why they are members of the same kind.
Aristotle's foundation was not sufficient: He ignored the unavoidable lim-
itations of communicating meaning via language and the ambiguities created
by implicit exchange of different "senses" of meaning. Understanding such
ambiguities was a topic considered by Gottlob Frege (1848-1925). Frege intro-
duced a distinction of two types of meaning: the concept and the referent. The
graphical interpretation of this distinction is commonly referred to as the mean-
ing triangle (cf. Figure 2.1) and has been introduced by (Ogden and Richards,
1923). The meaning triangle defines the interaction between symbols or words,
thoughts and things of the real world.
14 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Symbols I
Syntactic structures
Thoughl$1
Semantic structures
: a speCifIC
; demaIn, e,9, Things ;n the
: animais real world
Figure 2.2 depicts the overall setting for communication between human and
machine agents. Three layers are distinguished:
• First, one can consider things that exist in the real world, including in this
example human and machine agents, cars, and animals.
• Secondly, one can consider symbols and syntactic structures that are ex-
changed.
16 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
• Thirdly, models with their specific associated thoughts and semantic struc-
tures are analyzed.
First consider the left side of Figure 2.2 without assuming a commitment to
a given ontology. Two human agents HAl and HA2 exchange a specific sign
(e.g. a word like "Jaguar"). Given their own internal model each of them will
associate the sign to their own concept (or thought) referring to possibly two
completely different existing things in the world, e.g. the animal vs. the car.
The same holds for machine agents: They may exchange statements based on a
common syntax, however, they may have different formal models with differing
interpretations.
Consider the scenario that both human agents commit to a specific ontology
that deals with a specific domain (e.g. animals). The chance they both refer
to the same thing in the world increases considerably. The same is true for
the machine agents MAl and MA2: They have actual knowledge and they
use the ontology to have a common semantic basis. When agent MAl uses
the term "Jaguar", the other agent MA2 may use the ontology just mentioned
as background knowledge and rule out incorrect references, e.g. ones that let
"Jaguar" stand for the car. Human and machine agents use their concepts and
their inference processes, respectively, in order to narrow down the choice of
referents (e.g., because animals do not have wheels, but cars have).
Subsequently, our notion of ontology is defined. However, in contrast to
most other research about ontology languages it is not the purpose to invent a
new logical language or to redescribe an old one. Rather it is a way of modeling
and structuring the elements contained in an ontology that inherently considers
the special role of signs (mostly strings in current ontology-based systems)
and references. Especially for this work signs playa major role for ontology
learning for the Semantic Web, because existing, available web data as starting
point for ontology learning is considered as written signs.
An important aspect is that there exists the conflict that ontologies are for
human and machine agents, but logical theories are mostly for mathematicians
and inference engines. Formal semantics for ontologies is a sine qua non.
However, in addition to the benefits of a logical rigor, user and developer of an
ontology-based system profit from ontology structures that explain the possible
misunderstandings and compromise a direct mapping to natural language.
For instance, one might specify the sign "Jaguar" refers to the union of the
set of all animals that are jaguars and the set of all cars that are jaguars. Alterna-
tively, one may describe that "Jaguar" is a sign that may either refer to a concept
ANIMAL-JAGUAR or to a concept CAR-J AGUAR. The second way is preferred.
In conjunction with appropriate GUI modules one may avoid presentations
of 'funny symbols' to the user like "animal-jaguar", while avoiding 'funny
inference' that may arise from artificial concepts like (ANIMAL-JAGUAR U
CAR-JAGUAR).
Ontology - Definition & Overview 17
The Ontology Structure O. Within the scope of the research described in this
book a definition of an ontology structure, describing its core "elements", their
interaction and a mapping to formal semantics of these elements is required.
Starting the research several existing definitions such as the widely accepted
OKBC model 1(see (Chaudhri et aI., 1998)) or the ISO standard (see (ISO 704,
1987)) of principles and methods of terminology 2 were analyzed.
The existing definitions did not fulfill the requirements of a comprehensive
ontology structure for the work described here. Based on the human point
of view that meaning is established through communication, e.g. through ex-
changing words or more general signs, existing definitions have been extended3 .
The reader may note that in the approach of ontology learning it is dealt with in
this work one considers the "discovery of semantics" that is implicitly contained
in existing data that has been generated by humans through exchanging signs.
Therefore, the focus on the interaction between natural language and formal se-
mantics is formed. The ontology structure 0 introduced here extends existing
ontology definitions such as OKBC and ISO-704 with an "explicit sign level"
and introduces a new model of ontologies. Looking into existing literature the
research area of semiotics, as the study of signs and the ways in which sign
systems convey (and are used to convey) meaning (Eco, 1981; Euzenat, 2000),
has to be mentioned. Semiotics is a good starting point of the required ontology
structure. In semiotics (also called the theory of signs) one distinguishes three
interlinked parts according to Peirce (see (Peirce, 1885)):
The different levels are not separated. A linkage between the different levels
is introduced by particular connection relations (cf. the relationship with the
meaning triangle). For instance, syntax and semantics are connected through
a reference relation that links a sign with a set of statements. We consider
ontologies as models that are used to communicate meaning between machines
and human beings. Based on the semiotic view of ontologies described above
an ontology structure as given in the following Definition 2.1 is defined.
18 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
The model introduced above constitutes a core structure that will be conse-
quently used in this book. It is quite straightforward, well-agreed and may be
easily mapped onto existing ontology representation languages.
What is missing so far is an explicit representation of a lexical level that is
typically restricted to ontologies for natural language applications 5 - in spite
of its general usefulness. Therefore, a lexicon for the ontology structure 0 is
defined as follows:
and for
.1'- 1 (C) = {L E £cl(L, C) E .1'}
g and g-1 are defined analogously.
In general, one lexical entry may refer to several concepts or relations and one
concept or relation may be referred to by several lexical entries. An ontology
structure with lexicon is a pair (0, £), where 0 is an ontology structure and £
is an lexicon.
Ontology - Definition & Overview 19
///----------", ---- ~
" Organization -.>,------- ~
,I,' '\\ '"
{ organization 1 _-------------------
~ : ~---
\
"\
\
\
Person -
------~
: .
,,
\" EmpIOYee---,7'--------------------------___ ~
Up till now the semantics of the primitives introduced in the ontology struc-
ture 0 have not been elaborated. In the next chapter it will be shown how
this ontology structure may be realized in different, concrete representation
languages (e.g. the well-understood logical framework F-Logic (Kifer et al.,
1995) or the W3C standard resource description framework RDF(S)). The se-
mantic pattern approach as described in (Staab and Maedche, 2000; Staab et al.,
2001b; Staab et al., 2001a) will be used for mapping the ontology structure into
different representation languages. Semantic patterns are used for communica-
tion between Semantic Web developers on the one hand, but also for mapping
and reuse to different target languages on the other hand, thus bridging between
different representations and different ways of modeling knowledge. Develop-
ing the approach of semantic patterns, the wheel from scratch is not invented
from scratch, instead insights from software engineering and knowledge rep-
resentation research are picked and integrated for use in the Semantic Web.
Further details on semantic patterns are provided in the subsequent chapter 3.
The Knowledge Base Structure JCB. It was already introduced that a knowl-
edge base6 may be defined and instantiated using an ontology structure. The
20 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
general distinction between ontology and knowledge base is that one tries to
capture the conceptual structures of a domain of interest in the ontology, while
the knowledge base aims to specify a given concrete state. Thus, the ontology
is (mostly) constituted by intensional logical definitions, while the knowledge
base comprises (mostly) the extensional parts. The ontology is mostly devel-
oped during the set up (and maintenance) of an ontology-based system, while
the facts in the knowledge base may be constantly changing. These distinctions
("general" vs. "specific", "intensional" vs. "extensional", "set up" vs. "con-
tinuous change") indicate that for purposes of development, maintenance and
good design of an information system it is reasonable to distinguish between
ontology and knowledge base. Based on this idea the knowledge base structure
feB may be defined as follows:
Again, one may also define a lexicon for a given knowledge base structure
feB.
• a set CI whose elements are called lexical entries for instances, respectively.
In general, one lexical entry may refer to several instances and one instance
may be referred to by several lexical entries. A knowledge base structure with
lexicon is a pair (feB, C ICB ), where feB is a knowledge base structure and CICB
is an lexicon
Ontology - Definition & Overview 21
In this section the ontology and knowledge base structure have been intro-
duced. Both will be used within the overall book, providing a background for
the ontology engineering approach and layered representation that is introduced
in chapter 3. The ontology learning framework (see chapter 4) with its associ-
ated data & import processing techniques and algorithms also relies on these
definitions.
• Top-Level ontologies describe very general concepts like space, time, event,
which are independent of a particular problem or domain. It seems rea-
sonable to have unified top-level ontologies for large communities of users.
Recently, these kinds of ontologies have been also introduced under the
name "foundational ontologies".
Semantic Web. The development of the World Wide Web is about to mature
from a technical platform that allows for the transportation of information from
Web sources to humans (albeit in many syntactic formats) to the communication
of knowledge from Web sources to machines (Berners-Lee et aI., 2001). The
Semantic Web should be able to support automated services based on formal
descriptions of semantics. The semantic is seen as a key factor in finding a
way out of the growing problems oftraversing the expanding web space, where
currently most web resources can only be found by syntactic matches (e.g.,
keyword search).
The Semantic Web relies heavily on formal ontologies that structure under-
lying data for the purpose of comprehensive and transportable machine under-
standing. They properly define the meaning of data and metadata (cf. e.g.
(Staab et aI., 2000a; Decker et aI., 2000b)). In general one may consider the
Semantic Web more as a vision than a concrete application.
24 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
II
H,snd%~'buh
URl
Figure 2.5 illustrates the use of the terms "ontology" and "relational meta-
data". It depicts some part of the SWRC 7 (semantic web research community)
ontology. Furthermore it shows two homepages, viz. pages about Siegfried and
Steffen (http://www.aifb.uni-karlsruhe.deIWBS/sha) and (http://www.aifb.uni-
karlsruhe.deIWBS/sst), respectively) with annotations given in an XML se-
rialization of RDF facts. For the two persons there are instances denoted by
corresponding URIs (person_sha and person_sst). The SWRC: NAME of person_sha
is "Siegfried Handschuh". In addition, there is a instantiated relationship
COOPERATEWITH between the two persons: thus, they cooperate.
identified early e.g. by (Hobbs, 1993) and is now applied on real-world data
(Staab et aI., 1999).
3. Conclusion
In this chapter an introduction and persuasion why computer science requires
the "concept" of ontology has been provided. Subsequently, the roots of ontol-
ogy in philosophy and its core paradigm based on the meaning triangle have
been introduced. It has been shown how ontologies may support communi-
cation between humans and machines. Therefore, the idea that underlies the
meaning triangle has been combined with a "semiotics view" on ontologies
leading to a ontology and knowledge base structure. A formal introduction of
these structures describing their core elements and their interaction relevant for
this book have been given. As mentioned earlier the layered structures will
accompany the reader through the whole book. The next chapter introduces
26 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Notes
1 The OKBC model provides formal semantics for a large set of ontology
primitives enabling machine-processable semantics of information.
2 An interesting aspect is that ISO-704 is divided into three major sections:
concepts, definitions and terms. Concepts are seen as units of thought that
conforms to our view. The idea of a definition is that it fixes the concept in the
proper position of the system. Terms represent natural language representa-
tions of concepts. ISO-704 recommends that one concept should ideally be
represented by one natural language term.
3 Recently this form of meaning establishing communication has been coined
by the term "emergent semantics"
4 In this generic definition one does not distinguished between relations and
attributes.
5 The distinction of lexical entry and concept is similar to the distinction of
word form and synset used in WordNet (Fellbaum, 1998). WordNet has been
conceived as a mixed linguistic / psychological model about how people
associate words with their meaning.
6 In the Semantic Web the knowledge base is given by the relational metadata
that is defined for the different information sources.
7 http://ontobroker.semanticweb.org/ontos/swrc.html
8 http://www.semanticweb.org
9 The EU funded projects OnToKnowledge, http://www.ontoknowledge.org,
and, OntoLogging, http://www.ontologging.com. pursue exactly this goal.
10 http://www.ontology.org
11 http://www.hr-xml.org/channels/home.htm
12 http://www. papinet.org/
II
Da eben, wo Begriffefehlen, Da stellt ein Wort zur rechten Zeit sich ein.
-(Goethe, Faust I)
In recent years the development of ontologies has been moving from the
realm of Artificial Intelligence laboratories to the desktop of domain experts and
knowledge officers. Ontologies have become common to the World Wide Web,
ranging from very light weight topic hierarchies, such as large categorizations of
web sites (see Yahoo! 1 ) and product hierarchies (like the UNSPSC standard2 )
for use in B2C and B2B applications (see (Fensel, 2001» to heavy weight
ontologies, as described in (Staab and Maedche, 2001) for setting up knowledge
portals or in developing natural language understanding systems as described
by (Hahn and Romacker, 2000).
The development of the Semantic Web as a Meta-Web based on the World
Wide Web will require a large number of domain-specific ontologies. New
methods supporting this ontology engineering task are required, mainly because
of two facts: First, not only trained ontology engineers will develop ontologies,
also domain experts will model and maintain ontologies. Second, the fact that
the ontologies will be developed for the Semantic Web used by humans and ma-
chines in different applications, requires the establishing of new engineering
paradigms. A couple of methodologies (Usc hold and Gruninger, 1996; Guarino
and Welty, 2000; Gomez-Perez, 1996), representation languages (MacGregor,
1991; Scholze and Woods, 1992; Kifer et aI., 1995; Decker et aI., 2000a) and
ontology modeling tools (Swartout et aI., 1996; Fikes et aI., 1997; Grosso et aI.,
1999) have been developed that allow for the development, representation and
engineering of ontologies. In fact, these languages and tools have matured con-
siderably over the last few years. Nevertheless, the approaches cited above have
A. Maedche, Ontology Learning for the Semantic Web
© Kluwer Academic Publishers 2002
30 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
not been developed with the application "Semantic Web" in mind. Thus, the
work described in this chapter builds on and extends existing work supporting
ontology representation and engineering for the Semantic Web. In this chapter
an ontology engineering framework for the Semantic Web 3 is presented. It is
based on the layered ontology structure 0 and the knowledge base structure
KB introduced in the Definitions 2.1 and 2.3. The framework supports the de-
velopment of ontologies for a wide range of Semantic Web applications, three
examples for ontologies that have been developed on the basis of this frame-
work and the applications in which they have been used are also given in this
chapter. The ontology engineering framework described in this chapter forms
the basis for the concrete implementation of the ontology engineering system
ONTOEDIT that is described in further detail in chapter 7.
This chapter starts with section 1, where the overall ontology engineering
framework supporting the incremental and cyclic development of ontologies
is explained. The interaction between the ontology and its lexicon and the
ontology and the knowledge base is introduced. As mentioned above, it will be
shown how different types of ontologies may be developed using this framework
by giving three real-world modeling examples and application scenarios.
The basis for the underlying representation within the ontology engineer-
ing framework is the Resource Description Framework (RDF)4, a data model
developed by the WWW consortium W3C 5 as a foundation for the Semantic
Web. Different layers within RDF(S)6 will be distinguished and it will be shown
how specific representation languages, domain-specific ontologies and associ-
ated knowledge bases may be represented using the different RDF(S) layers.
Additionally, it will be discussed how mappings of the specific representation
language primitives (or representation vocabularies) into a logical layer to pro-
vide formal semantics are established. Finally, the semantic pattern approach
is introduced and it is shown how the ontology and knowledge base structure
may be mapped into a logical language and processed by an inference engine.
Concluding this chapter open questions regarding the ontology engineering
task will be discussed, including aspects on cooperative and multi-user engi-
neering, ontology federation and evolving ontologies and knowledge bases.
Finally, several requirements for a framework for ontology learning from an
engineering point of view will be summarized and defined. Additionally, the
concluding part of this chapter will prepare the subsequent chapter 4, where the
general framework for ontology learning for the Semantic Web is introduced.
• Second, though the ontology (see Definition 2.1) and the knowledge base
structure (see Definition 2.3) were defined separately, in reality there exists
no strict separation between the ontology and the knowledge base.
In the following the elements contained in the ontology and knowledge base
structure are embedded into a comprehensive picture.
AO
-- ~~~-
--------- --- -
---
1{C(C],C2 ) R(C1 , C 2 )
-........
R(Ib 12 )
C=g-- - - - - - - - -
C(I)
, .......
F
_f
~I:,C I:,R .J
KB
I:,
KB
Figure 3.1. Layered Ontology Engineering
Figure 3.1 depicts the overall layered and cyclic ontology engineering frame-
work. On the left side the ontology primitives as defined in Definition 2.1 and
on the right side the knowledge base primitives according to Definition 2.3 are
depicted.
In the bottom layer, one finds lexical entries representing signs for concepts
U:,c), relations (C n ) and instances (C/(6). The second layer comprises on the
left side the set of concepts C referenced by CC , the set of relations R referenced
by Cn, the concept taxonomy defined by statements such as 1{C(C1 , C2 ) and
non-taxonomic relations defined by statements such as R(C1 , C2 ). On the
right side the middle layer comprises the set of instances I referenced by C/(6,
32 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
the set of concept (by statements such as C (1) and relation instantiations (by
statements such as R(h, 12 ). The set of ontological axioms AO is defined on
top of the two existing layers. The reader may recognize the connection between
the ontology and the knowledge base: Instances are defined as members of
concepts via concept instantiation, relation instantiations refer to the set of
relations R defined in the ontology.
The overall incremental model consists of two main interactions: The first,
vertical interaction represents the dependency between the layers, the second,
horizontal interaction represents the dependency or overlap between ontology
and knowledge base. An example dependency for the first point may be th
requirement of the definition of a concept while defining a domain-specific
axiom. An example for the second point may be the definition of "default"
instances, such as specific instances of the concept CITY in the tourism domain.
Typically, the ontology engineer starts with a collection of lexical entries.
These lexical entries may be given by interviews or protocols as typical in knowl-
edge acquisition. Based on this collection of lexical entries the formalization
process starts: On the one hand, she defines concepts, conceptual relations and
a set of ontology axioms. On the other hand, the ontology engineer typically
develops the ontology with a specific application domain in mind. Therefore,
knowledge base entries may already be defined during ontology engineering.
These entries may be defined not only for debugging the ontology, but also to
constraint possible entries within the application (e.g. in a human resource sce-
nario the knowledge base may contain instance objects describing employees).
An implementation of the ontology engineering framework introduced in this
section has been realized by the ontology engineering environment ONTO EDIT,
which is described in the "system" chapter 7. Currently ongoing work on
a comprehensive methodology extends ONToEDIT with a component called
ONTO KICK, that supports the kick-off phase in ontology engineering (see
(Boyens, 2001».
As first scenario the SEAL-II framework (Hotho et al., 2001b) for devel-
oping knowledge portals and its instantiation by the HR-TopicBroker (Kuehn,
2001) application is considered. The application builds a company- and domain-
specific portal for human resource topics. Thus, it supports the allocation of
relevant topics (in the form of web pages) by using a focused web crawler,
it presents the crawled documents along the concept hierarchy and allows for
the joint definition of a simple knowledge base according to the given "topic
Layered Ontology Engineering 33
ontology" (e.g. it allows to define the contact information of people that are
experts for specific topics).
U sing the layered approach, lexical entries representing human resource
topics (e.g. "working hours model", "E-Learning") are first collected. The set
of concepts is derived by generalizing and formalizing the lexical entries, a
mapping between concepts and lexical entries is established. The application
is based on a set of concepts that are organized in a the taxonomy He. In
the next step, potential non-taxonomic relations between concepts have been
explored and discussed with the users (e.g. leading to relations such as the
relation CONTACT between TOPIC and PERSON).
Developing an ontology 0' for the TOPICBROKER application requires a set
of concepts C i E e, a set of lexical entries Li E Cp, (Ie I = I£ I) and a bijective
mapping F. The concepts e are embedded in the taxonomy He. Additionally,
a set of non-taxonomic relations R in the form of simple attributes has been
modeled to allow for ajoint definition of a knowledge base as described above.
All in all, 45 concepts, 76 lexical entries and five non-taxonomic relations have
been modeled in the ontology. Ontology 0' may be considered as a light-weight
ontology.
The second scenario of ontology engineering has been carried out in the
GETESS 7 application (see (Staab et al., 1999)). The application targets the
development of domain-specific search engines that support natural language
queries, and natural language processing of Web documents (e.g. generating
abstracts) using ontological background knowledge. The application is based
on an ontology consisting of a taxonomy of concepts and relations between
them. Bilingual (English I German) word stems (lexical entries that are used
by the natural language processing component) and external representations
(lexical entries that appear in the user interfaces) are mapped to the concepts
and the relations. Using the layered approach interfaces to the ontology as
well for the natural language processing component as for the user interface
component could be provided. Developing an ontology 0" for the GETESS
application requires a set of concepts Ci E e, a set of relations R j E R, lexical
entries Lk E £0 and mapping functions F, g, respectively. The concepts e
are embedded in the taxonomy He. All in all, in the tourism domain of the
GETESS project approx. 1120 concepts, 3200 lexical entries and 300 non-
taxonomic relations have been modeled in the ontology.
The third task of ontology development has been carried out in the devel-
opment of the ontology-based human resource management systems Proper
and OntoProper (see (Sure et al., 2000)) that uses means from decision the-
ory to allow for compensate skill matching. Additionally, these methods are
combined with intelligent means for inferencing of skill data. For the latter an
34 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
These three examples show what different types of ontologies used by dif-
ferent applications may be developed using the ontology and knowledge base
structures with the underlying layered and cyclic ontology engineering frame-
work. Thus, the special role of signs is considered (e.g. allowing the definition
of different types of lexicons) and the interaction between ontology and knowl-
edge base is supported (e.g. allowing the definition of predefined instances)
Translation Translation
Translation
into into Operation- Logical I
into
KIF
description logic
F-Logicwell alization ... Operationalization
axiomatic founded
semantics S1tIQ semantics
Layer
tI
t
I
t'-....... "'-Jt
I
DAML+ OIL OIL F-Logic Representation
Primitives Primitives Primitives
O/KB
Primitives
...
Vocabulary
------------------------------------ ----------------- -------- ----- ------------------------- Layer
Core representation vocabulary based on RDF-Schema
Data
Simple data model (resource description framework (RDF))
Layer
Syntax
Syntax based on XML
Layer
marriedWith
http://www.foo.comlW. Simth http://www.foo.com/S.Simth
a)
tt ://www.vatican.va/hoILfather
confirmed By
c) http://www.vatican.va/hoILfather
http://www.foo.comlW.Simth http://www.foo.com/S.Smith
by (Decker et aI., 2000c), the recently published book by (Hjelm, 2001) and
the more technical W3C specification of (Lassila and Swick, 1999).
RDF is an abstract data model that defines relationships between entities
(called resources in RDF) to a similar fashion as semantic nets. Statements
in RDF describe resources, that can be web pages or surrogates for real world
objects like publications, pieces of art, persons, or institutions. In a graphical
representation of an RDF statement, the source of the relationship is called the
subject, the labeled arc is the predicate (also called property) and the relation-
ship's destination is the object.
The RDF data model distinguishes between resources, which are objects
represented by URIs, and literals which are just strings. Resources may be
related to each other or to literal (Le. atomic) values via properties. Such a
relationship represents a statement that itself may be considered a resource, i.e.
reification is directly built into the RDF data model. Thus, it is possible to make
statements about statements. These basic notions can be easily depicted in a
graphical notation that resembles semantic nets. To illustrate the possibilities
of pure RDF the following statements are expressed in RDF and depicted in
Figure 3.3: 13
Layered Ontology Engineering 37
• Firstly, in part (a) of Figure 3.3 two resources are defined, each carrying
a FIRSTNAME and a LASTNAME property with literal values, identifying the
resources as William and Susan Smith, respectively. These two resources
come with a URI as their unique global identifier. They are related via the
property MARRIED WITH, which expresses that William is married with Susan.
