Semantic Integration: A Survey Of Ontology-Based

Approaches
Natalya F. Noy
Stanford Medical Informatics
Stanford University
251 Campus Drive, Stanford, CA 94305
[email protected]

ABSTRACT pressed in a language that can be used for reasoning. These

Semantic integration is an active area of research in several features of ontologies underscore the main trends that dis-
disciplines, such as databases, information-integration, and tinguish semantic-integration research in the ontology com-
ontologies. This paper provides a brief survey of the ap- munity: First, since the underlying goal of ontology devel-
proaches to semantic integration developed by researchers in opment is to create artifacts that different applications can
the ontology community. We focus on the approaches that share, there is an emphasis on creating common ontologies
differentiate the ontology research from other related areas. that can then be extended for more specific domains and ap-
The goal of the paper is to provide a reader who may not plications. If these extensions refer to the same top-level on-
be very familiar with ontology research with introduction to tology, the problem of integrating them can be greatly alle-
major themes in this research and with pointers to different viated. Second, since ontologies are developed for use with
research projects. We discuss techniques for finding corre- reasoning engines and semantics of ontology languages are
spondences between ontologies, declarative ways of repre- specified with reasoning in mind, inference and reasoning
senting these correspondences, and use of these correspon- takes center stage in ontology-integration approaches.
dences in various semantic-integration tasks Ontologies have gained popularity in the AI community as
a means for establishing explicit formal vocabulary to share
1. ONTOLOGIES AND SEMANTIC INTE- between applications. Therefore, one can say that one of the
goals of using ontologies is not to have the problem of het-
GRATION
erogeneity at all. It is of course unrealistic to hope that there
Researchers and practitioners in the fields of databases will be an agreement on one or even a small set of ontologies.
and information integration have produced a large body of While having some common ground either within an appli-
research to facilitate interoperability between different sys- cation area or for some high-level general concepts could
tems. This research ranges from techniques for matching alleviate the problem of semantic heterogeneity, we will still
database schemas to answering queries using multiple sources need to map between ontologies, whether they extend the
of data. Ontology research is another discipline that deals same top-level ontology or are developed independently.
with semantic heterogeneity in structured data. We refer So, what are the types of differences between ontologies?
the reader to another article in this issue [24] for a detailed In part summarizing earlier surveys, Klein [16] categorizes
discussion on the uses of ontologies, their differences from different types of mismatches between ontologies. The first
database schemas, and challenges in semantic integration class of mismatches are mismatches at the language level—
that the ontology community faces. The goal of this paper mismatches in expressiveness and semantics of ontology lan-
is to discuss the major thrusts of approaches to semantic in- guage. The languages can differ in their syntax, but, more
tegration produced by various projects in the ontology com- important, constructs available in one language (e.g., stat-
munity and to provide readers with pointers to sources for ing that classes are disjoint) are not available in another.
additional information. We will focus on the approaches that Even semantics of the same language primitives could be
highlight the use of ontologies, their emphasis on knowledge different (e.g., whether declaration of multiple ranges of a
sharing, and their use in reasoning. Note that this paper does property have union or intersection semantics). The normal-
not attempt to provide a comprehensive review of the state of ization process therefore often precedes ontology-matching
the art in using ontologies for semantic integration. We refer [15] and translates source ontologies to the same language,
the reader to an excellent and thorough review by Kalfoglou resolving these differences.
and Schorlemmer [15] for that purpose. However, even for ontologies expressed in the same lan-
While there are many definitions of what an ontology is guage, possible ontology-level mismatches abound. A par-
[26], the common thread in these definitions is that an on- tial list of ontology-level mismatches includes using the same
tology is some formal description of a domain of discourse, linguistic terms to describe different concepts; using differ-
intended for sharing among different applications, and ex-

