Beruflich Dokumente
Kultur Dokumente
Ontologies for
Enterprise Knowledge
Management
Alexander Maedche, Boris Motik, Ljiljana Stojanovic, Rudi Studer, and Raphael Volz,
Research Center for Information Technology, University of Karlsruhe
information, including knowledge management and enterprise-wide OKMSs is still in the early stages.
e-business applications. Next-generation knowledge First, despite much research on ontology represen-
Several challenges management systems will likely rely on conceptual tation, engineering, and reasoning, features such as
models in the form of ontologies to precisely define scalability, persistency, reliability, and transactions—
exist related to the meaning of various symbols. For example, standardized and widely adopted in classical data-
FRODO (a Framework for Distributed Organiza- base-driven information systems—are typically not
applying ontologies tional Memories) uses ontologies for knowledge available in ontology-based systems. So, using
description in organizational memories,1 CoMMA requirements analysis from several applications, we
in real-world (Corporate Memory Management through Agents) recently developed a conceptual-modeling approach4
investigates agent technologies for maintaining suitable for business-wide applications. Our design
environments. The ontology-based knowledge management systems,2 goals were to express and access ontology-based
and Steffen Staab and his colleagues have discussed conceptual models in a natural and easily under-
authors present an the methodologies and processes for building ontol- standable way, with a small gap between the model
ogy-based systems.3 conceptualization and its implementation. At the
integrated enterprise- Here we present an integrated enterprise-knowl- same time, we wanted to realize an enterprise-wide
edge management architecture for implementing an ontology-based system using existing and well-
knowledge ontology-based knowledge management system established technologies such as relational databases.
(OKMS). We focus on two critical issues related to Second, a large body of information in an enter-
management working with ontologies in real-world enterprise prise typically already exists outside the knowledge
applications. First, we realize that imposing a single management system—for example, in other applica-
architecture, focusing ontology on the enterprise is difficult if not impossi- tions such as groupware, databases, and file systems.
ble. Because organizations must devise multiple To reuse information from these sources, we must
on how to support ontologies and thus require integration mechanisms, integrate the sources and the OKMS, which typically
we consider means for combining distributed and het- involves creating wrappers that lift these sources’con-
multiple ontologies erogeneous ontologies using mappings. Additionally, tent to the ontology level—not an easy task.5
a system’s ontology often must reflect changes in sys- Third, providing metadata is a time-consuming
and manage ontology tem requirements and focus, so we developed guide- and difficult process that users tend to avoid. So, we
lines and an approach for managing the difficult and must develop ways to easily assign fine-grained
evolution. complex ontology-evolution process. metadata (based on ontologies) to the different
resources available in the enterprise. This would
Bringing ontologies to real-world require enhancing tools used daily (such as text
enterprise applications processors) so that they could extract as much meta-
For several reasons, the development of real-world data as possible semiautomatically.
pings. Two execution modes are provided: assumptions made when the system was built. types of change generation in OKMSs.11 In
static and virtual. The static-execution mode For example, acquiring a new subsidiary in top-down change generation, the knowledge
transforms the source ontology’s instances an enterprise adds new business areas and officer or end user explicitly defines the
once and then stores them in the associated functionalities to the existing system. requirements for ontology changes. These
knowledge management system—changes changes cover business strategy evolution,
to instances in the source ontology are not modification in the application domain, new
visible in the mapped ontology. user needs, additional functionality, and so
In the virtual-execution mode, the map- forth and are captured in a variety of ways:
ping-ontology instance transforms every Knowledge management systems direct discussion or interviews, customer
query from the target ontology into a query specifications, surveys, or observations. Alter-
over the source ontology. The system then generally are not developed to natively, some changes might be discovered
executes the transformed query and trans- by analyzing log files that track interaction of
forms the obtained information back into the remain stable. Rather, several users with the system, which is known as bot-
target ontology. With this approach, changes tom-up change generation.
to the source ontology’s instances are imme-
diately visible in the target ontology. How-
factors make them subject to We might also view the processing of
these changes as ontology evolution, which
ever, because the system performs a trans- we can define as timely adaptation of the
formation at each request, performance is
continual change. ontology to changing requirements and the
worse than in the static case. consistent propagation of changes to the
dependent artifiacts. Modifying one part of
Postprocessing the ontology might generate subtle inconsis-
Postprocessing applies only to static exe- Second, user requirements often change tencies in other parts of the same ontology,
cution, where the goal is to improve the after the system has been built, warranting in the instances, dependent ontologies, and
results of the execution phase. For example, system adaptation. For example, hiring new applications. For example, Michel Klein and
it deals with the problem of object identity employees might lead to new competencies his colleagues12 discuss this variety of causes
(identifying, for example, that “A. Maedche” and greater diversity in the enterprise, which and consequences of the ontology changes.