• Part (b) of the illustration shows a convenient shortcut for expressing more
complex statements, i.e. reifying a statement and defining a property for the
new resource. The example denotes that the marriage between William and
Susan has been confirmed by the resource representing the Holy Father in
Rome.
• The RDF data model offers the predefined resource rdf: statement and the
predefined properties rdf:subject, rdf:predicate, and rdf:object to reify a
statement as a resource. The actual model for the example (b) is depicted in
part (c) of Figure 3.3. Note that the reified statement makes no claims about
the truth value of what is reified, i.e. if one wants to express that William
and Susan are married and this marriage has been confirmed by the pope.
Then the actual data model must contain a union of part (a) and part (c) of
the example illustration.
S subClassOf (rdfs:subClassOf)
R
D
domain (rdfs:domain)
range (rdfs:range)
~~
VI ..
T instanceOf (rdf:type) !'S
"".a
f ~
Ul 0
LL >
C c
a: 0
Li:;
c ..
a:
u "g
"'''"
.r; ::
~E
VI ..
. ..
C C
0"0
., C
.!! ..
"D.E
Uu
VI
··································11··
http://www.foo.comlW.Smith
appl:marriedWith
http://www.foo.com/S.Smith
~~
~ ti
C"!:""O
.2 ca !!!
~~E
c."ii
"",,
cui
ontology (see Figure 3.4) in the abstract data model. RDF-Schema offers a
distinguished representation vocabulary defined on top of RDF to allow the
modeling of object models. The most relevant RDF-Schema primitives are
given in the following list:
• The most general class is rdf: Resource. It has two subclasses, namely
rdfs: Class 14 and rdf: Property (see Figure 3.4 15 ). When specifying a do-
main specific schema for RDF(S), the classes and properties defined in this
schema will become instances of these two resources.
• The resource rdfs:Class denotes the set of all classes in an object-oriented
sense. That means that classes like appl: Person or
appl: Organisation are instances of the meta-class rdfs: Class.
• The same holds true for properties, i.e. each property defined in an applica-
tion specific RDF-Schema is an instance of rdf: Property, e.g.
appl:marriedWith
• RDF-Schema defines the special property rdfs: subClassOf that defines the
subclass relationship between classes. Since rdfs: subClassOf is transitive,
definitions are inherited by the more specific classes from the more general
classes. Resources that are instances of a class are automatically instances
Layered Ontology Engineering 39
<rdf:RDF
xmlns:rdf ''http://www.w3.org/1999/02/22-rdf-syntax-ns#''
xmlns:rdfs = ''http://www.w3.org/2000/01/rdf-schema#''
xmlns "example-ontology.rdfs#">
<rdfs:Class rdf:ID="Person"/>
<rdfs:Class rdf:ID="Organization"/>
<rdfs:Class rdf:ID="Han">
<rdfs:subclassOf rdf:ID=IIPerson">
</rdfs:Class>
<rdfs:Class rdf:ID="Woman ll )
<rdfs:subclassOf rdf:ID="Person")
</rdfs:Class>
<rdf:Property rdf:ID="firstName">
<rdfs:domain rdf:resource="#Person"/>
<rdfs:range rdf:resource=''http://www.w3.org/TR/xmlschema-2/#string''/>
</rdf:Property>
<rdf:Property rdf:ID=lIlastName ll )
<rdfs:domain rdf:resource="#Person"/>
<rdfs:range rdf:resource=''http://www.w3.org/TR/xmlschema-2/#string''/>
</rdf:Property>
<rdf:Property rdf:ID="marriedWith">
<rdfs:domain rdf:resource="#Person li />
<rdfs:range rdf:resource="#Person ll />
</rdf:Property>
</rdf:RDF>
resource. The next level of nested tags gives properties of the resource denoted
by the ID. Figure 3.5 gives an example for the RDF schema serialization syn-
tax. The example represents in XML the classes and property types defined
in Figure 3.4. Additionally, domains and ranges of the properties are defined
using the RDF constraint properties rdfs: domain and rdfs: range.
<rdf:RDF
xmlns:rdf = ''http://www.w3.org/1999/02/22-rdf-syntax-ns#''
xmlns:app = "example-ontology.rdfs#">
Figure 3.6. XML Serialization of RDF instances, their literals, and relations between them.
Figure 3.6 gives an example for RDF instances defined according to the
ontology given in Figure 3.5. The application specific classes <app : Man> and
Layered Ontology Engineering 41
<rdfs:Class rdf:ID;"LexicalEntry"/>
<rdfs:Class rdf:ID=IILanguage ll )
<rdfs:Class rdf:ID=IIDE H )
<rdfs:subClassOf rdf:resource="#Language"/>
</rdfs:Class>
<rdf:Property rdf:ID=ltreferences">
<rdfs:range rdf:resource;''http://www.w3.org/2000/01/rdf-schema#Resource"/>
<rdfs:domain rdf:resource="#LexicalEntry ll/>
</rdf:Property>
<rdf:Property rdf:ID=lIlanguage ll )
(rdfs:range rdf:resource="#Language"l>
<rdfs:domain rdf:resource="#LexicalEntry"/>
</rdf:Property>
<rdf:Property rdf:ID="value")
<rdfs:range rdf:resource;''http://w,,,,.w3.org/2000/01/rdf-schema#Literal"/>
<rdfs:domain rdf:resource="#LexicalEntry"/)
</rdf:Property>
</rdf:RDF>
<rdf:Description rdf:ID="A189">
<rdf:type resource=lIhttp://ontoserver.aifb.uni-karlsruhe.de/
schema/oevoc.rdfs#LexicalEntryli/>
<references resource="#Organization ll />
<value>Organisation</value>
<language resource=lIhttp://ontoserver.aifb.uni-karlsruhe.de/
schema/oevoc.rdfs#DE")
</rdf:Description>
</rdf:RDF>
An example for defining and representing the OIL description logics vocab-
ulary on top of RDF-Schema has been given in (Decker et aI., 2000a; Fensel
et aI., 2000). Figure 3.9 depicts the derived OIL extensions, using as many
as possible RDF-Schema constructs. The usage of RDF-Schema primitives as
a basis for the OIL representation vocabulary has the advantages that agents
that do not understand OIL, at least may understand the classes, properties and
subclass relationships.
rdfs.:Resource
otl:ExpteSsian
,--_ _ _ _.--_..1.'- - - - - - , I
oil:TranSili~eProperty oil:FLlftet~.IPropN1)'
~
tdfa:Ctass
I
oU:SaoieanExp'e.ston
oil:p,opert~Re$f1ietlon o":SymmetrieP/oJ)etty
oil:ConetereTypeExpr.ssion
The core idea. A rough outline of how semantic patterns may be developed
and used is given in the following. For a comprehensive introduction the inter-
Layered Ontology Engineering 45
ested reader is referred to (Staab et aI., 200la). The work on semantic patterns
has been motivated first by axiom schemata (Gruber, 1993a). While axiom
schemata already go into the direction of abstracting from formal model char-
acteristics (see (Staab et aI., 2001b)), by definition they are developed for one
language only. Hence, one part of the high-level idea was to allow for (an
open list of) new epistemological primitives (see (Brachman, 1979)) that can
be instantiated in different representation languages for modeling particular
semantic entailments and which are, thus, similar to named axiom schemata
working in one language.
However, one needs a more flexible paradigm better suited to apply to a larger
range of representation languages and able to abstract more from particular
formal models. As described above, the general problem does not allow to
come up with a completely formal and ubiquitously translatable specification
of semantics. Hence, the other part of the high-level idea is to require extra
efforts from Semantic Web developers. To support them in their efforts, it
appeared to be a prerequisite that they could communicate more efficiently
about these new epistemological primitives - similar to the way that software
engineers talk about recurring software designs.
Design patterns have been conceived for object-oriented software develop-
ment to provide (i) a common design vocabulary, (ii) a documentation and
learning aid, and (iii) support for reorganizing software. Likewise to the nam-
ing and cataloguing of algorithms and data structures by computer scientists,
design patterns are used by software engineers to communicate, document and
explore design alternatives by using a common design vocabulary or a design
pattern catalog. This way, they also decrease the complexity of developing
and understanding of software systems. Additionally, design patterns offer so-
lutions to common problems, help a novice "acting" more like an expert and
facilitate the reverse-engineering of existing systems.
Though bridging between formal representations seems to be a formal task
only, very often quite the contrary becomes true, when not everything, but
only relevant aspects of knowledge can or need to be captured, when not all
inferences, but only certain strains of semantic entailments can or need to be
transferred. The development of new semantic primitives should not only al-
lude to the formal definition of translations into target languages, but also to
informal explanations. Therefore a semantic pattern does not only comprise
new epistemological primitives, but likewise to design patterns, it also serves
as a means for communication, cataloguing, reverse-engineering, and problem-
solving. Thus, it may contribute to a more efficient exploitation of Semantic
Web techniques.
Figure 3.10 summarizes our approach for modeling axiom specifications in
RDF(S) in an overall picture. It depicts the core of the RDF(S) definitions
and our extension for some example semantic patterns (i.e. our ontology meta
46 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
_ subClassOf (rdfs:subClassOf)
.............
......
~:-.:........... .,..
., I",
~ ,.',',,\
" I I, 1
., I I I 1
" I 'I ,
/:
, , I I 1
I
,,' I,
",
" "
I I II ,
1
I I ' ,
I I J ,
,,I / : \
I I I ,
'". '" .
U
<LeU
w ~ ~
~E=
~~~
a.
w
~
••••••••••••• •••• :~ .. ~ • ••••••••••••••••••••••••• •••••• ~.:.-.:-.:- .... - ......:!... :................................................ .
http://www.foo.comNV.Smith
,
appl:marriedWith
--.
http:/twww.foo.comlS.Smith
The definition of the axiom in the fonn of a semantic pattern has the ad-
vantage, that it captures the axiom in a very generic way that is independent
of the concrete representation language. The instantiated semantic pattern IN-
VERSEREL may than be easily translated in F-Logic.
The next example concerns composition of relations. For instance, consider
if a first person is FATHEROF a second person, who is MARRIED WITH a third person
then one may assert that the first person is the FATHERINLAWOF the third person.
The definition such an axiom may be denoted in F-Logic as follows:
future one may develop libraries of semantic patterns that are driven by the
requirements of Semantic Web developers and specific application domains.
Pattern Libraries. With the engineering of ontologies on the Web new ideas
will come up about what type of inferencing should be supported and, hence,
made interchangeable between representation systems. Since this development
is in its infancy right now, a number of semantic patterns that seem widely
applicable have been collected:
• Gruber's Frame Ontology (Gruber, 1993a) includes a set of over 60 prim-
itives, some of which are found in core RDF(S), e.g. rdf: type, and some
of which are more sophisticated, e.g. symmetry of relations or composition
(database joins).
• Medical knowledge processing often relies on the engineering of part-whole
reasoning schemes that appear or do not appear when considering the fol-
lowing examples: (i), the appendix is part of the intestine. Therefore, an
appendix perforation is an intestinal perforation. And, (ii), the appendix is
part of the intestine, but an inflammation of the appendix (appendicitis) is
not an inflammation of the intestine (enteritis). In (Staab et aI., 2000b) it is
described how to represent structures that allow for expressing (for (i)) and
preventing (for (ii)).
• Inheritance with exception is a semantic pattern that is very often useful.
Its application is tractable, even efficient, and the reasoning part has been
described, e.g., in (Morgenstern, 1998). The core idea is that one considers
the inheritance of properties, allows for the non-inheritance of certain prop-
erties, and uses a particular, unambiguous strategy for resolving conflicts
between paths of inheriting and non-inheriting a particular property. A sim-
ple example is that a PATIENT'S treatment may be covered by medical in sur-
Layered Ontology Engineering 49
Note that often there is no translation into particular target languages for
this pattern. For instance, it can be achieved in Prolog or F-Logic, but not
in the standard description logics systems .
Multiple and Federated Ontologies & Knowledge Bases. It has been shown
that ontology-based community websites work well (Staab et aI., 2000a) with
a central community ontology. However, a problem to this approach may be
the global categorization schema. In the form of the central ontology it may
not be understood and accepted by all community members in the same way
(see (Lacher and Groh, 2001). In real life communities, community members
would like to keep their own perspective and view on the community repository.
For representing such specific views one needs a distributed ontology setting,
where explicit knowledge is exchanged via ontology mappings. In (Lacher
and Groh, 2001) an approach for automatically performing ontology mappings
via analyzing document categorizations done by users has been described. The
approach is very similar to the approach for ontology merging described in
the next chapter, building on an extensional description of concepts and deriv-
ing mappings or merging by comparing extensional descriptions. Supporting
Layered Ontology Engineering 51
• How fast can you develop an ontology? How much time does it take to build
an ontology for the target application?
• How do you know the ontology has a good coverage for a given domain of
interest?
• How does the ontology develop over time? Are there any mechanisms that
support the problem of knowledge maintenance?
In fact, these problems of time, difficulty, confidence, and maintenance were
similar to what knowledge engineers have dealt with over the last two decades
when they elaborated on methodologies for knowledge engineering or work-
benches for defining ontologies and knowledge bases.
A generic method that proved extremely beneficial for the knowledge engi-
neering task was the integration of knowledge acquisition with machine learning
techniques (see the seminal work described in (Skuce et aI., 1985; Reimer, 1990;
Szpakowicz, 1990; Buntime and Stirling, 1991; Morik et aI., 1993a; Nedellec
and Causse, 1992; Webb, 1996». Current work on applying machine learn-
ing for knowledge acquisition and engineering mainly restricted its attention to
structured knowledge, normally in the form of instances, given in the form of
more or less complex knowledge bases.
The World Wide Web uses a wide range of different data (see the data taxon-
omy depicted in Figure 4.1). Thus, in this work it is targated at the integration
of a multitude of disciplines in order to facilitate the construction of ontolo-
gies, in particular machine learning. Because the fully automatic acquisition of
know ledge by machines remains in the distant future, the process of ontology
learning is considered as semi-automatic with human intervention, adopting the
paradigm of balanced cooperative modeling (Morik, 1993) for the construction
of ontologies for the Semantic Web. With this objective in mind, an architec-
ture that combines knowledge acquisition with machine learning, feeding on
the resources that we nowadays find on the syntactic Web, has been built. The
layered ontology engineering approach introduced in this section builds the ba-
sis for human intervention in the ontology engineering process. Thus, thinking
about how to integrate ontology learning for supporting manual engineering a
number of requirements had to be fulfilled. In the following these requirements
that make up a generic framework supporting the idea of Ontology Learning
are defined:
• The framework for ontology learning should define relevant phases in that
one would like to support the ontology engineer. Ontology learning should
be embedded in a process-oriented view.
• The framework should offer ways for user interaction. The user should
easily access and import data. Data may be transformed and directly passed
to an ontology learning algorithm.
• The user should be guided in finding relevant data for ontology learning.
The access and the process of this data should be managed by the system. A
combination of different data should be enforced and compensate different
aspects of knowledge and different data quality.
Notes
1 http://www.yahoo.com
2 http://www.unspsc.org
3 Part of the work described in this chapter has been published in (Maedche
et aI., 2000; Staab and Maedche, 2000; Erdmann et aI., 2000).
4 http://www.w3c.orgIRDF
5 http://www.w3c.org
6 "RDF(S)" is used to refer to the combined technologies of RDF and RDF-
Schema.
7 German Text Exploitation and Search System - http://www.getess.de
8 http://www.w3.org/XML/
9 As depicted in Figure 3.2 DAML+OIL and OIL do not built on all RDF-
Schema primitives and exclude some of them, e.g. reification. There is
still an ongoing discussion on the primitives that may be contained in future
versions of RDF-Schema
10 http://www.ontoknowledge.org/oill
11 http://www.damI.orgI200l/03/daml+oil-index
12 http://www-sop.inria.fr/acacialdrdfs/
13 Resources are represented by shaded ovals, literal values by rectangles, and
properties by directed, labeled arcs.
14 The reader may note that term "class" used in RDF(S) is used as a synonym
of the term "concept", as given in the ontology structure definition.
15 The reader may note that only a small part of RDF(S) is depicted in the
RDFIRDFS layer of the figure. Furthermore, the relation APPL:MARRIED WITH
in the data layer is identical to the resource
APPL:MARRIEDWITH in the schema layer.
16see http://www.w3.org/1999/02/22-rdf-syntax-ns
17 see http://www.w3.org/2000/01lrdf-schema
18 The namespace is accessible at
http://ontoserver.aifb.uni-karlsruhe.de/schema/oevoc.rdfs
19The reader may note that the discussion of the semantics of RDF(S) is still
going on, e.g. see the W3C RDF-Logic mailing list, archive available at
http://lists. w3.org/ArchiveslPublic/www-rdf-logic/.
20 http://www.ontoknowledge.org/oilldownllsemantics.pdf
21 http://www.damI.org
22DAML+OIL is a joint initiative of the DARPA-DAML project and EU-
funded projects.
23 see http://www.damI.org/200 1I03/daml+oil-index.html
24 see http://logic.stanford.edulkif/kif.html
Layered Ontology Engineering 55
ontology learning algorithms will put life into these core components. The
comprehensive architecture introduced here has been implemented in the on-
tology learning and engineering environment TEXT- TO-ONTO (Maedche and
Volz, 2001) that will be presented in chapter 7.
In the third part of this chapter a process-oriented view on ontology learn-
ing will be introduced. This process-oriented view embeds four main phases,
namely import, extract, prune and refine into an ontology learning cycle.
Based on these four phases, an overall picture of how to apply ontology learning
for extraction and maintenance of ontologies will be provided. One important
aspect in the Semantic Web, besides building up an ontology by importing
existing knowledge structures and extracting new structures is to consider the
"maintenance phases" necessary for the pruning of an ontology and for the
refining of the ontology supported by ontology learning techniques.
semi-
structured
•
data
NL docs with
pure semi-structured
NLtext inf01tion
ling.
ontologies
(Wordnet)
reI. ER 00 DTD XML-S
model dictionaries
Figure 4.1 provides an overview and a classification of the relevant data for
ontology learning in the form of a taxonomy. Five main types of relevant data
are distinguished: ontologies, schemata, instances, web documents and semi-
structured data. These five generic types are further refined into more specific
kinds of data and are explained in further detail in the following section.
Ontology Learning Framework 61
Data in the form of ontologies. The first type of data considered is the left
branch of the taxonomy labeled "ontologies". Ontologies may also be con-
sidered to be a specific kind of data. As introduced in the ontology overview
chapter, different kinds of ontologies are available, e.g., linguistic ontologies
(such as WordNet (Fellbaum, 1998) and its German counterpart GermaNet
(Hamp and Feldweg, 1997) discussed in Section 1), thesauri (Wersig, 1985) or
domain-specific web ontologies 2 . The topic of reusing already existing on-
tologies was a topic of discussion long before large libraries of ontologies were
available (see (Pirlein, 1995». This has changed and is becoming increasingly
important: The Semantic Web is being built on top of domain-specific schemata
in the form of ontologies. Currently and continuing on in the future more and
more ontologies will be available (by using the XML namespaces mechanism)
and the quick adaptation of an ontology from one domain to another or the
extension of a given ontology will become critical.
As mentioned above in the context of this book, reuseable ontology structures
are considered to be available in different kinds of ontologies. Examples are the
large linguistic lexical-semantic nets, e.g., WordNet or its German counterpart
GermaNet, and, domain-specific ontologies, e.g., switching from a tourism on-
tology to a finance ontology as done in the GETESS project (Staab et aI., 1999).
One may also consider thesauri 3 (Wersig, 1985) to be light weight ontologies
that may be reused to derive more comprehensive ontologies 4 • The interested
reader is referred to (Amann and Fundulaki, 1999), where an interesting ap-
proach for the re-engineering of thesauri to ontologies has been described. In
their work an ontology represented in RDF(S) has been derived by exploiting
existing domain-specific ontologies and thesauri.
Further elaboration on how to import existing ontologies and similar struc-
tures will be discussed in the the next chapter.
Thus, their main target is to provide more sophisticated query and retrieval
for documents than current systems provide.
data graph and represent a compact structural description of the given set of
semi-structured data.
The interested reader can refer to the comprehensive overview on the con-
nection between ontologies and semi-structured data (with specific focus on
XML) presented in (Erdmann, 2001).
Data in the form of natural language text. Natural language documents are
data that are freely available in large amounts on the Web. This kind of data
is considered to be the most important source of data for ontology learning
for the Semantic Web. Therefore, the applicability of the techniques laid out
in this book can be guaranteed, because great amounts of this kind of data are
available for all domains of interest. In general, the following distinction can
be made:
• Pure natural language text: Natural language text exhibits morphological,
syntactic, semantic, pragmatic and conceptual constraints that interact in
order to convey a particular meaning to the reader. Thus, the text transports
information to the reader and the reader embeds this information into his
background know ledge.
Linguistic constraints found in the language serve as an important input for
ontology learning. Real-world data on the web requires specific techniques
for recognizing and analyzing these constraints. Shallow text processing
and the underlying techniques that are applied in the overall framework are
introduced in chapter 5.
• Natural language documents enriched with semi-structured informa-
tion: As mentioned above when referring to the success of new standards
for document publishing on the web, there exists a proliferation of semi-
structured data on the web. Formal descriptions of semi-structured data are
available freely and widely, e.g., HTML data adds more or less expressive
semantic information to documents. Documents on the WWW typically
combine natural language text with semi-structured information, e.g., the
information contained in tables, lists, etc. Dealing with this kind of infor-
mation is a challenge and requires specific techniques.
One specific kind of a semi-structured document is found in online dictio-
naries on the web 9 . In (Litkowski, 1978) it is argued that the definitions
contained in dictionaries hold a great deal of information about the seman-
tic charateristics that should be attached to a lexeme. Dictionaries serve as
stable resources of domain knowledge that provide a good starting point for
a core ontology by supporting subsequent ontology learning from pure text.
ing chapters attention will only be paid to two specific kinds of data, namely
ontologies and natural language documents. However, the architecture that is
presented in the following section is independent of specific input data.
66 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Algorithm
Library
The overall architecture and the interaction between the four components
introduced above is graphically depicted in Figure 4.2. The reader may note
that the ''functional components" (the data import & processing and the algo-
rithm library component) are hidden by the ''interface components" (graphical
68 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
One important aspect of ONTO EDIT is the account of the layered language
approach, ontologies in different representation languages (e.g., representation
vocabularies like OIL 12, DAML+OIL13, F-Logic or OntoEdit specific language
primitives) may be developed. New representational requirements (e.g., new
modeling primitives) can easily be defined on account of the layered approach:
New ONTO EDIT specific ontology representation primitives are represented
in the OntoEdit namespace. 14
An important advantage of the overall ontology learning architecture is that
ontologies in different representation languages may be accessed. On the other
hand, ontology elements that are extracted (e.g., a concept taxonomy He) may
70 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
All the techniques and mechanisms introduced above support the manual
generation of high quality web ontologies. However, it was already mentioned
that there is a large conceptual gap between the ontology engineering tool and
input (and often legacy data) such as Web documents, which ultimately deter-
mine the target ontology. The next three components support the extension of
the manual ontology engineering environment ONTO EDIT to a comprehensive
framework for extracting and maintaining ontologies for the Semantic Web.
As depicted in Figure 4.2 the final result of the data import & processing
module is either an instantiated ontology structure (e.g. through importing or
merging) or "blocks" of algorithm-specific preprocessed data (depicted in the
lower left part of Figure 4.2).
(Kietz et ai., 2000a) and finance, it is to be expected that there are some kinds
of ontology structures available for almost any commercially significant do-
main. Thus, mechanisms and strategies are required to import & reuse these
structures. The import & reuse process may be roughly separated into two
parts:
• In the first part, relevant ontologies have to be selected and importing strate-
gies have to be defined, e.g., an ontology wrapper supporting the transfor-
mation from one representation language to another has to be defined.