SIGMOD Record, Vol. 33, No. 4, December 2004 65

ent terms to describe the same concept; using different mod- becoming accepted standards. The explicit goal of these on-
eling paradigms (e.g., using interval logic or points for tem- tologies is to have domain-specific ontologies extend them,
poral representation); using different modeling conventions thus providing the grounding in common vocabulary for these
and levels of granularity; having ontologies with differing ontologies. Note that this scenario is different from the tra-
coverage of the domain, and so on. ditional information-integration scenario where the global
We discuss three dimensions of semantic-integration re- schema—the common view on different schemas to be integ-
search in this paper: rated—is usually developed after the schemas themselves
Mapping discovery: Given two ontologies, how do we find are developed and its design is therefore guided by the in-
similarities between them, determine which concepts dividual schemas to be integrated. The implication of this
and properties represent similar notions, and so on. difference is that in the information-integration scenario the
Declarative formal representations of mappings: Given two global schema is only general enough to provide access to
ontologies, how do we represent the mappings between all the schemas that it integrates. The common top-level or
them to enable reasoning with mappings. reference ontology is usually more general since it needs to
Reasoning with mappings: Once the mappings are defined, encompass the top level for ontologies yet to be developed.
what do we do with them, what types of reasoning are Two of the ontologies that are built specifically with the
involved? purpose of being formal top-level ontologies are the Sug-
gested Upper Merged Ontology (SUMO) [19] and DOLCE
In the rest of this paper, we explore these dimensions.
[8]. SUMO is an effort by the IEEE Standard Upper Ontol-
2. DISCOVERING MAPPINGS ogy Working Group aimed at developing “a standard upper
ontology that will promote data interoperability, information
Many researchers agree that one of the major bottleneck search and retrieval, automated inferencing, and natural lan-
in semantic integration is mapping discovery. There are sim- guage processing.” The SUMO ontology defines such high-
ply too many ontologies and database schemas available and level concepts Object, ContinousObject, Process, Quantity,
they are too large to have manual definition of correspon- Relation, and so on, providing axioms in first-order logic that
dences as the primary source of mapping discovery. Fur- describe properties of these concepts and relations among
thermore, in the world where software agents will roam the them. Similarly, the DOLCE ontology is a formal foun-
(semantic) web, they will need to map structures they know dational ontology developed as a top-level ontology in the
about to new structures they come across on-the-fly. Hence, WonderWeb project, which comprises a large number of Eu-
the task of finding mappings (semi-) automatically has been ropean research groups. The goal of DOLCE is to provide
an active area of research in both database and ontology a common reference framework for WonderWeb ontologies
communities [22, 15]. to facilitate sharing of information among them. In its repre-
We identify two major architectures for mapping discov- sentation, DOLCE aims at capturing “ontological categories
ery between ontologies. For the first approach, recall that underlying natural language and human common-sense.”
the goal of ontologies is to facilitate knowledge sharing. As While many researchers hope that domain- and application-
a result, ontologies are often developed with the explicit goal specific ontologies will reuse the foundational ontologies,
of providing the basis for future semantic integration. Here, like SUMO and DOLCE, and that such reuse will indeed fa-
the vision is that a general upper ontology is agreed upon by cilitate semantic interoperation between applications based
developers of different applications, who then extend this on these ontologies, we do not yet have enough experience
general ontology with concepts and properties specific to reports with such approaches to claim it a success. There
their applications. As long as this extension is performed are reports on both the successes [21] and difficulties [25] of
in a way consistent with the definitions in the shared ontol- such reuse. The Workshop on Core Ontologies in Ontology
ogy, finding correspondences between two extensions can Engineering 1 in October 2004 will discuss both successful
be facilitated by this common “grounding.” The second set and unsuccessful cases and best practices on reusing foun-
of approaches comprises heuristics-based or machine learn- dational ontologies for specifying domain content.
ing techniques that use various characteristics of ontologies, There are also implemented semantic-integration tools that
such as their structure, definitions of concepts, instances of exploit the idea that if two ontologies extend the same ref-
classes, to find mappings. These approaches are similar to erence ontology in a consistent way, then finding correspon-
approaches to mapping XML schemas or other structured dences between their concepts is easier. For example, the
data (e.g., Cupid [17]) but tend to rely more heavily on fea- Process Specification Language (PSL) [11], developed at the
tures of concept definitions or on explicit semantics of these National Institute for Standards and Technology, is an on-
definitions. tology that is endorsed as an International Standard within
2.1 Using a Shared Ontology the International Organization of Standardisation (ISO). PSL
was designed to “facilitate correct and complete exchange of
A number of very general ontologies formalizing notions process information among manufacturing systems such as
such as processes and events, time and space, physical ob-
jects, and so on, are being developed and some of them are www.loa-cnr.it/core_onto.html