and “Alexander Mädche” are the same per- the system must reflect. So, we observe that ontology evolution is a
son). For virtual execution, this phase is not Third, some changes in the domain are complex operation that should be realized as
applicable for two main reasons. First, the implicit and can be discovered only by ana- an organizational process (see Figure 3). Fur-
system performs the execution only on small lyzing user interaction with the system. For thermore, ontology evolution is clearly dis-
subsets of information, so it is not possible example, let’s say many users are interested tinguished from the problem of schema evo-
to apply complex algorithms that require in two topics in conjunction (for example, lution in relational databases, which has
access to entire source and target ontologies. “debugging” and “Java”) and no knowledge already been extensively studied.13,14 For
Second, the virtual mapping’s performance is resource matches this criterion. In this case, example, Natalya Noy and her colleagues
already critical, and introducing another an efficient knowledge management system have discussed these differences and identi-
phase would further aggravate it. should signal that it needs a knowledge fied that they stem from different knowledge
resource about the combination of these top- models and usage paradigms of ontologies,
Managing evolving ontologies ics (for example, a document on how to contrary to relational databases.15
Knowledge management systems gener- debug Java code). Our ontology-evolution process comprises
ally are not developed to remain stable. Ad hoc management of the changes in the following six phases.
Rather, several factors make them subject to knowledge management systems might work
continual change. in the short term, but to avoid unnecessary Representation
First, the environment in which the knowl- complexity and failures in the long run, man- To resolve changes, we must identify and
edge management systems operate can agement must be interpreted at the concep- represent them in a suitable format. Elemen-
change unpredictably, thereby invalidating tual level. Therefore, we can identify two tary changes in the ontology (such as adding
a concept, removing a property, or setting a When a user modifies the ontology, ontol- of some change and thus approve a change
property range) are derived from our con- ogy instances must change to preserve con- that shouldn’t be performed. Moreover,
ceptual model. However, this granularity of sistency with the ontology. An ontology changing the ontology for experimental pur-
ontology changes is not always appropriate. update might also corrupt ontologies that poses might be desirable. To enable recover-
Often, the intent of a change might be better depend on the modified ontology and, con- ing from these situations, we introduce the
expressed at a higher level. sequently, all artifacts based on these ontolo- validation phase. Validation concerns an
For example, the user might need to move gies. We can solve this problem recursively ontology’s truthfulness regarding its prob-
a concept form one parent to another. He or by applying ontology evolution process to lem domain—does the ontology correctly
she might bring the ontology into the desired these ontologies. However, apart from syn- represent reality and user requirements? It
state through a successive application of a tax inconsistency, semantic inconsistency lets the user acknowledge performed changes
list of elementary evolution changes (such as can also arise when, for example, the depen- and request that they be undone.
“remove subConceptOf” and “add subCon- dent ontology already contains a concept that Reversibility means undoing a change’s
ceptOf” elementary changes). However, the is added to the original ontology. effects, which might not be the same as sim-
system might perform unnecessary changes When an ontology changes, applications ply requesting an inverse change manually.
if it applies each change alone. To avoid that based on the changed ontology might not For example, if a user deletes a concept from
drawback, it should be able to express work correctly anymore. An ontology evo- a concept hierarchy, the subconcepts will
changes on a coarser level, with the intent of lution approach must recognize which either need to be deleted as well or attached
making a change directly visible. So, we to the root concept or the deleted concept’s
introduce composite changes (such as “move parent. Reversing such a change is not equal
concept”) representing a group of elemen- to recreating the deleted concept—we also
tary changes applied together. need to revert the concept hierarchy to the
Changing an ontology can induce original state. Creating evolution logs typi-
Semantics of change cally solves the problem of reversibility. An
Changing an ontology can induce incon- inconsistencies in other parts of evolution log tracks information about each
sistencies in other parts of the ontology. change, allowing the reconstruction of the
Semantic inconsistency arises if an ontology the ontology. Semantic sequence of changes leading to the ontol-
entity’s meaning changes. Syntactic incon- ogy’s original state.