• In the second part of the import & reuse step, imported conceptual struc-
tures need to be merged. This should for the basis for the subsequent ontol-
ogy learning phases of extracting, pruning and refining.
While the general research issue of merging and aligning is still an open prob-
lem, recent proposals (e.g., (Noy and Musen, 2000)) have shown how to improve
the manual process of merging/aligning. Existing methods for merging/aligning
mostly rely on matching heuristics to propose merging concepts and similar
knowledge-base operations. As mentioned earlier, the FCA-MERGE method
is used for generating a common ontology, such as input for learning. This
mechanism follows an application data-oriented, bottom-up approach based
on formal concept analysis (Ganter and Wille, 1999). A detailed introduction
and examples of this approach are given in the next chapter of this book.
3.2 Extract
In the ontology extraction phase of the ontology learning cycle, major parts
(i.e., the complete ontology or large chunks reflecting a new sub domain of
the ontology) are modeled with learning support to exploit the various types of
relevant data. During this process, ontology learning techniques partially rely
on given ontology parts. Thus, encountering an iterative growing model during
previous revisions to the ontology learning cycle may result in subsequent ones.
More sophisticated algorithms may work best with the structures in combination
with the more straightforward ones, which where previously applied.
An example for this iterative, bootstrapping model is instantiated with the
non-taxonomic relation extraction technique described in chapter 6. Relations
between concepts may be extracted based on a given set of concepts C and an
associated lexicon C. However, if a taxonomic order He of these concepts
is given, more comprehensive, generalized relations between concepts may be
extracted. Thus, the existing background knowledge allows for the generaliza-
tion of the results. The algorithms for extracting the ontological structures are
described in detail in chapter 6.
76 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
3.3 Prune
A common theme for modeling in various disciplines is the balance between
the completeness and the incompleteness of the domain model. It is a widely
held belief that targeting completeness for the domain model appears to be
practically inmanagable and computationally intractable. Targeting the incom-
plete model, on the other hand, is overly limiting with regard to expressiveness.
Hence, striving for a balance between these two models is one of the greatest
challenge to successful ontology learning. The import & reuse of ontologies,
as well as the extraction of ontologies can result in imperfect (often too large
and comprehensive) models . Therefore pruning the ontology to diminish its
size is of great importance. Thus, there are at least two dimensions to the
problem of pruning that need to be examined:
• First, one needs to clarify how the pruning of particular parts of the ontol-
ogy (e.g., the removal of a concept or a relation) affects the ontology as a
whole. For instance, (Peterson et aI., 1998) describe strategies that leave
the user with a coherent ontology (i.e. no dangling or broken links). A sim-
ilar strategy has been described by (Swartout et aI., 1996) where ontology
pruning is considered to be the task of "intelligent" deletion of ontological
structures. The ontology engineering environment ONTOEDIT employs a
specific technique for the elimination of concepts & relations: If a user wants
to eliminate a specific concept, all the ontological structures concerned are
computed and proposed to the user. Finally, the user decides if the specific
concept or the relation should be eliminated.
• Second, one may consider strategies for proposing ontology items that
should be either kept or pruned. Several mechanisms for generating propos-
als from application data have been investigated (see (Kietz et aI., 2000b)
and chapter 6, subsection 2.1). These proposals follow an application data-
driven approach: Concepts and relations that are never instantiated in a set of
relevant documents are not considered to be necessary and their elimination
is proposed.
Given a set of application-specific documents, there are several strategies for
pruning the ontology:
• First, one may count lexical entry frequencies. Concepts referring to lexi-
cal entries that seldom appear are deleted from the ontology. Lexical entries
may also be deleted or mapped to more generic concepts that remain in the
ontology. Thereby, one may substitute simple lexical entry frequencies with
more sophisticated information retrieval measures, such as term frequency
/ inverse document frequency (tfidf), which seems to offer a more balanced
estimation of the importance of a lexical entry and its corresponding con-
cept. Additionally, using the concept hierarchy 1I.c
one may propagate the
Ontology Learning Framework 77
frequencies via their concepts through the taxonomy. This method has the
advantage that top level concepts which serve as structuring concepts are
not removed from the ontology.
• Second, a more sophisticated pruning technique compares the lexical en-
try frequencies of a domain specific document collection (e.g., reports about
electronic products) with those of a generic reference document collection
(e.g., general newspaper reports). Thus, one may avoid having very inter-
esting, but rare domain-specific lexical entries and their concepts pruned
from the ontology.
A detailed description of the pruning algorithms is given in the second part
of chapter 6.
3.4 Refine
Refining plays a role similar to extracting. The difference between these
methods and when to use them cannot be laid down with a clear-cut distinction.
While extracting serves mostly the cooperative modeling of the overall ontology
(or at least very significant chunks of it), the refinement phase is about fine
tuning the target ontology and the support of its evolving nature. In principle,
the same algorithms may be used for extraction as for refinement. However,
during refinement one must consider the existing ontology in detail and the
existing connections to the ontology, while extraction works more often than
not practically from scratch.
A prototypical approach to refinement (though not to extraction!) has been
presented by (Hahn and Schnattinger, 1998». They have introduced a method-
ology for automating the maintenance of domain-specific taxonomies. An
ontology is incrementally updated as new concepts are acquired from text. The
acquisition process is centered around the linguistic and conceptual "quality" of
various forms of evidence underlying the generation and refinement of concept
hypotheses. In particular, semantic conflicts and analogous semantic structures
from the ontology are considered in order to determine the quality of a partic-
ular proposal. Thus, an existing ontology is extended with new lexical entries
for £, new concepts for C and new taxonomic relations for He.
The approach to ontology refinement described in this book is mainly based
on the assumption that unknown lexical entries share similar "conceptual be-
haviour" with respect to already known lexical entries (lexical entries that are
assigned to concepts via F). In addition to this assumption, frequency distri-
butions are compared using different statistical similarity measures.
The refinement phase may also use data that comes from a concrete Semantic
Web application, e.g., log files of user queries or generic user data. Adapting
and refining the ontology with respect to user requirements plays a major role
in the acceptance of the application and its further development. This basic
78 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
idea is also called "Semantic Web Mining" and will be briefly introduced in
the concluding chapter of this book.
4. Conclusion
This chapter defined and explained the overall ontology learning framework
that underlies the work presented in this book. The framework is based on the
foundations and definitions of ontologies and knowledge bases and the layered
ontology engineering paradigm introduced in the first part of this book.
In the first part of this chapter, relevant data for ontology learning was de-
fined providing a basis to roughly distinguish between ontologies, schemata,
instances, web documents and semi-structured data. On account of the com-
plexity of the overall topic ontology learning, attention was restricted to data
in the form of natural language text and existing ontologies. Data in the form of
instances and web and database schemata and semi-structured data has been pur-
posely neglected. The reader may note that the generic architecture introduced
here is mainly abstracted from the concrete data and offers core components to
support ontology learning. Thus, the purpose of the second part of this chapter
was to give an overview of the generic architecture for learning ontologies for
the Semantic Web and its relevant components.
Moreover, ontology learning was embedded in a process-oriented cycle con-
sisting of the core phases import, extract, prune and refine. These four main
phases reflect the fact that ontology learning can be applied to ontology building
on the one hand and to ontology maintenance, on the other hand.
Chapters 5 and 6 which will follow and belong to this portion of this book,
will instantiate the ontology learning components of the architecture and fur-
ther elaborate on data import & processing, as well as on ontology learning
algorithms which support extracting and maintaining ontologies. Chapters 7
and 8 will describe the implementation of the architecture in an actual system
and its evaluation.
Ontology Learning Framework 79
Notes
1 Parts of this chapter has been published in (Maedche and Staab, 2001c;
Maedche and Staab, 200Ia).
2 An example of a large library of ontologies is the collection developed in the
DAML project, that is available at
http://www.damI.org/ontologies/
3 A large number of domain-specific thesauri are available online at
http://www.thesaurus.com!
4 Classification schemes in the form of thesauri are standardized in ISO 2780,
see "Documentation - Guidelines for the establishment and development
of monolingual thesauri. International Organization for Standardization, 11
1986, Ref. No. ISO 2788-1986".
5 http://www.w3.org/XML/1998/06/xmlspec-report
6 http://www.w3.org/XML/Schema
7 http://www.w3.orgIXML/
8 The system is available for download at http://www.isi.edulisd/LOOMI.
9 A directory of domain specific dictionaries is accessible at
http://www.dictionary.com.
10 Similar to the area knowledge discovery in databases (KDD) the preprocess-
ing task is considered as difficult and time-intensive.
11 Examples and a detailed description of the comprehensive modeling envi-
ronment ONTO EDIT are given in section 2.
12 http://www.ontoknow ledge.org/oil
13 http://www.damI.org/daml+oil
14 The namespace is online available at
http://ontoserver.aifb.uni-karlsruhe.de/ontorep.
15 An introduction to the main memory deductive, object-oriented database sys-
tem is given in (Decker et aI., 1998). The engine is available for download at
http://www.ontoprise/download/. Its successor, called TRIPLE, is currently
under development, see
http://www.dfki.uni-kl.de/frodo/triple/index.html.
16The approach is motivated by the multi-strategy learning idea.
17 CyC is a large common-sense ontology, cf. http://www.cyc.com.
Chapter 5
j
fktol0JyOntoIOS'! Cmw1er Do!> NtP TUl1),,-
Wroppef Metgin@ Wmpper System lCfMiMkm
and indexing from the Web, where a relevant set of document data is "compiled"
by applying a semi-supervised algorithm, is introduced. As mentioned earlier
in the last chapter, one central sUb-component for data import & processing is a
natural language processing (NLP) system. The architecture of the system and
the underlying techniques for shallow text processing are introduced in sub-
section 2.2. In particular, extensions of the system supporting the interaction
between shallow linguistic processing and the ontology are described. Domain-
specific dictionaries are considered as a stable source of know ledge for deriving
ontologies. Our approach of document wrapping described in subsection 2.3
allows a fast import and normalized representation of a given dictionary. Subse-
quently, the normalized dictionary is given directly to the learning component.
Finally, one important issue of ontology learning or machine learning in gen-
eral is to find the right representation for the application of a given algorithm.
In subsection 2.4 it is formally defined what type of relational structures are
generated from the linguistically normalized data. The concluding section 3
summarizes the content of this chapter and defines a list of further work that
has not been approached here.
In the following a short example is given how the lexical semantic nets
WordNet and GermaNet have been imported into the framework.
• Holonym: the name of the whole of which the meronym names a part. Y
is a holonym of X if X is a part of Y.
• Meronym: the name of a constituent part of, the substance of, or a member
of something. X is a meronym of Y if X is a part of Y.
There exist more primitives in WordNet, like reason, link and pertainym.
However, on account of their lexical motivation it was decided not to integrate
them. Table 5.1 lists the mappings that have been performed for the GermaNet
and WordNet lexical-semantic ontology to the ontology structure D.
Figure 5.2 depicts an example excerpt from tic extracted from WordNet on
the left side and GermaNet on the right side visualized with the ONTOEDIT
concept hierarchy view. The conversion of GermaNet results in an ontology
consisting of approx. 20, 000 concepts and 2 relations (meronym, holonym)
with 2,713 domain and range restrictions. The taxonomy tic has an average
depth of 7.19 and a maximum depth of 18.
Ji,'f~z~t~z:~H"k\ A~;t.'m:$t>¥
C{f.:rH#:,:)~r,:~t f-Xfl;NN':$j:;·:.'?iH4iWN, r:i::wWn~
:$--
:(.: ~&~:~;t~b:t
:t-nat:t~
"'?+f$
" ,'*
?!i;,:Iw%*b~""
\'hNk'i"",ub; L&lrw
'@ h%~!eNMlI\, hKhkkm
GblWtMd"wlh, U","owtil&lr
merging two ontologies into one target ontology is required. The process of
ontology merging takes as input two (or more) source ontologies and returns a
merged ontology based on the given source ontologies.
A new method, called FCA-MERGE, has been developed for merging on-
tologies following a bottom-up approach which offers a structural description
of the merging process 3 . The method is guided by an extensional description
of the set of concepts C of two given source ontologies, that are to be merged.
The extensional description of concepts is derived by using the natural language
processing core system4 . Extensional descriptions are stored in so-called con-
texts that serve as input to a theory called formal concept analysis (cf. (Ganter
and Wille, 1999». Formal concept analysis derives a lattice of concepts as a
structural result of FC A-MERGE. The result is then explored and transformed
into the merged ontology with human interaction.
The set of all formal concepts of a context K together with the partial order
::; is always a complete lattice,5 called the concept lattice of K and denoted by
Il3(K).
A possible confusion might arise from the double use of the word "concept"
in FCA and in Definition 2.1. This comes from the fact that FCA and Definition
2.1 are two models for the concept of "concept" which arose independently. In
order to distinguish both notions, the FCA concepts will always be referred to
as 'formal concepts'. The concepts in ontologies are referred to as "concept"
or as 'ontology concepts '.
There is no direct counter-part of fonnal concepts in Definition 2.1. Concepts
as defined in Definition 2.1 are best compared to FCA attributes, as both can
be considered as unary predicates on the set of objects.
m
Vl------------------~--------~+
n1
D IL--
cB=
~
~
Linguistic
°
Processing
FCA- IBp(lK) Lattice
Merge ~ Exploration -+
new
~
Linguistic
Processing
~
• Second, the documents have to cover all concepts from the source ontolo-
gies. Concepts which are not covered have to be treated manually after the
merging procedure (or the set of documents has to be expanded).
• And last but not least, the documents must separate the concepts well enough.
If two concepts which are considered as different always appear in the same
documents, FCA-MERGE will map them to the same concept in the target
ontology (unless this decision is overruled by the knowledge engineer).
When this situation appears too often, the knowledge engineer might want
to add more documents which further separate the concepts.
In the following we introduce the three core steps for deriving a common
ontology based on a set of input ontologies using FCA-MERGE. For purpose
of explanation, a small example based on two simple ontologies is taken from
the tourism domain is used. The method has been empirically evaluated in a
larger scenario, using two tourism ontologies, each containing approx. 300
concepts.
6
·rl
J.J
11)
5 '0
.r<
'"
~
~
0 M
'" ..,
'u'""' '-'
..,'"
U
rl Q) rl
c u
'0'""'
'-' 0 .r<
II
:n
> ::c '"
:>
tu
c
0
u
0
0
0::
1;.:
0
::c
U
t!
..;
III
:>
:l1:
0
0
0::
docl X X X X X docl X X X X
doc2 X X X X X doc2 X X X
doc3 X X X X doc) X X X X
doc4 X X X X X doc4 X X X X
docS X X X docS X X
doc6 X X X X doc6 X X X X
doc7 X X doc7 X X X
docS X X X X X doc8 X X X X
doc9 X X X X doc9 X X X
doclO X X X X doclO X X X
docll X X X X X docll X X X X
doc12 X X doc12 X X X
doc13 X X X X doc13 X X X X
doc14 X X X X doc14 X X X
Hotel_l
Hotel_2
Accommodation_2 Event 1
concert_l
sical_2
o
Figure 5.5. The Pruned Concept Lattice
The computation of the pruned concept lattice is done with the algorithm
TITANIC 8 . However for the specific task described here, it is modified and
adopted to allow the pruning of a derived concept lattice. Compared to other
algorithms for computing concept lattices, TITANIC has - for the purpose -
the advantage that it computes the formal concepts via their so-called key sets
(or minimal generators). A key set is a minimal description of a formal concept.
We refer the reader to (Stumme et aI., 2000) where a detailed introduction of
the algorithm is given. In this application of the algorithm, key sets serve two
purposes. First, they indicate if the generated formal concept gives rise to a
new concept in the target ontology or not. A concept is new if and only if it has
no key sets of cardinality one. Second, the key sets of cardinality two or more
can be used as generic names for new concepts and they indicate the arity of
new relations.
The result from the last step is a pruned concept lattice. From it the target
ontology has to be derived. Each of the formal concepts of the pruned concept
lattice is a candidate for a concept, a relation, or a new subsumption in the target
ontology. There is a number of views which may be used to focus on the most
relevant parts of the pruned concept lattice. These views are discussed after the
description of the general strategy - which follows now. Of course, most of
the technical details are hidden from the user.
The documents are not needed for the generation of the target ontology.
Therefore, the attention is restricted to the intents ofthe formal concepts, which
are sets of (ontology) concepts of the source ontologies. For each formal concept
of the pruned concept lattice, the related key sets are analyzed. For each formal
concept, the following cases can be distinguished:
1 It has exactly one key set of cardinality 1.
2 It has two or more key sets of cardinality 1.
3 It has no key sets of cardinality 0 or 1.
4 It has the empty set as key set. 9
The generation of the target ontology starts with all concepts being in one of the
two first situations. The first case is the easiest: The formal concept is generated
by exactly one ontology concept from one of the source ontologies. It can be
included in the target ontology without interaction of the knowledge engineer.
In the example, these are the two formal concepts labeled by VACATION_I and
by EVENT_I.
In the second case, two or more concepts of the source ontologies generate
the same formal concept. This indicates that the concepts should be merged
into one concept in the target ontology. The user is asked which of the names
to retain. In the example, this is the case for two formal concepts: The key
sets {CONCERT _I} and {MUSICAL2} generate the same formal concept, and
are thus suggested to be merged. The key sets {HOTELI}, {HOTEL2}, and
{ACCOMMODATION_2} also generate the same formal concept.1° The latter
case is interesting, since it includes two concepts of the same ontology. This
means the set of documents does not provide enough details to separate these
two concepts. Either the knowledge engineer decides to merge the concepts (for
instance because he observes that the distinction is of no importance in the target
application), or he adds them as separate concepts to the target ontology. If there
are too many suggestions to merge concepts which should be distinguished, this
is an indication that the set of documents was not large enough. In such a case,
the user might want to re-Iaunch FCA - MERGE with a larger set of documents.
When all formal concepts in the first two cases are dealt with, then all concepts
from the source ontologies are included in the target ontology. Now, all relations
from the two source ontologies are copied into the target ontology. Possible
conflicts and duplicates have to be resolved by the ontology engineer.
Data Import & Processinl{ 93
In the next step, it is dealt with all formal concepts covered by the third case.
They are all generated by at least two concepts from the source ontologies, and
are candidates for new ontology concepts or relations in the target ontology.
The decision whether to add a concept or a relation to the target ontology (or
to discard the suggestion) is a modeling decision, and is left to the user. The
key sets provide suggestions either for the name of the new concept, or for the
concepts which should be linked with the new relation. Only those key sets
with minimal cardinality are considered, as they provide the shortest names for
new concepts and minimal arities for new relations, respectively.
Example. For instance, the formal concept in the middle of Figure 5.5 has
{HOTEL2, EVENT_l}, {HOTEL_I, EVENT_l}, and
{ACCOMMODATION_2, EVENT _I} as key sets. The user can now decide if she
wants to create a new concept with the default name HOTELEvENT (which
is unlikely in this situation), or to create a new relation with arity (HOTEL,
EVENT), e. g., the relation ORGANIZESEvENT.
There is exactly one formal concept in the fourth case (as the empty set is
always a key set). This formal concept gives rise to a new largest concept in
the target ontology, the ROOT concept. It is up to the knowledge engineer to
accept or to reject this concept. Many ontology tools require the existence of
such a largest concept. In the example, this is the formal concept labeled by
ROOT _1 and ROOT _2.
Finally, the taxonomic order on the concepts of the target ontology can be
derived automatically from the pruned concept lattice: If the concepts CI and
C2 are derived from the formal concepts (AI,B I ) and (A 2 ,B2 ), resp., then
He (CI , C2) if and only if B I :2 B2 (or if the user explicitly modeled it based on
a key set of cardinality 2).
out of a single ontology might be subject to merge. The user might either
conclude that some of these concept pairs can be merged because their dif-
ferentiation is not necessary in the target application; or he might decide that
the set of documents must be extended because it does not differentiate the
concepts enough. In the small example, the list for 0 1 contains only the pair
(HOTEL_I, ACCOMMODATION_I). In the larger scenario that has been car-
ried out, additional interesting pairs like (RAUMLICHES, GEBIET) and (AuTo,
FORTBEWEGUNGSMITTEL) are have been introduced. For the target applica-
tion, RAUMLICHES [spatial thing] and GEBIET [region] have been merged, but
not AUTo [car] and FORTBEWEGUNGSMITTEL [means of travel].
The number of suggestions provided for the third situation can be quite high.
There are three views which present only the most significant formal concepts.
These views can also be combined:
• First, one can fix an upper bound for the cardinality of the key sets. The
lower the bound is, the fewer new concepts are presented. A typical value is
2, which allows the retention of all concepts from the two source ontologies
(as they are generated by key sets of cardinality 1), and to discover new
binary relations between concepts from the different source ontologies, but
no relations of higher arity. If one is interested in having exactly the old
concepts and relations in the target ontology, and no suggestions for new
concepts and relations, then the upper bound for the key set size is set to 1.
• Second, one can fix a minimum support. This prunes all formal concepts
where the cardinality of the extent is too low (compared to the overall number
of documents). The default is no pruning, i. e., with a minimum support of
o%. It is also possible to fix different minimum supports for different
cardinalities of the key sets. The typical case is to set the minimum support
to 0 % for key sets of cardinality 1, and to a higher percentage for key sets
of higher cardinality. This way all concepts are retained from the source
ontologies, and generate new concepts and relations only if they have a
certain (statistical) significance.
• Third, one can consider only those key sets of cardinality 2 in which the
two concepts come from one ontology each. This way, only those formal
concepts are presented which give rise to concepts or relations linking the
two source ontologies. This restriction is useful whenever the quality of
each source ontolology per se is known to be high, i. e., when there is no
need to extend each of the source ontologies alone.
In the small example, there are no key sets with cardinality 3 or higher.
The three key sets with cardinality 2 (as given above) all have a support of
~~ ~ 78.6 %. In the larger application, 2 has been fixed as upper bound
for the cardinality of the key sets. Key sets like (TELEFON_l [telephone],
Data Import & Processing 95
The experiences have shown that different selection strategies of specific learn-
ing corpora heavily influence the final, target ontology. All in all, the reader
may note that for ontology learning from web documents "intelligent support"
for the definition of a representative learning corpus D is required. Having this
target in mind, an ontology-focused document crawler has been developed
In general, a crawler is a program that retrieves Web pages, commonly used
by a search engine (Pinkerton, 1994) or a Web cache. Roughly, a crawler starts
off with the URL for an initial page Po. It retrieves Po, extracts any URLs in it,
and adds them to a queue of URLs to be scanned. Then the crawler gets URLs
from the queue (in some order), and repeats the process. Every page that is
scanned is given to a client that saves the pages, creates an index for the pages,
or summarizes or analyzes the content of the pages. With the rapid growth of
the world-wide web new challenges for general-purpose crawlers are given (cf.
recent work done by (Chakrabarti et al., 1999».
The crawler builds on the general crawling mechanism described above and
extends it by using ontological background knowledge to focus the search in the
web space. Therefore, it supports the configuration of a learning corpus V. It
takes as input a user-given set A of seed documents (in the form ofURLs), a core
ontology 0, a maximum depth level dmax to crawl and a minimal document
relevance value r min. The resulting output of the crawling process is a focused
learning corpus V.
The algorithm. The crawler downloads each document contained in the set
A of start documents. Each document is analyzed using the same extraction
mechanism as used in FCA-MERGE. Based on the results of the extraction
mechanisms for each document a relevancy measure r(d) is computed. In its
current implementation this relevancy measure is equal to the overall number
of concepts referenced in the document, defined as follows:
(5.2)
Ifthe relevancy r(d) exceeds the user defined threshold rmin, the specific
document is added to the learning corpus Vll. All hyperlinks starting from
a document d are recursively analyzed. If the crawling process for a given
d does not automatically stop, the crawling process is additionally restricted
with a maximum depth level dmax for a given start document d. A detailed
description of the focused crawling approach and its evaluation areprovided in
(Ehrig, 2001).