66 SIGMOD Record, Vol. 33, No. 4, December 2004

scheduling, process modeling, [and] process planning” [12]. names that are similar to each other. The second consider-
The designers of PSL have developed it as an interlingua for ation are the words used in natural-language definitions of
ontologies representing these different process. All theories concepts. The matcher compares the number and the ratio
within the PSL ontology have been verified with respect to of shared words in the definitions to find definitions that are
the intended semantics of their terminology. Grüninger and similar. An experimentally determined formula for combin-
Kopena [12] developed an integration architecture with the ing these measures of similarity yields potential matchers
PSL ontology at the center and mappings between ontolo- that the user needs to examine and approve.
gies for specific manufacturing processes and the PSL ontol- The P ROMPT system [20] was originally developed to sup-
ogy. The mappings are defined semi-automatically by pre- port ontology merging, guiding users through the process
senting ontology developers with a set of questions (in nat- and suggesting which classes and properties can be merged.
ural language) helping them to map terms in their process- It records the mappings identified both by the system and
specific ontology to the terms in PSL. The system then gen- by the user during merging to create a declarative mapping
erates two-way mappings between the task-specific ontol- specification between source ontologies. To make sugges-
ogy, such as scheduling and the PSL interlingua. Note that tions, P ROMPT uses a mixture of lexical and structural fea-
the generation of these mappings is defined formally and is tures, as well as input from the user during an interactive
not based on heuristics. These mappings can be composed merging session to find the mappings. For instance, if a
to provide mappings between any task-specific ontologies. user said that two classes in two source ontologies are the
same (should be merged), then P ROMPT analyzed the prop-
2.2 Using Heuristics and Machine-learning erties of these classes, their subclasses and superclasses to
look for similarities of their definitions and suggest addi-
It is certainly helpful to have ontologies that we need to tional correspondences. Another algorithm in the toolset–
match to refer to the same upper ontology or to conform to A NCHOR P ROMPT [20]—treats an ontology as a graph with
the same reference ontology. However, we often do not have classes as nodes and slots as links. The algorithm analyzes
this “luxury” and need to create mappings between ontolo- the paths in the subgraph limited by the anchors and deter-
gies that perhaps use the same specification language but do mines which classes frequently appear in similar positions
not have any vocabulary beyond the specification language on similar paths. These classes are likely to represent se-
in common. Most researchers agree that automatic mapping mantically similar concepts.
between ontologies in this context is beyond our grasp at the Recently, the W3C has approved a standard for represent-
moment, but many techniques have produced good results. ing ontologies on the Semantic Web—the OWL language. 2
Heuristic-based approaches to ontology mapping are simi- Acceptance of a standard encouraged researchers to propose
lar to heuristic-based approach to matching database schemas algorithms that rely more heavily on features of the ontology
and XML structures ([22, 17]) and use lexical and structural language to compare ontologies. For example, a similar-
components of definitions to find correspondences. How- ity metric between concepts in OWL ontologies developed
ever, ontology-based approaches often go further, exploiting by Euzenat and Volchev [7] is a weighted combination of
semantics of relationships in ontologies, such as, for exam- similarities of various features in OWL concept definitions:
ple, the semantics of the subclass-of or part-of relationships, their labels, domains and ranges of properties, restrictions
attachment of property to a class, domain and range defini- on properties (such as cardinality restrictions), types of con-
tions for properties, and so on. Ontologies usually have a lot cepts, subclasses and superclasses, and so on.
more constraints specified than database schemas do, and the FCA-Merge [23] is a method for comparing ontologies
methods for finding mappings automatically tend to exploit that have a set of shared instances or a shared set of docu-
this larger number of constraints ments annotated with concepts from source ontologies. Based
We will start by reviewing several ontology-mapping tools on this information, FCA-Merge uses techniques from For-
and then summarize the different ontology features that they mal Concept Analysis [9] to produce a lattice of concepts
use. We would like to emphasize again that this paper presents which relates concepts from the source ontologies. The al-
only a sampling of such tools to give examples of different gorithm suggests equivalence and subclass–superclass rela-
approaches. Please see a paper by Kalfoglou and Schorlem- tions. An ontology engineer can then analyze the result and
mer [15] for a comprehensive review. use it as a guidance for creating a merged ontology.
Hovy [13] describes a set of heuristics that researchers at The IF-Map [14] system identifies mappings automatically
ISI/USC used for semi-automatic alignment of domain on- based on the theory of information flow [1]. Given two on-
tologies to a large central ontology. Their techniques are tologies, IF-Map generates a logic infomorphism—a map-
based mainly on linguistic analysis of concept names and ping between ontologies that is based on the above confor-
natural-language definitions of concepts. (There is a limited mance. The system then uses the channel theory to infer
use of taxonomic relationships as well). First, the matcher the mappings between different local ontologies using these
uses natural-language–processing techniques to split composite- logic infomorphisms.
word names (a common occurrence in concept names). It