sistency arises if undefined entities are used
at the ontology or instance level or ontology
inconsistency arises if an ontology Discovery
model constraints are invalidated. For exam-
ple, removing a concept that is the only ele-
entity’s meaning changes. The validation phase results in an ontol-
ogy that, although consistent, either might
ment of a domain set for some property contain redundant entities or could be better
results in syntactic inconsistency. We can structured with respect to the domain. For
resolve that problem by requesting a new example, multiple users might work on dif-
change in the ontology, which can induce change in the ontology can affect the func- ferent parts of an ontology without enough
new problems that cause new changes and so tionality of dependent applications and react communication. They could delete subcon-
on. If an ontology is large, fully compre- correspondingly. cepts of some concept at different times to
hending each induced change’s extent and fulfill their immediate needs. Thus, it is pos-
meaning might be difficult. Implementation sible that only one subconcept might remain.
The semantics-of-change phase aims to To avoid performing undesired changes, Because classification with only one subclass
resolve induced changes systematically, before the system applies a change to the violates the classification’s original purpose,
ensuring the consistency of the whole ontol- ontology, it should generate and present to we consider such an ontology to have a sub-
ogy. To help in better understanding the effects the user a list of all implications affecting the optimal structure. To help users detect such
of each, this phase should contribute to the ontology. The user should be able to com- situations, we investigated applying self-
maximum transparency, providing detailed prehend the list and approve or cancel the adaptive-system principles and proactively
insight into each change. For each change in changes. Once approved, the system could making suggestions for ontology refine-
the ontology, it is possible to generate a set of perform changes by successively resolving ments—changes to the ontology with the
additional changes, leading to different final changes from the list. Because it is necessary goal of improving it. This would make mak-
consistent states. Evolution strategies are to perform several changes together, they ing the ontology easier to understand and
mechanisms used in this phase that let users must be performed within a transaction. If cheaper to modify. We have identified three
customize ontology evolution according to the user cancels the changes, the ontology ways to discover changes.
their needs.11 Consequently, users can trans- should remain intact. Structure-driven change discovery exploits
fer the ontology in the desired consistent state. a set of heuristics to improve an ontology on
Validation the basis of the analysis of the ontology’s
Propagation When working on an ontology collabora- structure. For example, if all subconcepts
This phase should automatically bring all tively, different parties might have different have the same property, the property may be
dependent elements to a consistent state after ideas about how to change it. Furthermore, moved to the parent concept.
the system has updated the ontology. one party might fail to understand the effect Data-driven change discovery detects the
KAON—The Karlsruhe ontology changes applied to the ontology leave the The Engineering Server implementation
and Semantic Web framework ontology in a consistent state and prevents uses relational databases for ontology per-
We based the Ontologging OKMS’s core illegal changes. sistence. We call it the Engineering Server
ontology-management component—the • Change reversibility tracks ontology because the server is optimized for ontology
ontology server—on the Karlsruhe ontology changes in an evolution log so that the sys- engineering, where creating and deleting
and Semantic Web framework (http://kaon. tem can reverse them at a user’s request. concepts are common. So, the Engineering
semanticweb.org). KAON is an open-source The evolution log is based on the evolu- Server has a fixed number of tables in the
ontology-management infrastructure tar- tion ontology (http://kaon.semanticweb. schema, rather than allocating one table per
geted for semantics-driven business applica- org/ontos/evolution.kaon) that models concept for storing the concept’s extension.
tions, developed and maintained at the FZI what happens and by whom, and why, We have heavily optimized and tested the
(Research Center for Information Technolo- when, and how it happens. server on an ontology consisting of 100,000
gies) and AIFB (Institute of Applied Infor- • Concurrency conflict detection detects and concepts, 66,000 properties, and 1,000,000
matics and Formal Description Methods) at resolves conflicts resulting in concurrent instances, where loading related information
the University of Karlsruhe. It includes a updates from different users. For example, about 20 ontology entities takes less than 3
comprehensive tool suite allowing easy ontol- if one user updates the ontology, then the seconds, while deleting a concept in the mid-
ogy management and application. KAON system must notify the update’s other dle of the concept hierarchy takes less than 5
focuses on integrating traditional technolo- active users. Alternatively, if a user seconds. It works on a typical single-proces-
gies for ontology management and applica- attempts to update the ontology using stale sor desktop computer running Windows XP
tion with those typically used in business information, the system must detect the with 256 Mbytes of RAM.