Data Import & Processing 97
Figure 5.6 depicts the overall architecture of the natural language processing
component. As seen above in Figure 5.6 the architecture of the NLP framework
may be decomposed into four main components: (1) a linguistic knowledge pool
consisting of a large lexical database and finite state grammers, (2) a conceptual
knowledge module with access to the ontology and the associated domain-
specific lexicon according to Definition 2.1, (3) a shallow text processing engine
comprising different models for parsing at the lexical and clause level and (4)
a so-called text chart, a common data structure for navigation and storage of
results.
The following subsections provide an overview of the core parsing technol-
ogy, the linguistic knowledge pool, shallow text processing strategies at the
lexical and the clause level and heuristic processing strategies. A more detailed
description of aspects 0), (3) and (4) is given in (Neumann et aI., 1997; Pisko-
rski and Neumann, 2000).
grammers are available for shallow text processing at the lexical and sentence
level.
• Tokenizer: Its main task is to scan the text in order to identify bound-
aries of words and complex expressions like "$20.00" or "Mecklenburg-
Vorpommern,,13, and to expand abbreviations.
• POS Filter: The output returned from the morphological analysis com-
prises the word form together with all its readings. Considering words in
isolation - as is usually done in lexical analysis - , each word is poten-
tially ambiguous. In order to decrease the set of possible candidates for the
following components, local and very efficient disambiguation strategies
are applied such that implausible readings are filtered out. This is usually
performed through part-of-speech taggers as well as through the applica-
tion of case-sensitive rules l4 . The final task of a part-of-speech tagger is to
determine the unique part-of-speech of a current word in its current context
using local rules (see (Brill, 1993)).
. :V)))
{"die"
("d-detll
«(:TENSE . :NO) (:FORM . :NO) (:PERSON . 3) (:GENDER . :M)
(:NUMBER. :P) (:CASE . :NOM))
:DEF))
)))) )
«:TYPE • :NAME-NP)
(:SEM (:NAME • "Daimler Benz ") (:COMP-FORM • "AG"»)
:7 bieten
~bj
M6glichkeiten
Wir adV ~-attr
von
die "-!:obj
Kutschfahrten
~-.attr
v
In
obj
Wittenbeck
In the implementation of SMES that has been used in this work, mechanisms
for the recognition of grammatical functions (subject, object) such as depicted
in Figure 5.8 based on the dependency structures from previous steps were not
available. Recent developments for the recognition of grammatical functions
are described in (Piskorski and Neumann, 2000).
• The title heuristic combines the lexical entries between the starting and
ending HTML-title tags with those from the rest of the document.
• The table heuristic combines lexical entries found in HTML tables; here
the identification of table cells used in the same manner as the identification
of sentence boundaries.
Example. This examples refers to a case study that has been carried out in
the finance domain within the GETESS project22 . The overall task has been
the generation of a finance ontology supported by means of ontology learning
using different kinds of given input data. One important kind of input data
was a dictionary describing finance terms by natural language definitions. The
dictionary is freely available -on the web in HTML format 23 . Thus, for each
dictionary entry one HTML document served as input for generating our nor-
malized, RDF-based representation using the document wrapper. An example
of this representation is given in Figure 2.3.
<rdf:RDF
xmlns:rdf = ''http://www.w3.org/1999/02/22-rdf-syntax-ns#''
xmlns: diet = ''http://ontoserver.aifb.uni-karlsruhe . de/schema/diet. rdfs#">
In general one may apply ontology learning techniques on the lexical level, if
no ontological background knowledge is available, or at the conceptual level, if
some kind of ontological background knowledge is available. In the following
a number of relevant input relations on both levels is provided.
Lexical entry-lexical entry relation. This first relation that is presented here
is a relation restricted to the lexical level without having available a set of
108 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Table 5.2 depicts an example lexical entry-lexical entry relation. The reader
may note that for example the lexical entries "accomodation" and "hotel" show a
similar behaviour with respect to the lexical entries "accomodation", "hotel", and
"event".
Data Import & Processing 109
We mainly use the transaction view for deriving candidate conceptual re-
lations between lexical entries using the algorithm that will be introduced in
chapter 6, subsection 1.3.
The document space based on lexical entries has been introduced in Defini-
tion 5.5. The document space may be analogously defined based on a given set
of concepts C as follows:
:l;-~''l;:,:i:I~<:<-S<~
? :·-;-'::,:i:~*,m~*,,).
'* ~:-'l::;:;~~m~m
*':-:;C;'X~~~;
~:;~'~;:;~~'}
\;~:i::'~~ffi
~<: "_:i::,*';:;-'%i,~~
?~$_'l::::;:'~<1':1
~~$'l::':::'~'::Ji<
~~'~::>:-$'~:)«
relation and the result is depicted in Figure 5.12. On the left side of Figure 5.12
the user may select concepts along the documents should be described. On the
right side the resulting relation for a concrete set of documents is depicted.
The last relation that is defined here is the concept transaction relation ret.
The relation has been already introduced on the lexical level. It represents
binary pairs of concepts and is defined as follows:
ID ITEMA ITEMB
PERSON HOTEL
2 AREA EVENT
3 CITY EVENT
3. Conclusion
This chapter introduced the mechanisms for data import and processing de-
veloped in this book. The range of considered data was restricted to free natu-
rallanguage and semi-structured documents (e.g. in the form of dictionaries)
and ontologies. For importing and processing existing ontologies an ontology
wrapper approach that supports importing single ontologies into the system has
been introduced. If two or more ontologies are available, the data-oriented,
bottom-up ontology merging method FCA-MERGE may be used to derive a
common ontology. Natural language documents are linguistically annotated
using shallow linguistic processing with a mapping to the ontology if avail-
able. Semi-structured documents in the form of dictionaries are wrapped into
a specific format that may be further processed.
It has been seen that the task of preprocessing data requires complex and
difficult mechanisms for transforming data into different forms with varying
complexities. The reader may note that the quality of the preprocessing step
directly influences the quality of the results that may be generated by the algo-
rithms that are provided in the subsequent chapter.
Concluding this chapter three main issues are defined that have to be ap-
proached in the future for further improving data processing and importing:
natural language processing techniques, structure-aware document processing,
and the usage of multi-relational data for ontology learning.
These phrases should be analyzed as the "hotel" is linked to the "luxury rooms"
as well as to the "splendid bar".
In the future approaches that combine structure analysis with natural lan-
guage processing have to be researched. The importance of including struc-
tural features into the text extraction process has been recently recognized (e.g.,
relevant work in the same direction has been done by (Wang and Liu, 1998)).
Notes
20In its most generic sense a wrapper is a piece of software that accompa-
nies resources or other software for the purpose of improving convenience,
compatibility, or security.
21 The namespace is online available at
http://ontoserver.aifb.uni-karlsruhe.de/schemaldict.rdfs.
22 http://www.getess.de
23 http://www.boersenlexikon.de/
24according to (Abiteboul et aI., 1994)
25 http://www.w3c.orglMarkUp/
III
This chapter presents the ontology learning algorithms developed and used
in the context of the ontology learning framework. According to the phases
of the ontology learning cycle described in chapter 4 a bundle of algorithms
is presented that support ontology extraction and maintenance. An important
aspect of all algorithms is that they support the idea of an ''incremental growing
ontology structure". In the last chapter it has been introduced how existing
ontologies may be imported and used within our framework as background
knowledge. The reader may note that all algorithms presented also work without
any given conceptual structures in the form of a baseline ontology. However, if
some kind of conceptual structures are available, the algorithms profit from the
existing background knowledge (e.g. in the form of already existing conceptual
structures such as a concept hierarchy) for the generation of further conceptual
structures building on the existing ones.
This chapter is introduced with section 1, that presents the mechanisms for
extracting ontologies according to the ontology structure (J. The section is
separated along the ontology elements, namely lexical entries referring to con-
cepts and relations, taxonomic relations in the form of concept hierarchies and
non-taxonomic relations between pairs of concepts. For each of these elements
different algorithms are offered that operate on different kinds of input data,
such as free natural language texts and structured dictionaries as introduced in
the last chapter. The techniques applied within the framework may be roughly
separated into (i) statistical or machine learning-oriented algorithms and (ii)
pattern-matching based techniques. Within the work described here existing
A. Maedche, Ontology Learning for the Semantic Web
© Kluwer Academic Publishers 2002
118 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
different types of input data. Related work exists in the "infonnation retrieval"
and "tenninology" community, where one mainly talks about terms. We refer
the interested reader to a recently published book by (J aquemin, 2001) that gives
a comprehensive overview on the topic of "spotting and discovering tenns"
and to the related work chapter contained in this book. Before the different
frequency measures and algorithms for extracting lexical entries are introduced,
a short clarification is provided about the difference between words and tenns
and their relationship to lexical entries and concepts, respectively.
• A word is an arbitrary entity that has on one side a concept and on another,
a fonn (see (de Saussure, 1916)).
• A word is a single unit of language which has meaning and can be spoken
or written (see Cambridge English Dictionaryl).
• A word is a sound or sign that, if written down, corresponds to a string
of letters that has spaces or punctuation marks on either side (see (Miller,
1996)).
• A word is a syntactic atom, something that can be a member of a category
(noun, verb) and that can be a product of a morphological rule (see (Bloom,
2000)).
It becomes obvious there exist at least two opinions about what a word con-
tains: On the one hand, a word is considered as something that has a meaning,
on the other hand it is considered as a pure syntactic unit. As mentioned above
in several communities one often talks about tenns instead of words, e.g. as
given as follows:
A specific kind of tenns are the so-called multi-word terms 2 that are espe-
cially important for identifying concepts.
Finally, the decision is reached there is neither a clear definition of what a
word is nor of what a tenn is. In general one may say that an important aspect
of tenns is they are more specialized than words of the everyday language. In
120 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
(Frantzi et aI., 2000) it is argued that terms are mainly noun phrases with an
average length of two words. Therefore, no further distinction between terms
and words will be made in the following and so-called lexical entry are used
as a general description of both "words" and "terms".
The reader may note that in general dfl ~ efl and L:d lefl,d = cfl . The
extraction of lexical entries is based on the information retrieval measure tfidf
(term frequency inverted document frequency). Tfidf combines the intro-
duced quantities as follows:
DEFINITION 6.1 Let leid l be the lexical entry frequency of the lexical entry 1
in a document d. Let dft b~ the overall document frequency of lexical entry 1.
Ontology Learning Algorithms 121
Then tfidfi ,d of the lexical entry l for the document d is given by:
Tfidf weighs the frequency of a lexical entry in a document with a factor that
discounts its importance when it appears in almost all documents. Therefore
terms that appear too rarely or too frequently are ranked lower than terms that
hold the balance. A final step that has to be carried out on the computed tfidf1,d is
the following: A list of all lexical entries contained in one of the documents from
the corpus V without lexical entries that appear in a standard list of stopwords
is produced. 3 The tfidf values for lexical entries l are computed as follows:
DEFINITION 6.2
The user may define a threshold k E R+ that tfidfl has to exceed. The
lexical entry approach has been evaluated in detail (e.g. for varying k and
different selection strategies). A detailed description of the evaluation results
is given in subsection 4.2.
for i:=l to n do
k i := Xi·
end for
K:= {k l , ... ,kn }.
j := n + l.
while IKI > 1 do
(knl' kn2 ) :=arg max(ku,kv)EKxKsim(ku , kv)·
k j = knl U k n2 .
K := K\{k nl , k n2 } U {k j }.
j := j + 1
end while
containing all objects has been formed. The most important aspect in clus-
tering is the selection of an appropriate computation strategy and a similarity
measure. We will introduce a number of computation strategies (e.g. single-
link, complete link or group-average) and similarity measures (e.g. cosine,
kullback-Ieibler) later in this subsection.
Algorithm 2 (adopted from (Manning and Schuetze, 1999» roughly de-
scribes the top-down algorithm. It starts out with one cluster that contains all
objects. The algorithm then selects the least coherent cluster in each itera-
tion and splits it. Clusters with similar objects are more coherent than clusters
with dissimilar objects. Thus, the strategies single-link, complete link and
group-average can also serve as measures of cluster coherence (function coh)
in top-down clustering.
The reader may note that splitting a cluster (function split) is also a clustering
task (namely the task of finding two sub-clusters of a cluster). Thus, there is
a recursive need for a second clustering algorithm. Any clustering algorithm
may be used for the splitting operation, including bottom-up algorithms.
_ ;7\ :ExEX,YEY xy
cos ( x, Y J = --;========= (6.3)
J:ExEX x 2 :EyEY y2
U sing the cosine measure it is computed how well the occurrence of a specific
lexical entry correlates in x and y and then divided by the Euclidean length of
the two vectors to scale for the magnitude of the individual length of x and y.
Though, the following measure is not a metric in the strong sense, it has been
quite successfully applied in statistical NLP. The kullback leibler divergence
has its roots in information theory and is defined as follows:
DEFINITION 6.4 For two probability mass functions p(x), q(x) their relative
entropy is computed by
~ p(x)
D(pllq) = ~ p(x)log (x) (6.4)
xEX q
126 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Using the cosine measure one may compute the similarity between the con-
cepts HOTEL and ACCOMODATION as follows. The vector of the concept
HOTEL is given by iT = (0,14,7,4,6), the vector of the concept
ACCOMODATION is given by if = (14,0,11,2,5).
_ _)
cos (X,Y
+4.2+6.5
= 7 . 11 101 ~
0 93
. (6.5)
. 150
For computing the kullback leibler divergence one has first calculate the prob-
ability mass functions for each concept and its corresponding frequencies. The
probability mass functions for HOTEL are given as (0,0.45,0.22,0.13,0.19)
the probability mass functions for the concept ACCOMODATION are given as
(0.44,0,0.34,0.06,0.16)
Based on these values one can compute the kullback leibler divergence as
follows
We refer the reader to (Manning and Schuetze, 1999) where a detailed in-
troduction into further similarity measures between two sets X and Y such as
the matching coefficient X n Y, the dice coefficient ~1~~11' the Jaccard or
.lammoto
'T' •
Clent IxnYI
coefIi' XuY or the overI ap coefIi Clent
. IxnYI··
minC!XI,IYI) IS gIven.
accomodation 1 eM='«'<
(hote' accQmodatioR, res urant. beer garden) $""
$ t(~¢o/~w
e ~,:~ e;,y:;..~:,.:
area
aeco odation
service company :=
(restaurant, beer garden)
organization: =
hotel aoooR>oda'ion restaurant beer garden city .......... (accomodation, service company)
Figure 6.1 depicts an example scenario. On the left side the hierarchical
clustering result, on the right upper side the existing background know ledge
and on the right lower side the new, manually added ontological structures
based on interpreting the clustering result are depicted. In this simple example
three nodes (ACCOMODATION, AREA, ROOT) of the clustering tree could
be labeled using the existing background knowledge. The labeling strategy
introduced above has the advantage that it narrows down the length of presented
node names (that are typically conjunctions of lexical identifiers representing
concepts). In this small example two new concepts (SERVICE COMPANY,
128 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Result
H C (C21 Cd, where
C 1 = F("Electronic service") and
C2 = F("Automatic Debit Transfer").
may further encourage the sale of these items together within single visits to the
store). In (Agrawal et al., 1993) concrete examples for extracted associations
between items are given. The examples are based supermarket products that
are included in a set of transactions collected from customers' purchases. One
of the classical association rule that has been extracted from these databases is
that "diaper are purchased together with beer".
The reader may now ask the question what these kinds of applications de-
scribed above have in common with the ontology learning task of extracting non-
taxonomic relations between concepts. The underlying idea is that an adopted
association rule algorithm analyzes statistical information about the linguistic
output such as given in Definition 5.10. For instance, the linguistic processing
may find out the lexical entries "hotel", "guest house", and "youth hostel" all co-
occur with the lexical entries "costs" in sentences like
"Costs at the youth hostel amount to $20 per night." several times in the texts. Thus,
statistical data is derived indicating co-occurrences of the corresponding con-
cepts HOTEL, GUEST HOUSE, and YOUTH HOSTEL with COSTS. In this
example, the relation between HOTEL and COSTS may be proposed to the
engineer for inclusion in the ontology.
(6.7)
(6.8)
The basic association rule extraction approach may be transferred to the on-
tology learning setting with only slight modifications. The major modification
that needs to be taken care of is the input transaction data fed into the learning
algorithm. A baseline approach has been chosen that considers each linguistic
pair to constitute a transaction on its own. The most appropriate choice for
compiling transaction rules from sets of concept pairs found by linguistic pro-
cessing may depend on the actual set of domain texts. Let's consider a particular
scenario to explain why this is the case. Let's assume a set of 100 texts each de-
scribing a particular hotel in detail. Each hotel has an address, but it also has an
elaborate description of the different types of public and private rooms and their
equipment resulting in 10,000 concept pairs returned from linguistic processing.
A baseline choice for transactions considers each concept pair as a transaction.
Then support for the rule H OTEL-t ADDRESS is equal or, much more probable,
(far) less than 1%. while rules about rooms and their equipment or their style,
like ROOM-tBED, might achieve ratings of several percentage points. This
means that an important relationship between HOTEL and ADDRESS might get
lost among other conceptual relationships. In contrast, if one considers com-
plete texts to constitute transactions, an ideal linguistic processor will lead to
more balanced support measures for HOTEL-t ADDRESS and ROOM-tBED
of 100% each. In other scenarios it might be useful to constitute transactions
from the results of processing paragraphs of texts, e.g. when hotel descriptions
are listed within text documents, or to combine results from several texts, e.g.
when each hotel description is spread over several web pages.
Another important aspect is allowing input transactions such as
{PERSON, PERSON}, because for the ontology learning task one may derive
associations such as PERSON -t PERSON that may result in a non-taxonomic
relation like COOPERATEWITH(PERSON,PERSON). To realize the possibility of
extracting these kind of non-taxonomic relations the concept PERSON is splitted
into two "artificial concepts" (like PERSON-l and PERSON-2). The concepts
are later merged to PERSON, and thus not presented to the user.
The general underlying idea of computing association rules may be separated
into the following two steps:
Ontology Learning Algorithms 133
1 Compute all item sets that fulfill a user defined minimum support Smin.
These item sets are called frequent item sets.
2 Derive from these frequent item sets all associations rules that fulfill a user
given minimum confidence Cmin.
The basic algorithm for extracting frequent item sets from the preprocessed
data is given in Algorithm 3. As mentioned earlier either lexical entries or
concepts are considered as input to the algorithm. Thus, in general the term
"item" is used as a generalization of lexical entries and concepts within the
algorithm.
The algorithm is based on the idea that for item set X' ~ X it holds that
support(X') ::; support(X) (see step Hn+l ... ). Thus, it iteratively computes
frequent item sets h, 12 , ....
Finally, for generating association rules from the computed frequent item
sets one typically uses the following characteristic property of frequent item
sets: Let X be a frequent item set, than it holds for all X' ~ X
support (X)
confidence((X - X') -+ X') = (6.9)
support(X - X')
Thus, the underlying idea is that the confidence may be computed from the
support of the frequent item sets. An algorithm for deriving association rules
134 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
from frequent item sets based on the characteristic property introduced above
is given in (Agrawal et aI., 1993).
• Rules at lower levels may not have minimum support. Thus, if items are
organized in a taxonomy may interesting rules that would otherwise not be
found, due to the fine leaf-level granularity, can be detected.
the ontology. The usage of background know ledge supports the generation of
more general non-taxonomic relations that are inherited to the sub-concepts,
e.g. the relation HASCOST holds between ACCOMOCATION and COSTS, but
also between sub-concepts of ACCOMODATION as given above.
LJ ail! L2 ai,2
"Mecklenburgs" AREA "hotel" HOTEL
"hairdresser" HAIRDRESSER "hotel" HOTEL
"balconies" BALCONY "access" ACCESS
"room" ROOM "TV" TELEVISION
provider for tourist information has been processed 9 . The corpus describes
actual objects referring to locations, accomodations, furnishings of accomo-
dations, administrative information, or cultural events, such as given in the
following example sentences.
The algorithm for learning generalized association rules uses the concept
hierarchy, an excerpt of which is depicted in Figure 6.4, and the concept pairs
from above (among many other concept pairs). In our actual experiments, it dis-
Ontology Learning Algorithms l37
(Accomodation,Room) hasRoom(Accomodation,Room)
hasHoteIRoom(Hotel, Room)
And an axiom:
(Accomodation,DoubleRoom)
subRelation(hasHoteIRoom,
hasRoom)
F-Logic:
(Hotel,DoubleRoom) FORALL x,y hasRoom(x,Y) <-
hasHoteIRoom(x,y).
and for each concept a domain frequency is computed (via the mapping from
concrete lexicalizations of concepts contained in texts to concepts). Addition-
ally, frequencies are propagated to super-concepts. Thus, if a concept occurs
frequently in a domain-specific corpus, it and its super-concepts remain in the
ontology (see Algorithm 5 for a detailed description).
The Relative Pruning Algorithm. The underlying ideas of the relative prun-
ing algorithm are that domain relevance is considered as the relative frequency
of given ontological entities with respect to frequencies obtained from a general
corpus. The pruning algorithm is given in pseudo code in Algorithm 5. It takes
as input an ontology 0, a domain-specific corpus D, a general corpus G, a ratio
r E R+ and the selected frequency computation technique m.
The general idea of relative pruning is that first for all lexical entries frequen-
cies values are computed using a measure selected by the user (e.g. lef or tfidf).
Then, both frequencies values obtained in the two corpora are compared. All
existing concepts and relations that are more frequent 14 in the domain-specific
corpus remain in the ontology. The user can also specify whether or not con-
cepts that are neither contained in the domain-specific nor in the generic corpus
should be pruned from the ontology.
When pruning the mapping information in the lexicon cfJ has also to be
updated. In general we offer two ways of interaction. First the user has to decide
what happens with the lexical entries. Second, we offer an automatic strategy:
In order to minimize the loss of references in F, references are migrated the
closest super-concept of C E C that remains in the ontology. If mUltiple super-
concepts in distinct paths remain, the stem reference is deleted, because there
is no possibility for automatically selecting the correct super-concept. If, for
example, "chair" is pruned from the ontology, the lexical reference to "furniture"
might be updated to the closest super-concept of "chair" that remains in the
ontology. Using this strategy we obtain an underspecified semantics for lexical
entries (similar to the approach described in (Buitelaar, 1998»
142 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Example. A small example may illustrate the idea. For example looking at the
unknown lexical entry "weekend excursion", one may say that it shares the same
behaviour as a concept like EXCURSION. Sharing conceptual behaviour may be
computed using measures of distributional similarity based on concept vectors
as introduced in subsection 1.2.1. According to Definition 5.8 introduced in
chapter 5 the data for this learning task is represented in a concept/lexical entry-
concept matrix as given in Figure 6.4.
The counts for dependency between a lexical entry or concept and a concept
entered in the relation are computed using the different linguistic and heuristic
indicators introduced in chapter 5.
would mean, that if a lexical entry or a concept is coupled with e.g. HOTEL it
is automatically coupled with the concept ACCOMODATION, on account of the
taxonomic relation He (HOTEL, ACCOMODATION).
3. Conclusion
This section has presented a number of algorithms that have been adopted,
designed, and implemented for the ontology learning algorithm library. The
algorithms have been separated into algorithms that support ontology extraction
and that support ontology maintenance.
The algorithms build on the shallow linguistic processing and the relational
transformations introduced in chapter 5. An important aspect of all algorithms is
that they are based on the idea of an "incrementally growing ontology structure".
Thus, on the one hand the algorithms support developing an ontology from
scratch (e.g. by lexical entry extraction, hierarchical ordering of lexical entries
and the extraction of associations between lexical entries), on the other hand
the algorithms profit from existing conceptual background knowledge, e.g. in
the form of the lexical entry to concept mapping or a concept hierarchy.
The following is a list of open issues that have only been partially explored in
this work but are of highest relevance for future work. The attention is restricted
to three important points, mainly multi-strategy learning, the relation separation
problem and the learning and maintenance of axioms A o .
Ontology Learning Algorithms 145
Notes
1 The dictionary is online available at http://dictionary.cambridge.org.
2 A comprehensive overview on multi-term recognition and extraction is given
in the EAGLES-96 report on
http://www.ilc.pLcnr.itlEAGLES96/rep2/node38.html.
3 Available for German at
http://www.loria.frrbonhomme/sw/stopword.de
4 Hierarchical clustering has in the average quadratic time and space complexity.
5 A comprehensive survey on applying clustering in NLP is also available in
the EAGLES report, see
http://www.ilc.pi.cnr.itlEAGLES96/rep2/node37.htm
6 Parts of this work has been published in (Kietz et aI., 2000b; Maedche and
Staab,2000b)
7 This means that all nominal phrases N Pi contained in a definition of a dic-
tionary entry that are followed by a comma or by an and lor may be repre-
sent lexical entries that refer to super classes of the given dictionary entry,
respectively.
8 The work described in this subsection has been presented in (Maedche and
Staab, 2000c; Maedche and Staab, 2000a)
9 A detailed description of the text corpus and the overall application and
evaluation study is provided in section 4.
10The function Clos(H C , C) retrieves the set of concepts that are super- or sub-
concepts of a given C including the transitive closure based on the taxonomy
H C.
11 http://www.kun.nUcelex/
12http://www.taz.de/
13 The function UClos retrieves the set of concepts that are superconcepts of
C including transitivity based on H C .
14 The factor r can be provided by the user.
Chapter 7
Following the point of view that "the proof is in the pudding", the method-
ological and theoretical research results of this book have been implemented in
the comprehensive TEXT- TO-ONTO ontology learning environment. TEXT-
TO-ONTO is the "running result" of the research described in this bookl.
TEXT- TO-ONTO is roughly separated into two main, historically caused parts:
In part I of this book the fundamentals of ontologies, ontology engineering and
ontology-based applications were introduced. According to these fundamen-
tals an environment for manually engineering and managing ontologies, the
ONTOEDIT ontology engineering environment, has been designed.
In part II an architecture, components and concrete mechanisms for support-
ing ontology learning for the Semantic Web have been provided. Inspired by
the idea of balanced cooperative modeling ONTOEDIT has been extended with
data import and processing techniques and algorithms that support the extrac-
tion and maintenance of ontological structures. Soon it became obvious that
for using ontology learning as a "plug-in" in an ontology engineering environ-
ment like ONTOEDIT, the mechanisms and user interfaces for data import and
processing and algorithm execution are extremely important when using it in a
real-world scenario.
TEXT- TO-ONTO builds on the capabilities of a comprehensive ontology
management and engineering environment and provides means for semi-auto-
matic ontology extraction and maintenance building on different kinds of input
data. This chapter describes TEXT- TO-ONTO, the ontology learning environ-
A. Maedche, Ontology Learning for the Semantic Web
© Kluwer Academic Publishers 2002
152 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
The main underlying idea of the overall approach is that each modeling step
may be done by the user or by an ontology learning algorithm. Thus, section 3
explains the functionality of the TEXT- TO-ONTO ontology learning environ-
ment building on ONTOEDIT. This section is started by explaining the main
features of the data import and processing component that may be accessed by
the management component of TEXT- TO-ONTO. Subsequently, some exam-
ples are shown how the algorithms for ontology extraction and maintenance
contained in the library may be accessed by the user. An important aspect is
the adequate presentation of extraction and maintenance suggestions of an al-
gorithm. The different views available and implemented in TEXT- TO-ONTO
for result presentation are explained.
1. Component-based Architecture
The main components of the TEXT- TO-ONTO environment and the inter-
actions between them are depicted in Figure 7.1. The reader may note the
relation to the ontology learning framework and its accompanying architecture
introduced in chapter 4. As mentioned above in the implementation of the
architecture it is internally distinguished between the data structures, the pro-
cessing modules and the graphical user interfaces. The reader may note that
all components are connected with the ontology, knowledge base and lexicon
model that play an important role in the overall environment. These three mod-
els are implementations of the formal definitions introduced in chapter 2. Three
Text-To-Onto
Graphical User
Interface
OntoEdit
Graphical User
Interface
I F-Log~ I
I~O
As mentioned above this chapter will mainly focus on the user interface and
interaction component supporting learning and engineering ontologies for the
Semantic Web. The overall user interface of TEXT- TO-ONTO is depicted
in Figure 7.2. The reader may note that there is no strict separation between
manual engineeering ontologies using ONTOEDIT and the usage of ontology
learning techniques contained in TEXT- TO-ONTO. Again, this fact reflects
the paradigm of balanced cooperative modeling.
The Lexicons £,0 , £}(B. Ontology engineering typically starts with the col-
lection of lexical entries that refer to concepts, relations, and instances. Figure
7.3 depicts the view that supports the definition of lexical representations for
the core primitives C, n, I contained in the ontology 0 and the knowledge base
KB (following the Definitions 2.1 and 2.3). The reader may note that a m:n
mapping between the lexical entries and the core primitives has to be supported.
• Tags referring to natural language processing, like the part-of speech (POS)
information depicted in Figure 7.3.
fined are already morphologically reduced to their stems with respect to the
interaction with the natural language processing environment SMES.
Emp!O¥~~
$
e- !>z.d~m"Slaff
,&,:jmlr;;:5-"tr<:<'ti'ysSta:ft
$ M"n.9N
$ TethnicSlSt.IT
:;>$ student
k$prMuc1
:2·'-". Pr-;1jed
+'.-.$ Ftit:;ic~tin-n
+$ Topic
new relations with domain and range restrictions or add domain and range
restrictions to existing relations.
In Figure 7.5 a relation excerpt from the SWRC ontology that has been
developed for the Semantic Web Research Community is depicted. The con-
cept ACADEMICSTAFF has been selected in the concept hierarchy view. Re-
lations that are marked grey are inherited from super-concepts of the concept
ACADEMICSTAFF, e.g. the relation NAME has been inherited from more generic
concepts (e.g. the concept PERSON). A number of relations are in particular
attached to the concept ACADMICSTAFF, e.g the relation WORKSATPROJEC with
the range PROJEKT. In contrast to the conceptual definition of an ontology 0,
ONTOEDIT provides several pre-defined datatypes such as STRING and IN-
TEGER that reference the datatypes contained in the XML-Schema standard4 .
The ontology engineer may define meta-information to a concrete relation,
e.g. minimum and maximum cardinalities, multi-lingual documentation and
further lexicalizations.
given for instantiating two simple patterns, namely inverse relations and disjoint
concepts. The interested reader is referred to the comprehensive description of
the approach given in (Staab et aI., 200Ia).
Inverse Relations. A simple example for inverse relations has been given
in subsection 3.9 as follows: Consider there exists a relation WORKS_AT hold-
ing between the concept PERSON and the concept PROJECT, and a relation
HAS_PARTICIPANT between PROJECT and PERSON. Explicit modeling that these
two relations are inverses of each other has several advantages: (i) it ensures
consistency in the knowledge base (e.g. it is impossible to have a project with a
participant that doesn't work at that project) and (ii) it frees the user from pro-
viding redundant information. Figure 7.6 depicts the general view of modeling
"relation axioms" and some concrete inverse relation axioms for the tourism
domain. On the left side all relations with their corresponding domain / range
restrictions are listed and provided to the user. On the right side a table for
defining that two relations are inverse is provided to the user. The user may
add a relation from the left pane via pushing the "+" button. Inverse relations
may be defined globally, e.g. that IN _G EBIET and BIETET _UNTERKUNFT in general
are inverse, or, locally restricted, e.g. that IN_GEBIET(UNTERKUNFT,GEBIET)
and BIETELUNTERKUNFT (GEBIET,UNTERKUNFT) are inverses of each other.
Global and local inverse relations are distinguished by the selection of specific
domain restrictions, respectively.
Disjoint Concepts. Figure 7.7 depicts the general view for modeling so-called
"concept axioms". Disjoint axioms are specific concept axioms that enforce
consistency and quality for the ontology and the knowledge base. Using the
view for modeling disjoint concepts the ontology engineer may select concepts
The TEXT-To-ONTO Environment 159
from the left pane and explicit model their disjointness. If more than two
concepts are selected, the cartesian product of disjointness between the selected
concepts is computed and the particular disjoint concepts are represented in the
right pane of Figure 7.7.
Pe-rst)ef!i~Che:;:_Oi!1g
TeHw'Bis€-_Mat€:lie))$$_Dir:g
«t:K>J::stwerk
'J"'FG
.- LBb-e-V~~'Bsen
Person
H $I ,ie'
HG Fisch
e:--t): Saeugetlef
e$l Vogel
2",0, Organ1s.alh:m
The ontology engineer may check the consistency of the actual model con-
cerning disjoint concepts by clicking on the check button.
Ge:se-llst!,i$l'ter
On.::md:Ji)~sm:t:Jli~<1
A Haef10ler
J:: .. "A.'!e9:d~~~"
k "" ".~~2"l':2Ind2r 1i1ll*dJ::h1::'"
t: .. "+4,,··(\:.l}?11··!).{lS l):SS:S"
z .. "'CRicl:J.t~r'"
k ,. "Cot.nd:lia Ric'ht:l:l:t:"
.e x «~ 'nl $0$ 4063»
·:?hD~t.ud~nt.~~Wdent~G'-$uet~" 82546.0
&.lo1a:e:dche ,:St:..<dent.lJ.::&&~e.u:"" 626ao.O
A.."!ae<iC'~.e / Smder.t:., P1>.D$t:u.d.ent.... S2Sa7. Q
Figure 7.10. View for Querying the SILRI F-Logic Inference Engine
Using this query it is asked for all instances of the concept STUDENT with
their names and phone number. The results are given as substitutions of the
162 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
variables z, k, e. The left side of figure 7.lO depicts the explanation to the
evaluation results. It describes the rules that have been activitated for generating
the appropriate substitutions, e.g. the transitivity of subconcept relationships
(see rule "2.0"):
The dialog shown in the upper part of Figure 7 .11 allows to perform different
operations on the indexed documents. First, if necessary, HTML elements may
be eliminated and substitution rules may be applied. Second, the natural lan-
guage processing component for normalizing the documents may be accessed
from this dialogue. An important aspect is that the natural language processing
component accesses the actual ontology and its lexicon to refer the documents
to the available background knowledge. The linguistically pre-processed doc-
uments serve as input for the transformation module. An example view for
generating a document-concept relation has been given in Figure 5.12. Another
graphical component that is available is the document wrapper for importing
semi-structured documents, in particular domain-specific dictionaries.
164 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
,;1':t:AD • "~':~J
t:stlE {::S.£rf i:JliODS "'-v:n;2o!::g~""J- (:HEAD ~ "t:~e:n"l)
On the right side of Figure 7.12 the view for pattern engineering and debug-
ging is depicted. A specific selected pattern may be applied to a test document
and matching elements are presented to the user. Using this view, the user may
adjust a specific pattern to the available domain texts or to a specific dictionary.
The TEXT- TO-ONTO Environment 165
4. Conclusion
In this chapter parts of the implemented environment TEXT- TO-ONTO, the
ontology learning environment as the running result of the research described
in this book have been presented . Following the quote from the start of this
section, TEXT- TO-ONTO has been indeed successfully used for "real-world
knowledge acquisition" from existing Web data.
However, as usual in developing software environments, there is still a lot
of further work required for improving the ontology engineering and learn-
ing environment. In the following important possibilities for extending the
environment are sketched: First, as mentioned earlier the TEXT- To- 0 NTO
environment is still in an early stage with respect to providing methodologi-
cal guidelines in applying ontology learning supporting ontology engineering.
Thus, in the future the integration between a comprehensive methodology with
support of the application of semi-automatic means is required. This holds es-
pecially for the difficult tasks of data import and processing, where experiences
have to be collected and provided to the ontology engineer.
Second, at the current moment the TEXT- TO-ONTO ontology learning en-
vironment is restricted to textual data that is based on the German language.
It has already been mentioned that in general the natural language processing
component may be replaced or complemented by another NLP component, e.g.
for English. However, using different engines for processing natural language
requires clearly defined representation layers for abstracting from language
specific issues. Using TEXT- TO-ONTO in a multi-language setting would
require the development of a generic technique for layering NLP results gen-
erated by different language-specific engines.
Furthermore, a flexible plug-in architecture should be developed. Specific
applications typically demand extensions or adaptions of existing functionality
of the ontology learning and engineering environment. A plug-in architecture
is currently developed to support the location and the management ofplug-in's
that provide specific functionality. The architecture with its plug-in and service
mechanisms is described in (Handschuh et aI., 2001). The components are
dynamically plug-able to the core Ont-O-Mat component environment5 . The
plug-in mechanism notifies each installed component, when a new component
is registered. Through the service mechanism each component can discover
and utilize the services offered by another component (Handschuh, 2001). A
service represented by a component is typically a reference to an interface. This
provides among other things a decoupling of the service from the implementa-
tion and allows alternative implementations.
Finally, a server as a infrastructural kernel for Semantic Web develop-
ment and applications is required. The ontology engineering and learning
environment contains a lot of functionality that several other ontology-based
applications also require. Strictly spoken, the engineering environment de-
The TEXT-To-ONTO Environment 169
veloped within this book is only a client that access a server that acts as a
infrastructural kernel for semantic web applications: The system will be real-
ized as a component-based, extensible plug-in architecture tightly connected to
several core components: an ontology and fact repository for persistent storage
based on RDF as a basis for the other services, inference engines for offering
reasoning services, versioning, application programmer interfaces (API's) for
ontology engineering, maintenance, migration and integration and Semantic
Web applications6 . The challenge of this task is the provisioning of capable,
but open interfaces that allow the welding together of components that already
exist or are just about to be provided by the research community.
170 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Notes
1 Parts of this chapter have been published in (Maedche and Volz, 2001).
2 Concepts are defined by a unique identifier that is typically not visualized.
Instead, optionally, the possibility to visualize the lexical entries for onto-
logical elements is offered.
3 The ontology may be accessed at the namespace
http://ontobroker.semanticweb.org/swrc-30-1 O-OO.rdfs
4 see http://www.w3.orgITRlxmlschema-2/
5 http://ontobroker.semanticweb.org/annotation/ontomatl
6 Further information is available on http://kaon.semanticweb.org
Chapter 8
EVALUATION
based on structured data. The approach has shown that know ledge acquisition
with machine learning outperforms manual knowledge acquisition techniques.
Several experiences concerning the specific knowledge acquisition scenario
of ontology engineering and learning and the application of ontologies have
been collected within the projects and case studies carried out. The experiences
led to two different ontology evaluation approaches: (i) application-specific
evaluation and (ii) cross-comparing two ontologies using a gold-standard. In
the first approach ontologies are evaluated through the application in which
they are employed, e.g. for information retrieval or information extraction
using their standard measures l . The second approach follows an application-
neutral paradigm: To learn about the effects that different ontology learning
methods lead to by a given set of data, a hand-modeled ontology is compared
with ontologies that have been generated by a specific algorithm (and its spe-
cific parametrization). Here, the second approach is pursued, following an
application-neutral paradigm for evaluating the ontology learning algorithms
based on hand-modeled gold standards.
This chapter is organized as follows 2 . Section 1 introduces the general
evaluation approach that references to the foundations of chapter 3 and its
resulting definition of an ontology structure O. The approach is mainly based
on comparing a given ontology or parts of it with an extracted ontology or parts
of it, respectively. Comparison is done at two different levels, namely the lexical
level and the conceptual or semantic level. Section 2 introduces a number of
measures for comparing two ontologies or parts of them at the lexical and the
conceptual level. 3 Th~ measures are first applied in section 3. In this section
a case study for evaluating human performance using the evaluation approach
and measures are presented. The last section evaluates the ontology learning
techniques that have been introduced previously and provides methodological
guidelines for applying them in real-world scenarios. Additionally, the results
that may be generated by different ontology learning algorithms are compared
with human performance leading to the result that the combination of ontology
learning with human modeling capabilities delivers clearly better results than a
pure human modeling approach.
r-------------- ..
ll C (C ll C2 ) llC (C 1, C2 )
c R(CI, C 2 )
~ ~(C], C2 )
1 j
£:c eR
o
---------------
2
tp fp
Retrieval
fn
Figure 8.2 depicts this general setting. The overall set of positive and negative
documents is separated into so-called true positives tp, true negatives tn, false
positives f p and true negatives tn. 4 The retrieval naturally consists of true and
false positives that are related to the set of false negative documents. Based on
these counts of the number of items given in the classical definition, a definition
of precision and recall for ontology learning is adopted.
DEFINITION 8.1 (PRECISION) Precision is a measure of the proportion of
selected items the system got right:
. . tp
preClSlOn = tp+ f p
Adopted to the ontology learning task precision is given as follows:
. . IComp n Refl
preclslOnOL = ICompl
Ref is the set of elements that are given in the reference ontology. Comp
is the set of elements that are contained in the comparison ontology. Elements
Evaluation 175
in this sense are primitives according to the ontology structure 0, e.g. like the
concept lexical entries £c, concepts C, the concept hierarchy or non-taxonomic
relations between concepts.
recall = tp f
tp+ n
Recall adopted to the ontology learning task is given as follows:
II _ ICamp n Refl
reca OL - IRefl
The measures precisionoL and recalloL are inverses of each other, i.e.
precisionoL(Ref, Comp) = recalloL(Camp, Ref) and recallodRef,
Camp) = precisionoL(Camp, Ref). Hence, in the following it is only re-
ferred to precisionOL> but "both directions" will be evaluated. This means,
when cross-evaluating two ontologies it will be evaluated how precisely a sub-
ject agrees with another subject and vice versa. Since recallOL is the inverse
of precisionoL this way will also yield all agreement recall numbers.
be deceptive, when two strings resemble each other though there is no mean-
ingful relationship between them, e.g. "power" and "tower". In the case studies
performed, however, it has been found that in spite of this added "noise" SM
may be helpful for proposing good matches of lexical entries.
Averaging over all lexical entries one may thus compute a semantic simi-
larity for two given hierarchies. In addition, however, one must consider the
case where a lexical entry L" is in CC1, but not in CC 2. Then, the simplest
assumption is that the L" is simply missing from Cc 2 , but when comparing the
two hierarchies the optimistic taxonomic approximation is the one that searches
for the maximum overlap given a fictive membership of L" to CC 2. This may
be reached by solving the following maximization problem.
DEFINITION 8.9
Td'(L" (] (] ):=max{IF1 1 (SC(F({L"}),1{CJ))nF;I(SC(C),1{C 2 )1}.
, I, 2 CE C2 IFII(SC(F({L"}),1{cJ))UF21(SC(C),1{C2)1
with
ifLE.q
if L ~ C~
Consider the lexical entry "accomodation", which is only in the tax- cf,
onomic overlap is computed as follows: For the lexical entry "accomodation"
.1'1 1 (SC(.1'( {"accomodation"}), 1{C 1)) ={"youth hostel", "accomodation", "hotel"}
is computed. The concept referred to by "hotel" in C2 yields the best match re-
sulting in .1'2 1 (SC(.1'( {"hotel"}))) = {"wellness hotel", "hotel"} and
TO" ("accomodation", 1{C 1, 1{C 2) = :l-. The reader may note several properties
of TO:
DEFINITION 8.12
CM(C 0 CO).= IFl1 (UC(C 1 , HCd) n F:;I(UC(C2 , H C2))1
1, 1, 2, 2 . IFl1 (UC(C 1 , HCd) U F:;I(UC(C2 , HC 2 ))I·
F--- --person
------------------
student Jesearcher
.. ----
---------project
.---- research project
4l--
Based on upwards cotopy and concept match one may now approach the defi-
nition of the non-taxonomic relation overlap. The definition of RO is separated
in two parts: First, without considering lexical entries, RO' of two relations
R 1, R2 is defined as follows:
RO' is based on the geometric mean value of how close domain and range con-
cepts match such as given by the concept match CM6. Basically, this measure
reaches 100% when both concepts coincide (i.e., their distance in the taxonomy
He is 0); it degrades to the extent to which their distance increases; however,
this degradation is seen as relative to the extent of their agreement.
DEFINITION 8.14
182 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
An important aspect is that some of the lexical entries may only refer to
relations in R 1 , thus, this is reflected by the following definition, where one
maximizes over all relations contained in RE without considering their lexical
representation.
DEFINITION 8.15
reference
ontology
Figure 8.7 depicts the overall evaluation scenario for measuring human per-
formance. Several students generated complete ontologies or parts of it «OS)
from different starting points (see the detailed description in the following).
One ontology (Ogold) has been defined within the GETESS project using
standard knowledge acquisition techniques (e.g., questionnaire, competency
questions, etc.), serving as the gold standard for our evaluation. The "basic,
gold-standard" ontology consisted of 2,690 bilingual lexical entries, 1,087 con-
184 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Subject
o 2 3 4
310/310/310 340/310/310 474/310/310 338/310/310 300/310/310
31113111311 347/318/311 503/3141311 3111313/311 315/3111311
71171171 65/89/65 621103177 65170/82 39/69/65
Table 8.1 depicts the basic statistics that have been computed from the mod-
eled ontologies over the three pashes. Ontology 0 gold (generated by subject
0) served as gold standard and has been modeled by an expert ontology engi-
neer. It is obvious this ontology delivers identical statistics for all three phases
(because it had been modeled only once).
recall
1,00
0,80
0,60
1-2
0,40
0-1
0,20 •
0-2
;. ••0-4
2-3· ••
0,00
0,00 0,20 0,40 0,60
precision
values is approx. (0.2,0.2). Thus, the reader may note that on the simple basis
of a set of documents people do not tend to agree on on common lexical entries.
As mentioned earlier phase II is based on common lexical entries. There-
fore, better precision and recall values is expected. An example is given for
the elements contained in He. Figure 8.9 depicts the results that have been
obtained. The comparison of taxonomic relations lead to acceptable values for
both precision and recall.
recall
1.00
0.80
1-4
0,60
0,40
0,20
0.00
0,00 0,20 0,40 0.60 0.80 1,00
precision
For phase III ontologies only the computation of precision and recall values
of non-taxonomic relations makes sense. The results of the pairwise compar-
ison are given in the following table. Figure 8.10 depicts the results obtained
by comparing the non-taxonomic relation elements contained in the different
ontologies_ It can be seen that the recall values for modeling non-taxonomic
relations are very low_ Later, it will be seen how the proposed ontology learning
algorithms perform in the same task by applying the same evaluation measures_
Subject
i\j 0 2 3 4
0 1,1 0.3,0.24 0.07,0.03 0.14,0.15 0.08,0.07
1 0.24,0.3 1,1 0.25,0.13 0.22,0.3 0.07,0.07
2 0.03,0.07 0.13,0.25 1,1 0.05,0.14 0.02,0.04
3 0.15,0.14 0.3,0.22 0.14,0.05 1,1 0.15,0.12
4 0.07,0.08 0.07,0.07 0.04,0,02 0.12,0.15 1,1
recall
1,00
0,80
0,60
0,40
.1-3
.0-1
0,20
2-3. 0-3• •3-4 • 1-2
0,00 • :.0-4
As mentioned earlier, it has been experienced that the measures precision and
recall reduce the decision criteria to match ornot match and do not include nearly
perfect matches. In the following the measures proposed for the lexical and
conceptual comparison level for comparing the ontologies have been computed.
Subject
i\j 0 1 2 3 4
0 0.51,0.35 0.53,0.21 0.46,0.39 0.5,0.29
1 0.43,0.52 0.65,0.43 0.43,0.53 0.39,0.41
2 0.42,0.24 0.54,0.37 0.36,0.24 0.4,0.2
3 0.38,0.47 0.43,0.45 0.38,0.28 0.38,0.36
4 0.46,0.38 0.41,0.5 0.48,0.16 0.43,0.39
Table 8.3. SM(.ec i, .ec j), SM(.e"R. i, .e"R. j) for Phase I-Ontologies.