applications, such as relational databases. It is conflict. To achieve atomicity, the system The Legacy Server implementation lets us
based on an ontology model,4 derived as an performs all updates within a transaction. lift existing relational databases to the ontol-
extension of RDF Schema, with some pro- • Optimized loading loads and transports ogy level. To do so, tables in the database
prietary extensions (such as inverse, sym- several ontology entities to the client in must be mapped to concepts and relation-
metric, and transitive relations), cardinalities, one request, thus significantly improving ships in the ontology. Based on these map-
modularization, metamodeling, and repre- system performance. pings, the system must modify all operations
sentation of lexical information. Figure 4 pre- • Query answering locates the ontology’s on the KAON API according to these map-
sents the KAON architecture. elements according to given criteria and is pings into appropriate database operations.
The KAON architecture’s core is its ontol- the key to providing scalable systems. The virtual mapping engine is an imple-
ogy API (KAON API), consisting of a set of mentation realizing a virtual ontology mapped
interfaces for access to ontology entities. For KAON API has several implementations, from some source ontology. It transforms
example, there are Concept, Property, and each using a different back end for informa- queries into the mapped ontology using an
Instance interfaces, which contain methods tion storage. One implementation of the instance of the mapping ontology (see http://
for accessing ontology concepts, properties, KAON API is based on the RDF API and kaon.semanticweb.org/ontos/mapping.kaon).
and instances, respectively. The API incor- thus allows access to RDF repositories. Various applications have been realized on
porates other elements required for ontology Although it offers capabilities for accessing top of the KAON API. Within Ontologging,
management: remote RDF repositories, such as RDF the KAON portal presents and browses
Server, it primarily manages local RDF ontologies on the Web. The OI Modeler real-
• The evolution strategy ensures that all ontologies stored as files. izes a graph-based user interface for ontol-
ogy creation and evolution. Figure 5 shows practicality by applying it to project man- 3. S. Staab et al., “Knowledge Processes and
how we’ve implemented ontology evolution agement. From that we hope to elicit the Ontologies,” IEEE Intelligent Systems, vol.
at the user level. The figure shows a modeling technical requirements that are crucial for 16, no. 1, Jan./Feb. 2001, pp. 26–34.
session where the user attempted to remove successfully applying ontologies in knowl- 4. B. Motik, A. Maedche, and R. Volz, “A Con-
the Person concept. Before applying this edge management. ceptual Modeling Approach for Building
change to the ontology, the system computes Semantics-Driven Enterprise Applications,”
Proc. 1st Int’l Conf. Ontologies, Databases,
the set of additional changes that must be and Application of Semantics (ODBASE-
applied to keep the consistency. It presented Acknowledgments 2002), Springer-Verlag, 2002, pp. 1082–1099.
the tree of dependent changes to the user, thus
We thank our colleagues and students at the FZI 5. M.T. Roth and P. Schwarz, “Don’t Scrap It,
letting the user comprehend the change’s (Research Center for Information Technologies) Wrap It! A Wrapper Architecture for Legacy
effects before applying it. Only when the user and AIFB (Institute of Applied Informatics and Data Sources,” Proc. 23rd Int’l Conf. Very
agrees will the system apply the changes. Formal Description Methods) at the University of Large Databases, Morgan Kaufmann, 1997,
Karlsruhe. This research has profited from fruit- pp. 266–275.
ful discussions with our Ontologging project part-
ners: Archetypon (Greece), Deltatec (Belgium), 6. G. Wiederhold, “Mediators in the Architec-
Indra (Spain), Insead (France), and Meta4 (Spain). ture of Future Information Systems,” Com-
T h e A u t h o r s
Alexander Maedche is department manager of the Knowledge Management
Research department at the FZI (Research Center for Information Technologies),
University of Karlsruhe. His research interests cover knowledge discovery in data
and text, ontology management and learning, and using ontologies in enterprise
applications and the Semantic Web. He received a Diploma in industrial engi-
neering and his PhD in applied informatics, both from the University of Karl-
A vertical
sruhe. He is a member of the IEEE and Gesellschaft für Infomatik. Contact him
at the FZI Research Center for Information Technology, Univ. of Karlsruhe, Haid-
fill here
und-Neu-Str. 10-14, 76131 Karlsruhe, Germany; maedche@fzi.de.