Results. The results for computing SM(.cC 1, .cC 2 ) of matching lexical entries
referring to concepts vary between 0.38 and 0.65 with an average of 0.45.
Comparing lexical entries referring to relations SM(.cf, .cf) results in values
188 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
L1 L2 SMeLl, L2)
"Sehenswuerdigkeit" "SehenswOrdigkeit" 0.875
[see sight] [seesight]
"Verkehrsmittel" "Luftverkehrsmittel" 0.71
[vehicle] [air vehicle]
"Zeit" "Zeit" 0.75
[tent] [time]
"AnzahLBetten" "hat...AnzahLBetten" 0.77
[numbeLbeds] [has..number _beds]
between 0.16 and 0.53 with an average of 0.36. Several typical, though not
necessarily good, pairs for which high string match values were computed are
shown in Table 8.4. RelHit(£C 1, £C 2 ) ranged between 20 to 25%, i.e. this
percentage of lexical entries referring to concepts matched exactly. For lexical
entries referring to relations the results were much worse, viz. between 10 to
15%.
Interpretation. Analyzing the figures it can be seen that human subjects have
a considerable higher agreement on lexical entries referring to concepts than on
ones referring to relations. Investigating the auxiliary measures we have found
that SM values above 0.75 in general retrieve meaningful matches - in spite of
few pitfalls, like for example when comparing "tent" and "time" (see Table 8.4).
Results. Table 8.5 presents the results that are obtained for the phase I-onto-
logies using the similarity measures taxonomy overlap (TO) and relation overlap
(RO). The reader may note these ontologies have been built without any pre-
vious assumptions about the lexica £1 and £2, thus their similarity values are
well below those of later phases where the lexica for concepts were predefined.
Table 8.6 depicts the similarity measures computed for phase II-ontologies.
Values for TO range between 0.47 and 0.87, the average TO over all 20 cross-
comparisons results in 0.56. RO yields values from 0.34 to 0.82 with an average
of 0.47.
Subject
i\j 0 2 3 4
0 0.33,0.35 0.31,0.25 0.32,0.5 0.29,0.28
0.35,0.15 0.4,0.41 0.34,0.03 0.28,0.15
2 0.28,0.12 0.36,0.25 0.25,0.04 0.24,0.15
3 0.36,0.4 0.31,0.32 0.24,0.04 0.26,0.03
4 0.38,0.29 0.31,0.21 0.32,0.2 0.32,0.26
Subject
i\j 0 2 3 4
0 0.57,0.5 0.54,0.47 0.54,0.48 0.59,0.39
1 0.57,0.44 0.86,0.78 0.48,0.45 0.55,0.35
2 0.54,0.46 0.87,0.82 0.46,0.46 0.58,0.35
3 0.54,0.44 0.48,0.5 0.46,0.47 0.47,0.34
4 0.58,0.4 0.55,0.45 0.57,0.45 0.47,0.35
Table 8.6. TO( Oi, OJ), RO( Oi, OJ) for Phase II-Ontologies.
Results. Table 8.7 depicts the similarity measures computed for phase III-
ontologies, where only RO has been computed, because the taxonomy was
predefined. RO ranges between 0.23 and 0.71, the average RO over all 20
cross-comparisons achieving 0.5.
Subject
i\j 0 1 2 3 4
0 0.61 0.38 0.51 0.54
0.69 0.56 0.57 0.55
2 0.4 0.49 0.35 0.23
3 0.67 0.71 0.5 0.57
4 0.45 0.44 0.3 0.41
Conclusion. In this case study the proposed measures have been applied for
evaluating human ontology engineering. In the following section the same
measures will be applied to different scenarios for the application of ontology
learning algorithms.
1
varying background knowledge, ....
L
R R
t t
12 _____ ~ 1,2 _____ ~ 12 _____ ~ 1,2 _____ ~
Figure 8.11 depicts the overall evaluation scenario for measuring ontology
learning performance. A given algorithm may generate different parts of an
ontology (e.g. the lexical entry extraction algorithm may generate proposals
for the ontology lexicon). Typically an ontology learning algorithm may be
Evaluation 191
parameterized (e.g. with a threshold) or may be fed with different input texts.
Thus, one algorithm may generate different kinds of ontologies in dependency
from different parameters. A given gold standard ontology may be compared
using the measures with these different kinds of ontologies that are generated
by an algorithm. The comparison then leads to similarity values. Thus, a high
similarity between the gold standard and an extracted ontology represents a
successful application of a specific algorithm based on the given set of param-
eters.
Though human decisions in this matter should not be taken for pure gold, it is
necessary to have measures that allow the comparison of different approaches
and parameter settings - even when the bases of these measures depend to
some extent on the quality of and on rather arbitrary, but equally plausible,
choices between modeling decisions.
Condition
>0 >5 > 20 > 50 > 100 > 500 > 1000
lefi 10,775 2,790 920 427 248 56 31
tfidfi 10,775 10,759 3,605 1,753 892 150 46
recall
1,00
0,80
0,60 -v
•
0,40
lef >=5
0,20
•
tfidf>=20· • lef >=20
tfidf >=50
.lef>=100
0,00 ......~ •
tfidf >=1000
Precision & Recall Evaluation. Figure 8.12 depicts the results obtained by
comparing the automatically extracted set of lexical entries. The reader may
note that the well-known trade-off between precision and recall becomes ob-
vious again in this figure. An interesting aspect of this figure is the average
tfidf measure outperforms lef for the task of lexical entry extraction for ontol-
ogy generation. Another fact has been recognized in evaluating lexical entry
Evaluation 193
extraction: Recall values are very low, even if the extraction algorithms are
executed without any condition. Thus, one may conclude that the manually
engineered GETESS ontology does not optimally reflect the lexical content of
the corpus.
recall
1,00
0,80
0,60
large_single
• large_average
0,40
+ large_complete
0,20
• middle
• small
0,00
0,00 0,20 0,40
precision
To derive the required input matrix for hierarchical clustering we used our
linguistic and heuristic preprocessing that came up with approx. 51,000 lin-
guistically related pairs of concepts using the small reference ontology. The
preprocessing strategy for extracting these pairs of concepts has been described
in chapterS, subsection 2.2.5. The three "background knowledge" ontologies
are distinguished between 0small (55 concepts), 0medium (110 concepts),
194 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
and 0large (211 concepts). They have been manually derived from the simple
"evaluation" tourism ontology. Based on these three input ontologies, three
different computation strategies have been used to derive 9 ontologies that are
compared with the simple "evaluation" tourism ontology introduced earlier.
Figure 8.13 depicts the results obtained by computing precision and recall
for the He element. It is obvious that the three derived ontologies based on the
large background knowledge ontology 0large result in the best precision and
recall values. An interesting aspect however is that the computation strategies
single and average linkage outperform complete linkage. Another interesting
aspect is that the results based on the two background knowledge ontologies
0medium and 0small do not distinguish too much.
TO
0,5 .----.--------------.. -~
large_average
0,4
large_complete large_single
0,3 • • I
0,2
mediu~_average
medium single
0,1 . ~. small~omplele
medium_complete. ~
small_average small_s ngle
0,0
Figure 8.14 depicts the values obtained from computing TO graphically. The
computation strategy "average link" with 0large outperforms the remaining
models. Again, it can be seen that the results based on the two background
knowledge ontologies 0medium and 0small do not distinguish too much.
The evaluation strategy proved to be very useful for deriving computation
strategies, e.g. how clusters are computed (e.g. single link, complete link,
average link). The interested reader is referred to (Boch, 2001) where further
and more detailed evaluation results are provided.
Table 8.9. Overview of Evaluation Results (number of proposed non-taxonomic relations. RO.
recall, precision
Confidence
Support 0.01 0.1 0.2 0.4
0.0001 2429/0.55 865/0.57 485/0.57 238/0.51
66%/2% 31% / 3% 18% / 3% 2%/1%
0.0005 1544 / 0.57 651/0.59 380/0.58 198/0.5
59% / 3% 30%/4% 17%/4% 1% /1%
0.002 889/0.6 426/0.61 245/0.61 131/0.52
47%/5% 27%/6% 16% / 6% 1% /1%
0.01 342/0.64 225/0.64 143/0.64 74/0.53
31%/8% 19% / 8% 14% / 8% 1%/1%
0.04 98/0.67 96/0.67 70/0.65 32/0.51
13 % / 11 % 11%/10% 6%/7% 0%/0%
0.06 56/0.63 56/0.63 48/0.62 30/0.53
6%/9% 6%/9% 3%/6% 0%/0%
recall
1,00
0,80
0,0001/0,01
0,60
• "0,0005/0,01
.0,00210,01
0,40
•• .0,002/0,1
0,20
•• • :0,0410,01
0,00 .. •• •
0,0410,1
5. Conclusion
In this chapter an evaluation approach for ontology learning has been intro-
duced. The approach is based on the ontology structure definition of chapter 2
and follows the layered view on ontologies distinguishing between a lexical and
a conceptual level. The underlying idea of the overall approach is to compute
Evaluation 197
Notes
1 The reader may note the application-specific evaluation in a knowledge man-
agement scenario becomes more difficult with respect to the current unsolved
problem of measuring the success of a knowledge management initiative.
2 Parts of this section have been published in (Maedche and Staab, 2001 b).
3 The reader may note that measures for computing similarity between ontolo-
gies may open a wide range of more general applications, like for example
agent-based systems or for ontology merging & mapping tasks.
4 From the statistical point of view false negatives are type I errors, false
positives are type II errors.
5 The algorithm is based on a dynamic programing technique that is described
in detail in (Levenshtein, 1966). The algorithm builds a matrix for the two
strings that are to be compared. Each element (x,y) depends on the values of
(x-1,y), (x,y-l) and (x-l,y-l). Whenever the characters for x and yare the
same then the value of (x,y) will be equal to the minimum of the three values
it depends on. When the characters are different then the value of (x,y) will
be this minimum plus one. In order to be able to compute the (x,y) values the
program uses a virtual initial row with only larger integer values as elements
and a virtual initial column with the values 0,1,2,3,4, ...
6 The geometric mean reflects the intuition that if either domain or range
concepts utterly fail to match, the matching accuracy converges against 0,
whereas the arithmetic mean value might still turn out a value of 0.5.
7 see http://www.all-in-all.de
8 It contained four concepts referred to by THING, MATERIAL, INTANGIBLE,
and SITUATION organized in the hierarchical relationships llC(MATERIAL,
THING) and ll C(SITUATION,INTANGIBLE).
9 http://www.all-in-all.coml
lOThe UCI Machine Learning Repository provides different kinds of standard
data sets and is online available at
http://www.ics.uci.edul mlearnlMLSummary.html
IV
RELATED WORK
1993; Han and Kamber, 2001» have been investigated by the database com-
munity).
• The machine learning community has a long research tradition in learning
from all kinds of data. One may distinguish between propositional and
non-propositional algorithms. The latter ones are further researched by the
inductive logic programming community (see (Muggleton, 1992»).
• Research for extracting domain knowledge from web documents has also
been done in the information retrieval community targeting a better ac-
cess to documents. Especially the clustering of term hierarchies has been
researched (see for example (Sanderson and Croft, 1999».
• The research area of terminology has much experience in acquiring and
modeling terminologies (e.g., (Biebow and Szulman, 1999; Daille, 1996».
In their work they concentrate mainly on the extraction of terms from a given
set of document resources.
• The knowledge engineering and acquisition community is a classic field
that deals with modeling knowledge-based systems. Within the knowledge
engineering community mechanisms for semi-automatically acquiring con-
ceptual knowledge supporting knowledge acquisition have been researched
a long time, e.g. the work done by (Skuce et aI., 1985; Reimer, 1990; Sz-
pakowicz, 1990).
All of these research communities and areas have (mostly) independently
analyzed and explored methods and algorithms that may be subsumed under
the term ontology learning. The list given above has to be considered as non-
exhaustive.
As mentioned above there is a second possibility for organizing related work:
One may introduced related work according to the "organization of the overall
book", namely ontology engineering, data import & processing, algorithms,
etc. In this chapter the second approach is followed as a way to provide the
reader a comprehensive overview on existing and related work. Figure 9.1
depicts a taxonomy of related work. Related work is mainly distinguished
between work on ontology engineering, on knowledge acquisition (KA) and
machine learning (ML) frameworks, on data and import processing, on
algorithms and on evaluation.
The following five sections will elaborate further on these main categories
of related work.
Comparison application-
with semantic based
merging
terms prune
hierarchy association
Methodologies for Ontology Engineering. In the past years only a few re-
search groups proposed methodological approaches guiding the ontology de-
velopment process. U scholds generic suggestions were the first methodological
outlines proposed in 1995 on the basis ofthe experience gathered in developing
the Enterprise Ontology (see (Uschold and King, 1995)). The methodological
outlines may be separated into five core guidelines, namely the identification
of the purpose, the building of the ontology (separated into capturing, coding
and integrating), evaluation and documentation. On the basis of the Toronto
Virtual Enterprise (TOVE) project, Grueninger and Uschold described ontol-
ogy development steps in (Usc hold and Gruninger, 1996). At the same time
METHONTOLOGY by (Gomez-Perez, 1996) appeared. In parallel the more
philosophical viewpoint on ontology has been evolved towards an engineering
discipline. (Guarino and Welty, 2000) demonstrate how some methodology
efforts founded on analytic notions that have been drawn from philosophy can
be used as formal tools of ontological analysis. A more linguistic viewpoint on
ontology has been provided by Kathleen Dahlgren. She defends the choice of
a linguistically-based content ontology for NLP and demonstrates that a single
206 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
The Protege methodology, to which the tool belongs, allows system builders
to construct software systems from modular components including reusable
frameworks for assembling domain models and reusable domain-independent
problem-solving models that implement procedural strategies for solving tasks.
The idea behind this methodology is the ontology editor and the layout editor
are supporting tools for the final generation of a knowledge acquisition tool for
entering instances. In contrast to the work described here ONTO EDIT does
not focus on the automatic generation of knowledge acquisition interfaces. In
general there is a trade-off between the automatic generation of user interfaces
and the question how ergonomic these user interfaces are. One example that
may be given for this trade-off is the relation hierarchy. Naturally, hierarchi-
cal relations may be defined by a simple template based interface. However,
user will prefer having at least a tree-oriented user interface available for the
definition of a relation hierarchy.
Ontology Merging. Several systems and frameworks for supporting the know l-
edge engineer in the ontology merging task have recently been proposed. The
approaches mainly rely on syntactic and semantic matching heuristics which
are derived from the behavior of ontology engineers when confronted with the
task of merging ontologies, i.e. human behaviour is simulated. Although some
of them locally use different kinds of logics for comparisons (e.g. descrip-
tion logics), these approaches do not offer a structural description of the global
merging process.
A first approach for supporting the merging of ontologies is described in
(Hovy, 1998). There, several heuristics are described for identifying corre-
sponding concepts in different ontologies, e.g. comparing the names of two
concepts, comparing the natural language definitions of two concepts by lin-
guistic techniques, and checking the closeness of two concepts in the concept
hierarchy.
The OntoMorph system (Chalupsky, 2000) offers two kinds of mechanisms
for translating and merging ontologies: syntactic rewriting supports the trans-
lation between two different knowledge representation languages and semantic
rewriting offers means for inference-based transformations. It explicitly al-
lows to violate the preservation of semantics in trade-off for a more expressive,
flexible transformation mechanism.
In (McGuinness et aI., 2000) the Chimaera system is described. It provides
support for the merging of ontological terms from different sources, for check-
ing the coverage and correctness of ontologies and for maintaining ontologies
over time. Chimaera supports the merging of ontologies by coalescing two
semantically identical terms from different ontologies and by identifying terms
that should be related by subsumption or disjointness relationships. Chimaera
offers a broad collection of functions, but the underlying assumptions about
structural properties of the ontologies at hand are not made explicit.
Prompt (Noy and Musen, 2000; Noy and Musen, 2001) is an algorithm for
ontology merging and alignment embedded in Protege-2000. It starts with the
identification of matching class names. Based on this initial step an iterative
approach is carried out for performing automatic updates, finding resulting
conflicts, and making suggestions to remove these conflicts. The work is im-
plemented as an extension to the Protege-2000 knowledge acquisition tool and
offers a collection of implemented operations for merging two classes and re-
lated slots.
The tools described above offer extensive merging functionalities most of
them based on syntactic and semantic matching heuristics, which are derived
from the behaviour of ontology engineers when confronted with the task of
merging ontologies. OntoMorph and Chimarea use a description logics based
approach that influences the merging process locally, e. g. checking subsump-
tion relationships between terms. None of these approaches offers a structural
Related Work 209
While people coming from knowledge acquisition think that one can insert ma-
chine learning algorithms considered as black boxes in knowledge acquisition
environments and tools.
The experiences in this work have shown that a tight integration between
the task of manually engineering an ontology and automatically generating
concepts and conceptual relationships is required. In the following, a short
introduction is given into existing work of combining KA and ML on structured
data (such as the tuples in a given database), and an overview is provided of the
work that tries to combine KA and ML for its application on natural language
texts.
ulary have been defined" and shows a significant improvement to the overall
development tasks by using machine learning techniques.
In the following existing work is referred with respect to the points intro-
duced above, namely focused crawling, linguistic processing and the document
wrapper.
Focused Crawling. The need for focused crawling in general has recently
been conceived by several researchers. The main target of all of these ap-
proaches is to focus the search of the crawler and to enable goal-directed
crawling. (Chakrabarti et aI., 1999) present a generic architecture of a focused
crawler. The crawler uses a set of predefined documents associated with topics
in a Yahoo like taxonomy to build a focused crawler. Two hypertext mining
algorithms build the core of their approach: a classifier evaluates the relevance
of a hypertext document with respect to the focus topics, and a destiller that
identifies hypertext nodes that are great access points to many relevant pages
within a few links. The approach presented in (Diligenti et aI., 2000) uses
so-called context graphs as a means to model the paths leading to relevant web
pages. Context graphs in their sense represent link hierarchies within which
relevant web pages occur together in the context of these pages. (Rennie and
McCallum, 1999) propose a machine learning oriented approach for focused
crawling. Their crawler uses reinforcement learning to learn to choose the next
link such over time a reward is maximized. A problem of their approach may
be that the method requires large collections of already visited web pages.
merick et aI., 1997; Ashish and Knoblock, 1997). For the low-level task of
transforming a given more or less well-structured dictionary into the internal
representation described here one easily manually defines the required mapping
and extraction rules. However, the reader may note that recently a number of
tools (e.g. (Sahuguet and Azavant, 1999)) and approaches (see (Kushmerick
et aI., 1997; Ashish and Knoblock, 1997)) for helping the manual or (semi-
)automatic construction of wrappers have been developed.
Lexical Entry Extraction. At the lowest level of ontology learning one typ-
ically has to deal with the task of extracting lexical entries referring to concepts
and relations. A short overview is given how this task has been approached by
several people from different research communities. Much work has been done
in the area of lexical acquisition. Lexical acquisition deals with the task of
acquiring syntactic and semantic classifications of unknown words. A compre-
hensive overview on lexical acquisition in the context of information extraction
is given in (Basili and Pazienza, 1997). In this overview paper the authors
identify the following methodological areas for lexical acquisition:
• Statistical induction using collocations, syntactic features, or lexemes.
• Logical induction using symbolic representations at word, phrase or sen-
tence level.
• Machine readable dictionary (MRD) and lexical knowledge base extrac-
tion including all methods that deal with some systematic sources like dic-
tionaries (like LDOCE) or general purpose lexical knowledge bases (like
WordNet).
• Quantitative machine learning referring to all other inductive methods that
are not purely statistical (e.g. neural networks).
According to the methodological areas the attention is mainly restricted to the
first three points. In contrast to existing work these different methodological
areas are combined into a common view for ontology learning on multiple
sources. In the following a short overview is given of how these methodological
areas have been approached in the existing work. The terminology research
community focuses on the extraction of terminologies from given data. The tool
Related Work 215
In contrast to the research described above in this work the idea is pursued that
the construction of semantic knowledge requires the combination of information
Related Work 217
from multiple sources (according to (Ide and Veronis, 1995». Clearly, coupled
with information from other sources (like free texts) and subjected to by-hand
amelioration the structures extracted from dictionaries are a valuable resource
for building ontologies.
To contrast the proposed approach with the research just cited, the reader
may note that all the verb-centered approaches may miss important conceptual
relations not mediated by verbs. Regarding evaluation, they have only appealed
218 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
to the intuition ofthe reader (Byrd and Ravin, 1999; Faure and Nedellec, 1998)
or used application-dependent evaluation measures.
In the area of text mining, (Feldman and Hirsh, 1996) have presented an
approach for association mining in the presence of background knowledge. In
this paper the system FACT for knowledge discovery from text is presented.
It is key-word oriented and offers a query-centered mechanism for extracting
associations. Background knowledge is used to constraint the desired results
of the query process. The evaluation of their approach is restricted to efficiency
without considering the quality of the extracted associations.
Ontology Pruning. (Peterson et aI., 1998) have described strategies that leave
the user with a coherent ontology (i.e. no dangling or broken links). The un-
derlying system of their approach is called Knowledge Bus. It is a system that
generates information systems (databases and programming interfaces) from
application-focused subsets of the CyC ontology2. In their approach the fol-
lowing four major components are distinguished: The sub-ontology extractor
identifies a domain-relevant section of the ontology. The logic program gen-
erator takes the extraction and translates it into a logic program which can be
evaluated by a deductive query engine3 . Then the API generator takes the
logic-based model and exposes it to application developers as an object model
through strongly typed object-oriented APIs. Finally, the runtime system sup-
ports access to the generated databases.
Related Work 219
A similar strategy has been described by (Swartout et ai., 1996) where on-
tology pruning is considered as the task of "intelligent" deletion of ontological
structures that leave the user with a coherent ontology.
concepts, and relations mainly within one ontology. The nearest to the proposed
comparison between two ontologies come (Bisson, 1992) and (Weinstein and
Birmingham, 1999).
(Bisson, 1992) introduces several similarity measures in order to locate a new
complex concept into an existing ontology by similarity rather than by logic
subsumption. Bisson restricts the attention to the semantic comparison level.
In contrast to the work described here the new concept is described in terms
of the existing ontology. Furthermore, he does not distinguish relations into
taxonomic relations and relations, thus ignoring the semantics of inheritance.
(Weinstein and Birmingham, 1999) compute description compatibility in or-
der to answer queries that are formulated with a conceptual structure that is
different from the one of the information system. A comprehensive introduc-
tion into the approach of Weinstein is given in (Weinstein, 1990). In contrast
to the proposed approach their measures depend to a very large extent on a
shared ontology that mediates between locally extended ontologies. Also their
algorithm seems less suited to evaluate similarities of sets of lexical entries,
taxonomies, and relations.
Notes
1 A detailed introduction is online available at
http://www.ontoknowledge.org/downl/de115.pdf and given in (Staab et aI.,
2000a; Staab et aI., 200le; Maedche et aI., 200le)
2 http://www.cye.com!
3 In their approach they use the freely avalaible XSB system
http://xsb.soureeforge.neti
4 Contextual Acquisition Mechanism for Incremental Lexeme Learning
5 A comprehensive overview is available at
http://www.cse.unsw.edu.au. a web page maintained by Tim Menzies.
Chapter 10
1. Contributions
The contributions made by this book fall into the following three main areas:
A. Maedche, Ontology Learning for the Semantic Web
© Kluwer Academic Publishers 2002
224 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Preprocessing. It is difficult to find the right data representation for the ap-
plication of an ontology learning algorithm. A similar experience was found
in machine learning and knowledge discovery. A useful approach is the user
is guided in the difficult preprocessing task, e.g., by accessing an "experience
base" of successful preprocessing strategies (see also (Engels, 1999)).
3. Unanswered Questions
While a number of problems remain with the single disciplines, more chal-
lenges regarding the particular problem of Ontology Learning for the Semantic
Web arise. Any book raises new questions while it answers old ones. A book
like this one, the subject of which is novel and relatively unexplored, seems to
raise more questions than it answers. In this section the open questions that are
not answered are identified:
• First, as mentioned earlier we are still in an early stage with respect to pro-
viding methodological guidelines in applying ontology learning to support
ontology engineering. Thus, in the future the integration between a com-
prehensive methodology with support of the application of semi-automatic
226 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
means is required. This holds especially true for the difficult tasks of data
import and processing, where experiences have to be collected and have to
be provided to the ontology engineer.
• Second, the attention has been restricted in ontology learning to the con-
ceptual structures that are (almost) contained in RDF(S) proper. Additional
semantic layers on top of RDF (e.g., future OIL or DAML+OIL with ax-
ioms, AO) will require new means for improving ontology engineering with
axioms, too! Thus, one important open question is how axioms can be ac-
quired from existing Web data.
• Third, a tight integration of techniques for the extraction of ontological struc-
tures from databases, semi-structured data, and existing instances with
the techniques proposed in this book have to be established. Nevertheless,
it is expected the more available data resources are included in the ontology
learning process, the better the overall performance will be.
4. Future Research
A number of future research topics and challenges at the end of each chapter
have already been listed, e.g.,
• More comprehensive, multi-lingual natural language processing support,
e.g. for automatically deriving lexical entries referring to non-taxonomic
relations using on verb-centered approaches.
• Including structural properties, e.g. HTML tags, to use the explicit content
contained in HTML tables. An interesting and promising table mining
approach has been introduced by (Chen et aI., 2000)).
• Multi-relational representations for the application of more logic-oriented
machine learning techniques.
• Multi-strategy learning techniques that e.g. use voting for combining the
results generated by different algorithms.
• Comparison of the results generated by different evaluation measures and
techniques.
• Development of multi-lingual standard data sets for ontology learning.
Finally, the following three important tasks for the future development and
application of ontology learning are considered as especially important.
Semantic Web Mining. It has been already mentioned that looking at the
user's behaviour may indicate necessary ontology changes and updates. In
the research area of Web Mining one applies data mining techniques in the
web. In general it is distinguished between web usage mining analyzing the
user behavior, web structure mining exploring the hyperlink structure, and web
content mining exploiting the contents of the documents in the web. A problem
of the current approaches is that they operate on syntactic, often meaningless
structures such as hyperlinks. In the near future, approaches that exploit the
complex structures contained in the Semantic Web in combination with the
analysis of user behaviour should be researched.
References
Abecker, A., Bernardi, A., Hinkelmann, K., KUhn, 0., and Sintek, M. (1998). Towards a technol-
ogy for organizational memories. IEEE Intelligent Systems and Their Applications, 13(3 ):4~
48.
Abiteboul, S., Buneman, P., and Suciu, D. (1999). Data on the Web: From Relations to Semistruc-
tured Data and Xml. Morgan Kaufmann Publishers, CA.
Abiteboul, S., Hull, R., and Vianu, V. (1994). Foundations of Databases. Addison Wesley, Mas-
sachusetts.
Adam, N. and Yesha, Y. (1996). Strategic directions in electronic commerce and digital libraries:
Towards a digital agora. ACM Computing Surveys, 28(4):818-835.
Agirre, E. and Rigau, G. (1996). Word sense disambiguation using conceptual density. InProc.
ofCOLING-96, pages 16-22.
Agrawal, R., Imielinski, T., and Swami, A. (1993). Mining Associations between Sets of Items
in Massive Databases. In Proceedings of the 1993 ACM SIGMOD International Conference
on Management of Data, Washington, D.C., May 26-28, 1993, pages 688-692. ACM Press.
Amann, B. and Fundulaki, I. (1999). Integrating Ontologies and Thesauri to Build RDF Schemas.
In Proceedings ofthe European Conference ofDigital Libraries - ECDL' 1999, Paris, France,
1999, pages 234-253.
Angele, J., Schnurr, H.-P., Staab, S., and Studer, R. (2000). The times they are a-changin' -
the corporate history analyzer. In Mahling, D. and Reimer, U., editors, Proceedings of the
Third International Conference on Practical Aspects of Knowledge Management. Basel,
Switzerland, October 30-31,2000.
http://www.research.swisslife.chlpakm2000/.
Appelt, D., Hobbs, J., Bear, J., Israel, D., and Tyson, M. (1993). FASTUS: A finite state processor
for information extraction from real world text. In /JCAI-93: Proceedings of the J3Interna-
tional Joint Conference on Artificial Intelligence. Chambery, France, August 28 - September
3, 1993, pages 1172-1178, Chambery, France.
Ashish, N. and Knoblock, C. (1997). Semi-automatic wrapper generation for internet informa-
tion sources. In Proceedings of the Second IFCIS International Conference on Cooperative
Information Systems, Kiawah Island, South Carolina, USA, June 24-27, 1997, Sponsored by
IFC1S, The Intn'l Foundation on Cooperative Information Systems, pages 1~169. IEEE-CS
Press.
Assadi, H. (1999). Construction of a regional ontology from text and ist use within a documentary
system. In N. Guarino (ed.), Formal Ontology in Information Systems, Proceedings of FOIS-
98, Trento, Italy, 1999, pages 236-249.
230 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Baldwin, B., Morton, T., Bagga, A., Baldridge, J., Chandraseker, R., Dimitriadis, A., Snyder, K.,
and Wolska, M. (1998). Description of the UPENN CAMP system as used for coreference.
In (MUC7, 1998).
Barker, K., Delisle, S., and Szpakowicz, S. (1998). Test-driving TANKA: Evaluating a semi-
automatic system of text analysis for knowledge acquisition. In Mercer, R. and Neufeld, E.,
editors, Advances in Artificial Intelligence. Proceedings of the 12Biennial Conference of the
Canadian Society for Computational Studies of Intelligence (AI '98). Vancouver, Canada,
June 18-20, 1998, LNAI 1418, pages 60-71, Berlin. Springer.
Basili, R. and Pazienza, M. T. (1997). Lexical acquisition and information extraction. In SCIE
1997: Rome, Italy, 1997.
Basili, R., Pazienza, M. T., and Velardi, P. (1993). Acquisition of selectional patterns in a sub-
language. Machine Translation, 8(1):175-20l.
Bateman, J. A. (1993). Ontology construction and natural language. In Proceedings of the Inter-
national Workshop on Formal Ontology in Conceptual Analysis and Knowledge Representa-
tion, Padova, March 1993, pages 83-93.
Bateman, J. A. (1995). On the relationship between ontology construction and natural language:
a socio-semiotic view. International Journal on Human-Computer Studies, 43:929-944.
Benjamins, R., Duineveld, A. J., Stoter, R., Weiden, M. R., and Kenepa, B. (1999). Wondertools?
A comparative study of ontological engineering tools. In Proceedings ofthe 12th International
Workshop on Knowledge Acquisition, Modeling and Mangement (KAW'99), Banff, Canada,
October 1999.
Berners-Lee, T. (1999). Weaving the Web. Harper, San Francisco.
Berners-Lee, T., Hendler, J., and Lassila, O. (2001). The semantic web. Scientific American.
Biebow, B. and Szulman, S. (1999). TERMINAE: A a linguistics-based tool for the building of a
domain ontology. In EKAW '99 - Proceedings of the 11th European Workshop on Knowledge
Acquisition, Modeling, and Management. Dagstuhl, Germany; LNCS, pages 49-66, Berlin.
Springer.
Bisson, G. (1992). Learning in FOL with a similarity measure. In Proc. of AAA1-1992, pages
82-87.
Bisson, G., Nedellec, c., and Canamero, D. (2000). Designing clustering methods for ontology
building: The Mo'K workbench. In (Staab et al., 2oooc).
Bloom, P. (2000). How Children learn the Meanings of Words. MIT Press, Massachusetts.
Boch, T. (2001). Separating taxonomic from non-taxonomic relations discovered from text.
Master's thesis, University of Karlsruhe.
Boyens, K. (2001). OntoKick - Ignition for Ontologies. Master'S thesis, University of Karlsruhe.
Brachman, R. (1979). On the epistomoiogical status of semantic networks. Associative Networks,
pages 3-50.
Bray, T., Hollander, D., and Layman, A. (1999). Namespaces in XML. Technical report, W3C.
W3C Recommendation. http://www.w3.orgITRlREC-xml-names.
Brickley, D. and Guha, R. (2000). Resource Description Framework (RDF) Schema Specifica-
tion. Technical report, W3C. W3C Candidate Recommendation.
http://www.w3.orglTRl2ooo/CR-RDF-schema-20000508.
Brill, E. (1993). Automatic grammar induction and parsing free text: A transformation-based
approach. In ACL'93 - Proceedings of the Annual Meeting of the Associationfor Computa-
tional Linguistics, pages 259-265, Ohio.
Brown, A. (2000). Large-Scale Component-Based Development. Prentice Hall.
Buitelaar, P. (1998). CORELEX: Systematic Polysemy and Underspecification. PhD thesis, Bran-
deis University, Department of Computer Science.
REFERENCES 231
Buneman, P., Davidson, S. B., Fernandez, M. F., and Suciu, D. (1997). Adding structure to
unstructured data. In Afrati, F. N. and Kolaitis, P., editors, Proceedings ofthe 6th International
Conference on Database Theory-ICDT'97, Delphi, Greece, pages 336-350. Springer.
Buntime, W. and Stirling, D. (1991). Interactive Induction. J.E. Hayes and D.Michie and E.
Tyugu (Eds.), Clarendon Press, Oxford.
Byrd, R. and Ravin, Y. (1999). Identifying and extracting relations from text. In NLDB '99 - 4th
International Conference on Applications of Natural Language to Information Systems.
Campbell, K., Oliver, D. E., Spackman, K., and Shortliffe, E. H. (1998). Representing Thoughts,
Words, and Things in the UMLS. Technical report, SMI Stanford Medical Informatics.
Chakrabarti, S., van den Berg, M., and Dom, B. (1999). Focused crawling: a new approach to
topic-specific web resource discovery. In Proceedings of www-s.
Chalupsky, H. (2000). OntoMorph: A translation system for symbolic knowledge. In Proc. of
KR-2000, Breckenridge, CO, USA, pages 471-482.
Chaudhri, v., Farquhar, A., Fikes, R., Karp, P., and Rice, J. (1998). OKBC: A Programmatic
Foundation for Knowledge Base Interoperability. In Proceedings 15th National Conference
on Artificial Intelligence (AAAI-9S), pages 600-607.
Chen, H. (1999). Semantic research for digital libraries. D-Lib Magazine, 5(10).
Chen, H.-H., Tsai, S.-c., and Tsai, J.-H. (2000). Mining Tables from Large Scale HTML Texts
. In Proceedings of the 18th International Conference on Computational Linguistics, Saar-
bruecken, Germany, July 2000.
Chen, P. (1976). The entity-relationship model- toward a unified view of data. ACM Transactions
on Database Systems, 1(1):9-36.
Chinchor, N., Hirschman, L., and Lewis, D. (1993). Evaluating message understanding sys-
tems: An analysis of the third message understanding conference (MUC-3). Computational
Linguistics, 19(3).
Chomsky, N. (1965). Aspects of the Theory of Syntax. MIT Press, Cambridge.
Christophides, V. and Plexousakis, D., editors (2000). Proceedings of the ECDL-2000 Workshop
- Semantic Web: Models, Architectures and Management.
Conen, W. and Klapsing, R. (2000). A logical interpretation ofRDF. RDF Interest Group Mailing
List, http://nestroy.wi-inf. uni -essen.delrdfllogical_interpretationl.
Cooper, J. W. and Byrd, R. J. (1997). Lexical navigation: Visually prompted query expansion
and refinement. In Proceedings of the International Conference on Digital Libraries DL'97,
pages 237-246.
Corby, 0., Dieng, R., and Hebert, C. (2000). A conceptual graph model for w3c resource descrip-
tion framework. In ICCS 2000 - International Conference on Conceptual Structures. Darm-
stadt, Germany, August 2000, Lecture Notes in Artificial Intelligence LNAI-1867. Springer.
Craven, M., DiPasquo, D., Freitag, D., McCallum, A., Mitchell, T., Nigam, K., and Slattery, S.
(1999). Learning to construct knowledge bases from the world wide web. Artificial Intelli-
gence, lI8( 1-2):69-113.
Cumby, C. and Roth, D. (2000). Relational representations that facilitate learning. In Proc. of
KR-2000, Breckenridge, Colorado, USA, 12-15 April 2000, pages 425-434.
Dagan, I., Lee, L., and Pereira, F. (1999). Similarity-based models of word coocurrence proba-
bilities. Machine Learning, 34(1):43-69.
Dahlgren, K. (1995). A linguistic ontology. International Journal of Human-Computer Studies,
43(5).
Daille, B. (1996). Study and Implementation of Combined Techniques for Automatic Extraction
of Terminology, chapter 3, pages 49-67. Klavans J. L. and Resnik P. (ed.): The Balancing
Act, Combining Symbolic and Statistical Approaches to Language, MIT Press, Cambridge
Mass., London England.
de Saussure, F. (1916). Course in general linguistics. McGraw Hill, New York.
232 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Decker, S., Brickley, D., Saarela, 1., and Angele, J. (1998). A Query and Inference Service for
RDF. In Proceedings ofthe W3C Query Language Workshop (QL-98), Boston, MA, December
3-4.
Decker, S., Fensel, D., van Harrnelen, F., Horrocks, I., Melnik, S., Klein, M., and Broekstra,
J. (2000a). Knowledge representation on the web. In Proceedings of the 2000 International
Workshop on Description Logics (DL2000), Aachen, Germany.
Decker, S., Jannink, J., Mitra, P., Staab, S., Studer, R., and Wiederhold, G. (2000b). An infonna-
tion food chain for advanced applications on the www. In ECDL 2000 - Proceedings of the
Fourth European Conference on Research and Advanced Technology for Digital Libraries.
Lisbon, Portugal, September 18-20, 2000, LNCS, pages 490-493. Springer.
Decker, S. and Melnik, S. (2000). A layered approach to infonnation modeling and interoper-
ability on the web. In (Christophides and Plexousakis, 2(00).
Decker, S., Mitra, P., and Melnik, S. (2000c). Framework for the Semantic Web - An RDF
Tutorial. IEEE Internet Computing.
Delteil, A., Faron-Zucker, C., and Dieng, R. (2001). Learning ontologies from RDF annotations.
In Maedche, A., Staab, S., Nedellec, c., and Hovy, E., editors, Proceedings of /JCA/-OI
Workshop on Ontology Learning OL-200I, Seattle, August 2001, Menlo Park. AAAI Press.
Diligenti, M., Coetzee, F., Lawrence, S., Giles, C. L., and Gori, M. (2000). Focused Crawling us-
ing Context Graphs. In Proceedings of the International Conference on Very Large Databases
(VLDB-OO), 2000, pages 527-534.
Doan, A., Domingos, P., and Levy, A. (2000). Learning source descriptions for data integration.
In Proceedings of the International Workshop on The Web and Databases - WebDB-2000,
pages 81-86.
Eco, U. (1981). Zeichen. Einfiihrung in einen Begriffund seine Geschichte. Suhrkamp (edition
suhrkamp), FrankfunlM.
Ehrig, M. (2001). Ontology-based Focused Crawling of Documents and Relational Metadata.
Master's thesis, University of Karlsruhe.
Engels, R. (1999). Component-Based User Guidance in Knowledge Discovery and Data Mining.
PhD thesis, University of Karlsruhe.
Engels, R., Bremdal, B., and Jones, R. (2001). CORPORUM: a workbench for the Semantic
Web. In Proceedings of the First Workshop on Semantic Web Mining, Freiburg, Germany,
September 2001. online available at http://semwebmine200l.aifb.uni-karlsruhe.de/online/.
Erdmann, M. (2001). Ontologien zur konzeptuellen Modellierung der Semantik von XML (in
german). PhD thesis, University of Karlsruhe.
Erdmann, M., Maedche, A., Schnurr, H.-P., and Staab, S. (2001). From manual to semi-automatic
semantic annotation: About ontology-based text annotation tools. ETAI- Semantic Web Jour-
nal, Linkoeping Electronic Articles, 16(1).
Erdmann, M., Maedche, M., Staab, S., and Decker, S. (2000). Ontologies in RDF(S). Technical
Report 401, Institute AIFB, Karlsruhe University.
Esposito, F., Ferilli, S., Fanizzi, N., and Semeraro, G. (2000). Learning from Parrsed Sentences
with INTHELEX. In Proceedings ofCoNLL-2000 and LLL-2000-Intemational Conference
on Grammar Inference (ICGI-2000), to appearin : Lecture Notes in Artificial Intelligence,
Springer.
Euzenat,1. (2000). Towards fonnal knowledge intelligibility atthe semioticIevel. In Proceedings
of the ECAI-2000 Workshop on Applied Semiotics ASC-2000, Berlin, Germany, 2000.
Faure, D. and Nedellec, C. (1998). A corpus-based conceptual clustering method for verb frames
and ontology acquisition. In In LREC workshop on Adapting lexical and corpus resources to
sublanguages and applications, Granada, Spain, Mai 1998.
REFERENCES 233
Faure, D. and Poibeau, T. (2000). First experiments of using semantic knowledge learned by
asium for information extraction task using intex. In Proceedings ofthe ECAI '2000 Workshop
Ontology Learning.
Feldman, R. and Dagan, I. (1995). Knowledge discovery in textual databases (kdt). In Proceed-
ings ofKDD-95, pages 112-117. ACM.
Feldman, R. and Hirsh, H. (1996). Mining associations in text in the presence of background
knowledge. In Proceedings ofthe Second International Conference on Knowledge Discovery
from Databases, pages 343-346.
Fellbaum, C. (1998). WordNet - An electronic lexical database. MIT Press, Cambridge, Mas-
sachusetts and London, England.
Fensel, D. (2001). Ontologies: Silver Bullet for Knowledge Management and Electronic Com-
merce. Springer, Berlin - Heidelberg - New York.
Fensel, D., Van Harmelen, F., Decker, S., Erdmann, M., and Klein, M. (2000). OIL in a nutshell.
In Dieng, R., editor, Knowledge Acquisition, Modeling, and Management, Proceedings of
the European Knowledge Acquisition Conference (EKAW-2000), Lecture Notes in Artificial
Intelligence, LNAI, pages 1-16. Springer-Verlag.
Fikes, R., Farquhar, A., and Rice, 1. (1997). Tools for assembling modular ontologies in Ontolin-
gua. In Proc. of AAAI 97, pages 436-441.
Fikes, R. and McGuiness, D. (2001). An Axiomatic Semantics for RDF, RDF(S), and DAML+OIL.
Technical report, Stanford University, KSL.
Fisher, D., pazzani, M., and Langley, P. (1991). Concept Formation: Knowledge and Experience
in Unsupervised Learning. Morgan Kaufmann, San Francisco.
Fong, 1. (1997). Converting relational to object-oriented databases. SIGMOD Record, 26(1):53-
58.
Franconi, E. and Ng, G. (2000). The i.com tool for Intelligent Conceptual Modeling. In Proceed-
ings of 7th International Workshop on Knowledge Representation meets Databases KRDB-
2000. Berlin, pages 45-53.
http://sunsite.informatik.rwth-aachen.de/Publications/CEUR -W SlVol-29/.
Frantzi, K., Ananiadou, S., and Mirna, H. (2000). Natural language processing for digital libraries:
Automatic recognition of multi-word terms: the c-value/nc-value method. International Jour-
nal on Digital Libraries, 3(2).
Frege, G. (1922). Begriffschrift. Lubrecht & Cramer Ltd., London.
Freitag, D. (1998). Machine Learning for Information Extraction in Information Domains. PhD
thesis, Computer Science Department, Carnegie Mellon University, Pittsburgh, PA,USA.
Fuernkranz, 1., Mitchell, T., and Riloff, E. (1998). A Case Study in Using Linguistic Phrases
for Text Categorization on the WWW. In Proc. of AAAIIICML Workshop Learning for Text
Categorization, Madison, WI, 1998. AAAI Press.
Ganter, B. and Wille, R. (1999). Formal Concept Analysis: Mathematical Foundations. Springer,
Berlin - Heidelberg - New York.
Genesereth, M. R. (1998). Knowledge interchange format. draft proposed american national stan-
dard (dpans). ncits.t2/98-004. http://logic.stanford.edulkif/dpans.html seen at Sep 7,2000.
Ginsberg, M. (1991). Knowledge interchange format: the KIF of death. AI Magazine, 12(3):57-
63.
Goldman, R. and Widom, J. (1997). DataGuides: Enabling Query Formulation and Optimization
in Semistructured Databases. In Proceedings of the Conference on Very Large Databases-
VLDB'1997, Athens, Greece, 1997.
Gomez-Perez, A. (1996). A framework to verify knowledge sharing technology. Expert Systems
with Application, 11(4):519-529.
Grefenstette, G. (1994). Explorations in Automatic Thesaurus Discovery. PhD thesis, University
of Pittsburgh.
234 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Hovy, E. (1998). Combining and standardizing large-scale, practical ontologies for machine
translation and other uses. In Proc. of the First Int. Can! on Language Resources and Eval-
uation (LREC).
Hudson, R (1990). English Word Grammar. Basil Blackwell, Oxford, UK.
Ide, N., McGraw, T., and Welty, C. (1997). Representing technical documents in the classic
knowledge representation system. In Proceedings ofthe Tenth workshop of the Text-Encoding
Initiative. November, 1997.
Ide, N. and Veronis, J. (1995). Knowledge extraction from machine-readable dictionaries: An
evaluation. In P Steffens (ed.): Machine Translation and the lexicon, Lecture Notes on Arti-
ficial Intelligence LNAI-898, pages 19-34. Springer.
ISO 704 (1987). Principles and methods of terminology. Technical report, International Standard
ISO.
Jannink, J. and Wiederhold, G. (1999). Thesaurus entry extraction from an on-line dictionary. In
Proceedings of Fusion '99, Sunnyvale CA, July 1999.
Jaquemin, C. (2001). Spotting and Discovering Terms through Natural Language Processing.
MIT Press, Massachusetts.
Jasper, R and Uschold, M. (1999). A Framework for Understanding and Classifying Ontology
Applications. In Proc. the 12th International Workshop on Knowledge Acquisition, Modeling
and Mangement (KAW'99), Banff, Canada, October 1999.
Jones, R., McCallum, A., Nigam, K., and Riloff, E. (1999). Bootstrapping for Text Learning
Tasks. In Working notes of the IJCA/'99 workshop on Text Mining: Foundations, Techniques
and Applications, pages 52-63. MIT Pressl AAAI Press.
Kamada, T. and Kawai, S. (1989). An algorithm for drawing general undirected graphs. Infor-
mation Processing Letters, 31(7):7-15.
Kaufman, L. and Rousseeuw, P. (1990). Finding Groups in Data: An Introduction to Cluster
Analysis. John Wiley.
Kesseler, M. (1995). A Schema-based Approach to HTML Authoring. In Proceedings of the 4th
International World Wide Web Conference WWW-4, Boston, 1995.
Kietz, J.-U. and Morik, K. (1994). A polynomial approach to the constructive induction of
structural knowledge. Machine Learning, 14(1):193-217.
Kietz, J.-U., Volz, R, and Maedche, A. (2000a). A method for semi-automatic ontology acqui-
sition from a corporate intranet. In EKAW-2000 Workshop "Ontologies and Text", Juan-Les-
Pins, France, October 2000.
Kietz, J.-U., Volz, R., and Maedche, A. (2000b). Semi-automatic ontology acquisition from
a corporate intranet. In International Conference on Grammar Inference (ICGI-2000), to
appear: Lecture Notes in Artificial Intelligence, LNAI.
Kifer, M., Lausen, G., and Wu, J. (1995). Logical Foundations of Object-Oriented and Frame-
Based Languages. Journal of the ACM, 42(4):741-843.
Klettke, M. (1998). Acquisition of Integrity Constraints in Databases. DISDBIS 51. infix, Sankt
Augustin, Germany. In German.
Klettke, M., Bietz, M., Bruder, I., Heuer, A., Priebe, D., Neumann, G., Becker, M., Bedersdorfer,
1., Uszkoreit, H., Maedche, A., Staab, S., and Studer, R (2001). GETESS - Ontologien, objek-
trelationale Datenbanken und Textanalyse als Bausteine einer semantischen Suchmaschine.
Datenbank-Spektrum, 1(1): 14-24.
Kramer, R, Nikolai, R, and Habeck, C. (1997). Thesaurus federations: loosely integrated the-
sauri for document retrieval in networks based on Internet Technologies. Journal of Digital
Libraries, 1(2): 112-131.
Kuehn, L. (2001). Human resource topic broker (in german). Master's thesis, University of
Karlsruhe.
236 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Kushmerick, N., Weld, D., and Doorenbos, R. (1997). Wrapper Induction for Information Ex-
traction. In UCAI-1997 - Proceedings ofthe 15th International Joint Conference on Artificial
Intelligence, Nagoya, Japan, August 23-29, 1997., pages 729-737, San Francisco. Morgen
Kaufmann.
Lacher, M. and Groh, G. (2001). Facilitating the exchange of explicit knowledge through ontology
mappings. In Proc.of 14th International FLAIRS conference, May 2001, pages 305-309.
AAAI Press.
Lamping, L., Rao, R., and Pirolli, P. (1995). A focus+context technique based on hyperbolic
geometry for visualizing large hierarchies. In Proceedings of the ACM SIGCHI Conference
on Human Factors in Computing Systems, pages 401-408.
Lang, S. and Lockemann, P. (1995). Datenbankeinsatz. Springer, Berlin.
Lassila, O. and Swick, R. (1999). Resource Description Framework (RDF) Model and Syntax
Specification. Technical report, W3C. W3C Recommendation.
http://www.w3.orgITRlREC-rdf-syntax.
Lee, L. (1999). Measures of distributional similarity. In Proceedings ofthe ACL '99, pages 25-32.
Levenshtein, I. V. (1966). Binary Codes capable of correcting deletions, insertions, and reversals.
Cybernetics and Control Theory, 10(8):707-710.
Linster, M. (1992). A review of sisyphus 91 and 92: Models of problem-solving knowledge. In
Aussenac, N., Boy, G., Gaines, B., Linser, M., Ganascia, J.-G., and Kordratoff, Y, editors,
Knowledge Acquisition for Knowledge-Based Systems, pages 159-182. Springer-Verlag.
Litkowski, K. (1978). Models of the semantic structure of dictionaries. Journal of Computational
Linguistics, 81 :25-74.
MacGregor, R. (1991). Inside the LOOM description classifier. SIGART Bulletin, 2(3):88-92.
Maedche, A., Neumann, G., and Staab, S. (2001 a). Bootstrapping an Ontology-Based Informa-
tion Extraction System. Intelligent Exploration of the Web, Series "Studies in Fuzziness and
Soft Computing", Springer.
Maedche, A., Schnurr, H.-P., Staab, S., and Studer, R. (2000). Representation language-neutral
modeling of ontologies. In Frank, editor, Proceedings ofthe German Workshop "Modellierung-
2000". Koblenz, Germany, April, 5-7, 2000. FOlbach-Verlag.
Maedche, A. and Staab, S. (2000a). Discovering conceptual relations from text. In ECAl-2000 -
European Conference on Artificial Intelligence. Proceedings ofthe 13th European Conference
on Artificial Intelligence. lOS Press, Amsterdam.
Maedche, A. and Staab, S. (2000b). Mining ontologies from text. In Proceedings of EKAW-2000,
Springer Lecture Notes in Artificial Intelligence (LNAI-1937), Juan-Les-Pins, France, 2000.
Springer.
Maedche, A. and Staab, S. (2000c). Semi-automatic engineering of ontologies from text. In Pro-
ceedings of the 12th Internal Conference on Software and Knowledge Engineering. Chicago,
USA, July, 5-7, 2000. KSI.
Maedche, A. and Staab, S. (2001 a). Learning ontologies for the semantic web. In WWW-ll
Workshop on the Semantic Web, Hong-Kong, 2001.
Maedche, A. and Staab, S. (2001b). On Comparing Ontologies. Technical report, Number 408,
Institute AIFB, Univ. of Karlsruhe.
Maedche, A. and Staab, S. (2001c). Ontology learning for the semantic web. IEEE Intelligent
Systems, 16(2).
Maedche, A., Staab, S., Nedellec, C., and Hovy, E., editors (2001 b). Proceedings of the IJ-
CAl'2001 Workshop on Ontology Learning - OL'2001.
Maedche, A., Staab, S., Stojanovic, N., Studer, R., and Sure, Y (2001c). SEmantic PortAL -
The SEAL approach. to appear in: Creating the Semantic Web. D. Fensel, J. Hendler, H.
Lieberman, W. Wahlster (eds.) MIT Press, MA, Cambridge.
Maedche, A., Staab, S., and Studer, R. (2001 d). Ontologien. Wirtschaftsinformatik,43(4).
REFERENCES 237
Maedche, A. and Volz, R. (2001). The Ontology Extraction & Maintenance Environment Text-
To-Onto. In Proceedings of the ICDM-2001 Workshop on the Integration of Data Mining and
Knowledge Management, San Jose, USA, November, 31, 2001.
Manning, C. and Schuetze, H. (1999). Foundations of Statistical Natural Language Processing.
MIT Press, Cambridge, Massachusetts.
Markert, K. and Hahn, U. (1997). On the interaction of metonymies and anaphora. In Proc. of
IJCAI-97, pages lO10-1015.
McGuinness, D., Fikes, R., Rice, 1., and Wilder, S. (2000). The Chimaera ontology environment.
In Proc. of AAAI-2000, pages 1123-1124.
Michalski, R. and Kaufmann, K. (1998). Data mining and knowledge discovery: A review of
issues and multi strategy approach. In Machine Learning and Data Mining Methods and
Applications. John Wiley, England.
Mikheev, A. and Finch, S. (1997). A workbench for finding structure in text. In Proceedings
of the Conference on Applied Natural Language Processing (ANLP-97), Washington D.C.,
i997.
Miller, G. (1996). The science of words. Freeman, New York.
Morgenstern, L. (1998). Inheritance comes of age: Applying nonmonotonic techniques to prob-
lems in industry. Artificial Intelligence, 103(1-2):237-271.
Mori, M. (1997). Finite-state transducers in language and speech processing. Computational
Linguistics, 23(3).
Morik, K. (1990). Integrating manual and automatic knowledge acquisition - BLIP. In Readings
in, Knowledge Acquisition: Current Practices and Trends. Ellis Horwood series in artificial
intelligence, Horwood, i990.
Morik, K. (1993). Balanced cooperative modeling. Machine Learning, 11:217-235.
Morik, K. and Brockhausen, P. (1996). A Multistrategy Approach to Relational Knowledge
Discovery in Databases. In Proceedings of the AAAi Workshop on Multistrategy Learning
(MSL-96), Palo Alto, 1996.
Morik, K., Wrobel, S., Kietz, 1.-U., and Emde, W. (1993a). Knowledge acquisition and machine
learning: Theory, methods, and applications. Academic Press, London.
Morik, K., Wrobel, S., Kietz, J.-U., and Emde, W. (1993b). Knowledge acquisition and machine
learning: Theory, methods, and applications. Academic Press, London.
Morin, E. (1999). Automatic acquisition of semantic relations between terms from technical
corpora. In Proc. of the Fifth International Congress on Terminology and Knowledge Engi-
neering - TKE'99.
MUC7 (1998). MUC-7 - Proceedings of the 7th Message Understanding Conference.
Mueller, H. A., Jahnke, J. H., Smith, D. B., Storey, M.-A., Tilley, S. R., and Wong, K. (2000).
Reverse Engineering: A Roadmap. In Proceedings of the 22nd international Conference on
Software Engineering (ICSE-2000), Limerick, Ireland. Springer.
Muggleton, S. (1992). Inductive Logic Programming. Academic Press.
Nedellec, C. and Causse, K. (1992). Knowledge Refinement using Knowledge Acquisition
and Machine Learning. In Proceedings of the European Knowledge Acquisition Workshop
(EKAW-92), 1992. Springer.
Nestorov, S., Abiteboul, S., and Motwani, R. (1997). Inferring Structure in Semistructured Data.
volume 26, pages 39-43.
Neuhaus, P. and Hahn, U. (1996). Trading of completeness for efficiency: The Parse Talk perfor-
mance grammer approach to real world text parsing. In FLAIRS-96: Proceedings of the 6th
Florida Artificial Intelligence Research Symposium, Key West, Florida, May 20-22.
Neumann, G:, Backofen, R., Baur, 1., Becker, M., and Braun, C. (1997). An information extrac-
tion core system for real world german text processing. In ANLP '97 - Proceedings of the
Conference on Applied Natural Language Processing, pages 208-215, Washington, USA.
238 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Noy, N. and Hafner, C. (1997). The State of the Art in Ontology Design - A Survey and Com-
parative Review. AI Magazine, 36(3).
Noy, N. F., Fergerson, R. w., and Musen, M. (2000). The knowledge model of protege-20oo:
Combining interoperability and flexibility. In Proceedings of the Conference on Knowledge
Acquisition and Management (EKAW-2000), Juan-Les-Pins, France, pages 17-32.
Noy, N. F. and Musen, M. A. (2000). PROMPT: Algorithm and Tool for Automated Ontology
Merging and Alignment. In Proceedings of the 17th National Con! on Artificial Intelligence
(AAAI'2000), Austin, Texas., pages 450-455. MIT Press/AAAI Press.
Noy, N. F. and Musen, M. A. (2001). Anchor-PROMPT: Using non-local context for semantic
matching. In Proceedings of the UCAI-2001 Workshop on Ontologies & Information Fusion.
Seattle, August 03, 2001.
Noy, N. F., Sintek, M., Decker, S., Crubezy, M., Fergerson, R. w., and Musen, M. A. (2001).
Creating semantic web contents with protege-2ooo. IEEE Intelligent Systems, 16(2).
Nwana, H. S. (1995). Software agents: An overview. Knowledge Engineering Review, 11(2):205-
244.
Ogden, C. and Richards, I. (1923). The Meaning ofMeaning: A Study ofthelnfluence ofLanguage
upon Thought and of the Science of Symbolism. Routledge & Kegan Paul Ltd., London, 10
edition.
Omelayenko, B. (2001). Learning of Ontologies for the Web: the Analysis of Existent Ap-
proaches. In Proc. of the International Workshop on Web Dynamics, London, UK, 2001.
Papazoglou, M., Proper, H., and Yang, J. (2000). Landscaping the information space of large
multi-database networks. In ????
Papazoglou, M. P., Proper, H. A., and Yang, J. (1995). Knowledge navigation in networked digital
libraries. Data and Knowledge Engineering, 36(3):251-281.
Peirce, C. (1885). On the Algebra of Logic. American Journal of Mathematics.
Pereira, F., Tishby, N., and Lee, L. (1993). Distributation Clustering of English Words. In Pro-
ceedings of the ACL-93, 1993, pages 183-199.
Pemelle, N., Rousset, M. C., and Ventos, Y. (2001). Automatic Construction and Refinement
of a Class Hierarchy over Semistructured Data. In Maedche, A., Staab, S., Nedellec, c.,
and Hovy, E., editors, Proceedings of UCAI-Ol Workshop on Ontology Learning OL-2001,
Seattle, August 2001, Menlo Park. AAAI Press.
Peterson, B., Andersen, w., and Engel, J. (1998). Knowledge bus: Generating application--
focused databases from large ontologies. In Proc of KRDB 1998, Seattle, Washington, USA,
pages 2.1-2.10.
Pinkerton, B. (1994). Finding What People Want: Experiences with the WebCrawler. In WWW2
- Proceedings of the 2nd International World Wide Web Conference, Chicago, USA, May,
October 17-20, 1994.
Pirlein (1995). Wiederverwendung von Common Sense Ontologien im Knowledge Engineering
(in german). PhD thesis, University of Karlsruhe.
Piskorski, 1. and Neumann, G. (2000). An intelligent text extraction and navigation system. In
Proceedings of the 6th Conference on Computer-Assisted Information Retrieval, Paris, 2000.
Rada, R., Mili, H., Bicknell, E., and Blettner, M. (1989). Development and application of a metric
on semantic nets. IEEE Transactions on Systems, Man, and Cybernetics, 19(1).
Ramanathan, S. and Hodges, J. (1997). Extraction of object-oriented structures from existing
relational databases. SIGMOD Record, 26(1):59-M.
Reimer, U. (1990). Automatic knowledge acquisition from texts: Learning terminological knowl-
edge via text understanding as inductive generalization. In Proceedings of the Workshop on
Knowledge Acquisiiton and Knowledge-based Systems (KAW-90), Banff, 1990.
REFERENCES 239
Rennie,1. and McCallum, A. (1999). Using Reinforcement Learning to Spider the Web Effi-
ciently. In Proceedings of the International Conference on Machine Learning (ICML-99),
1999.
Resnik, P. S. (1993). Selection and Information: A Class-based Approach to Lexical Relation-
ships. PhD thesis, University of Pennsylania.
Richardson, S. D. (1997). Determing Similarity and Inferring Relations in a Lexical Knowledge
Base. PhD thesis, City University of New York.
Romacker, M., Markert, K., and Hahn, U. (1999). Lean semantic interpretation. In Proc. of
IlCAl-99, pages 868-875.
Rousselot, F., Barthelemy, T., de Beuvron, F., Frath, P., and Oueslati, R. (1996). Terminolog-
ical competence and knowledge acquisition from texts. In Proceedings of the EKAW'I996
Workshop.
Sahuguet, A. and Azavant, F. (1999). Building light -weight wrappers for legacy web data-sources
using w4f. In VLDB '99, Proceedings of 25th International Conference on Very Large Data
Bases, September 7-10, 1999, Edinburgh, Scotland, UK, pages 738-741. Morgan Kaufmann.
Salton, G. (1988). Automatic Text Processing. Addison-Wesley.
Sanderson, M. and Croft, B. (1999). Deriving Concept Hierarchies from Text. In Proceedings
of the International Conference on Information Retrieval- SIGIR'99, August 1999, Berkley
CA, USA.
Schlobach, S. (2000). Assertional mining in description logics. In Proceedings of the 2000
International Workshop on Description Logics (DL2000), pages 89-97.
http://SunSITE.Informatik.RWTH-Aachen.DElPublications/CEUR -WSIV01-33/.
Schmitt, I. and Saake, G. (1997). Merging inheritance hierarchies for database integration. In
Proc. of the 3rd International Conference on Cooperative Information Systems - CooplS'98,
pages 322-331. IEEE Computer Science Press.
Scholze, J. and Woods, W. (1992). The kl-one family. F Lehmann (ed .. ), Semantic Networks in
Artificial Intelligence, Pergamon Press.
Schreiber, A. T. and Birmingham, W. P. (1996). The Sisyphus-VT initiative. International Journal
of Human-Computer Studies, 44(3/4).
Sheth, A. and Larsen, J. (1990). Federated database systems for managing distributed, hetero-
geneous and autonomous databases. ACM Computing Surveys, 22(3).
Skuce, D., Matwin, S., Tauzovich, B., Oppacher, F., and Szpakowicz, S. (1985). A logic-based
knowledge source system for natural language documents. Data and Knowledge Engineering,
1:201-231.
Spanoudakis, G. and Constantopoulos, P. (1994). Similarity for analogical software reuse: A
computational model. In Proc. of ECAl-1994, pages 18-22.
Sparck-Jones, K. and Willett, P., editors (1997). Readings in Information Retrieval. Morgan
Kaufmann.
Srikant, R. and Agrawal, R. (1995). Mining generalized association rules. In Proc. ofVLDB '95,
pages 407-419.
Srikant, R., Vu, Q., and Agrawal, R. (1997). Mining association rules with item constraints. In
Proceedings of the AAAI'97.
Staab, S., Angele, J., Decker, S., Erdmann, M., Hotho, A., Maedche, A., Studer, R., and Sure,
y. (2000a). Semantic Community Web Portals. In Proceedings of the 9th World Wide Web
Conference (WWW-9), Amsterdam, Netherlands.
Staab, S., Braun, C., Dtisterhoft, A., Heuer, A., Klettke, M., Melzig, S., Neumann, G., Prager,
B., Pretzel, J., Schnurr, H.-P., Studer, R., Uszkoreit, H., and Wrenger, B. (1999). GETESS-
searching the web exploiting german texts. In Proceedings of the 3rd Workshop on Cooper-
ative Information Agents, LNCS-1652, Berlin. Springer.
240 ONTOLOGY LEARNING FOR THE SEMANTIC WEB
Staab, S., Erdmann, M., and Maedche, A. (2001a). Engineering Ontologies using Semantic Pat-
terns. In Proceedings of the UCAI-2001 Workshop on E-Business & Intelligent Web. Seattle,
August 03,2001.
Staab, S., Erdmann, M., and Maedche, A. (2001b). From manual to semi-automatic seman-
tic annotation: About ontology-based text annotation tools. ETAI - Semantic Web Journal,
Linkoeping Electronic Articles, 16(1).
Staab, S., Erdmann, M., Maedche, A., and Decker, S. (2000b). An extensible approach for
modeling ontologies in RDF(S). In (Christophides and Plexousakis, 2(00).
Staab, S. and Maedche, A. (2000). Ontology engineering beyond the modeling of concepts and
relations. In Benjamins, v., Gomez-Perez, A., and Guarino, N., editors, Proceedings of the
ECAI-2000 Workshop on Ontologies and Problem-Solving Methods. Berlin, August 21-22,
2000.
Staab, S. and Maedche, A. (200 1). Knowledge portals - ontologies at work. A1 Magazine, 21 (2).
Staab, S., Maedche, A., Nedellec, c., and Hastings, P., editors (2000c). Proceedings of the
ECAl'2000 Workshop on Ontology Learning - OL'2000.
Staab, S. and Schnurr, H.-P. (2000). Smart task support through proactive access to organizational
memory. Journal of Knowledge-based Systems, 13(5). Special issue on AI and Knowledge
Management.
Staab, S., Schnurr, H.-P., Studer, R., and Sure, Y. (2001c). Knowledge processes and ontologies.
IEEE Intelligent Systems, 16(1).
Stumme, G. and Maedche, A. (2001a). Developing Federated Ontologies for the Semantic Web
using FCA-MERGE. In Proceedings of the UCA/-2001 Workshop on Ontologies & Infor-
mation Fusion. Seattle, August 03, 2001.
Stumme, G. and Maedche, A. (2001b). FCA-Merge: Bottom-Up Merging of Ontologies. In
UCAI-2001 - Proceedings of the 17th International Joint Conference on Artificial Intelli-
gence, Seattle, USA, August, 1-6,2001, San Francisco. Morgen Kaufmann.
Stumme, G., Taouil, R., Bastide, Y., Pasquier, N., and Lakhal, L. (2000). Fast computation of
concept lattices using data mining techniques. In Proc. KRDB '00, Berlin, 2000. CEUR-
Workshop Proc.,
http://sunsite.inforrnatik.rwth-aachen.de/Publications/CEUR-WS/.
Sure, Y., Maedche, A., and Staab, S. (2000). Leveraging corporate skill knowledge - From ProPer
to OntoProper. In Mahling, D. and Reimer, U., editors, Proceedings ofthe Third 1nternational
Conference on Practical Aspects of Knowledge Management. Basel, Switzerland, October
30-31, 2000.
http://www.research.swisslife.chlpakm2000/.
Swartout, B., Patil, R., Knight, K., and Russ, T. (1996). Toward distributed use of large-scale
ontologies. In Proc. the 10th International Workshop on Knowledge Acquisition, Modeling
and Mangement (KAW'96), Banff, Canada, November 9-14, 1996.
Szpakowicz, S. (1990). Semi-automatic acquisition of conceptual structure from technical texts.
International Journal of Man-Machine Studies, 33(4):385-397.
Tari, Z., Bukhres, 0., Stokes, 1., and Harnmoudi, S. (1998). The Reengineering of Relational
Databases based on Key and Data Correlations. In Proceedings of the 7th Conference on
Database Semantics (DS-7), 7-10 October 1997, Leysin, Switzerland. Chapman & Hall.
Uschold, M. and Gruninger, M. (1996). Ontologies: Principles, methods and applications. Knowl-
edge Sharing and Review, 11(2):93-155.
Uschold, M. and King, M. (1995). Towards a Methodology for Building Ontologies. In Proceed-
ings of the IJCAl'95 Workshop on Basis Ontological Issues in Knowledge Sharing.
van Heijst, G. (1995). The Role ofOntologies in Knowledge Engineering. PhD thesis, Universiteit
van Amsterdam.
REFERENCES 241
van Heijst, G., Schreiber, A. T., and Wielinga, B. J. (1997). Using explicit ontologies for kbs
development. International Journal of Human-Computer Studies, 42(2): 183-292.
Vanderwende, L. H. (1995). The Analysis of Noun Sequences using Semantic Informnation
Extractedfrom On-Line Dictionaries. PhD thesis, Georgetown University.
Visser, P. and Tamma, V. (1999). An Experience with Ontology Clustering for Information
Integration. In Proceedings of the IJCAI'99 Workshop on Intelligent Information Integration,
Stockholm, Sweden, July 31, 1999.
Volz, R. (2000). Akquisition von Ontologien mit Text-Mining-Verfahren (in german). Master's
thesis, University of Karlsruhe.
Wang, K. and Liu, H. (1997). Schema Discovery for Semi-Structured Data. In Proceedings of
the Third International Conference on Knowledge Discovery and Data Mining (KDD-97),
Newport Beach, California, USA, August 14-17, 1997. AAAI Press.
Wang, K. and Liu, H. (1998). Discovering typical structures of documents: A road map approach.
In Proc of SIGlR-98.
Webb, G. (1996). Integrating Machine Learning with Knowledge Acquisition through direct
interaction with domain experts. volume 9, pages 236--252, Berlin. Springer Verlag.
Webb, G., Wells, J., and Zheng, Z. (1999). An Experimental Evaluation of Integrating Machine
Learning with Knowledge Acquisition. Machine Learning, 35(1):5-23.
Wei, F. (1999). F-logic Semantics and Implementation of Internet Metadata. Master's thesis,
University of Freiburg.
Weinstein, P. (1990). Integrating Ontological Metadata: algorithms that predict semantic com-
patibility. PhD thesis, Computer Science and Engineering Department, University of Michi-
gan, Pittsburgh, USA.
Weinstein, P. and Birmingham, W. (1999). Comparing concepts in differentiated ontologies. In
Proc. of KAW-99, Banff, Canada, 1999.
Welty, C. and Ide, N. (1999). Using the right tools: enhancing retrieval from marked-up docu-
ments. Journal Computers and the Humanities, 33(10):59-84.
Wersig, G. (1985). Thesaurus-Leitfaden. K.G.Saur Verlag KG, MUnchen.
Wiederhold, G. (1992). Mediators in the architecture offuture information systems. IEEE Com-
puter, 25(3):38-49.
Wiemer-Hastings, P., Graesser, A., and Wiemer-Hastings, K. (1998). Inferring the meaning of
verbs from context. In Proceedings of the Twentieth Annual Conference of the Cognitive
Science Society.
Wittgenstein, L. (1922). Tractuatus Logico-Philisophicus. Routledge & Kegan Paul Ltd., Lon-
don.
Wrobel, S. (1994). Concept Formation and Knowledge Revision. Kluwer, Dordrecht, Boston,
London.
Yangarber, R., Grishman, R., Tapanainen, P., and Huttunen, S. (2000). Automatic Acquisition
of Domain Knowledge for Information Extraction. In Proceedings of COLlNG '2000, Saar-
brucken, Germany, 2000.
Index