then compares substrings of different lengths to find concept http://www.w3.org/TR/owl-features/

SIGMOD Record, Vol. 33, No. 4, December 2004 67

 GLUE [5] is an example of a system that employs machine- as a single theory by a theorem prover optimized for ontology-
learning techniques to find mappings. GLUE uses multiple translation task. The theorem prover runs either in forward-
learners exploiting information in concept instances and tax- chaining or backward-chaining mode depending on the task
onomic structure of ontologies. GLUE uses a probabilistic at hand.
model to combine results of different learners. The learn- Several researchers use ontologies themselves to represent
ers that GLUE uses currently relies on ontologies having in- mappings declaratively, as instances in an ontology. The
stances and they work much better if many slot values have mapping ontology by Crubézy and colleagues [4] or the Se-
text in them rather than references to other instances. mantic Bridge Ontology of the MAFRA framework [18], for
Researchers have also addressed the issue of finding com- instance, define the structure of specific mappings and the
plex mappings, such as determining that a concepts in one transformation functions to transfer instances from one on-
ontology is a specialization of a concept in another ontology. tology to another. This ontology can then be used by tools
For example, Giunchiglia and Shvaiko [10] start by ground- to perform the transformations. Such an ontology usually
ing their source ontologies in WordNet terms but then run provides different ways of linking concepts from the source
a SAT prover on the mappings to determine other types of ontology to the target ontology, transformation rules to spec-
mappings (such as generalization, specialization or disjoint- ify how values should be changed, and conditions and effects
ness): the authors reformulate the matching problem as that of such rules. Then a mapping between two ontologies con-
of propositional satisfiability. stitutes a set of instances of classes in the mapping ontology
To summarize, the tools for automatic and semi-automatic and can be used by applications to translate data from the
ontology alignment use the following features in ontology source ontology to the target. The mapping ontology men-
definitions (to various extent): tioned above [4], for example, provides declarative means
for defining many-to-one or many-to-many aggregation re-
¯ concept names and natural-language descriptions lationships between concepts in the source and target on-
¯ class hierarchy (subclass–superclass relationships) tologies, as well as one-to-many concept-decomposition re-
¯ property definitions (domains, ranges, restrictions) lations. It allows specification of recursive mappings, com-
plex mappings between that collect information from several
¯ instances of classes related concepts, and other mechanisms.
¯ class descriptions (as in DL-based tools). Finally, researchers also used views to define mappings
between ontologies, similar to defining mappings in infor-
3. REPRESENTATIONS OF MAPPINGS mation integration, both in global-as-view (GAV) and local-
as-view (LAV) setting. The OIS framework [2] is a good
While developing tools for automatic and semi-automatic example of such approach. In OIS, a global ontology is used
ontology matching is a large thrust of semantic-integration to provide access to local ontologies. Both global and local
research in the ontology community, it is definitely not the ontologies are defined using Description Logics. The map-
only one. The higher expressive power of ontology lan- pings are defined as views over either the global or the local
guages provides the opportunity for representing mappings ontologies. In other words, a predicate from one ontology
themselves in more expressive terms. Mappings between el- is defined as a query (and DL expression) over predicates in
ements in schemas are usually expressed either as queries another ontology.
and views or as pairs of related terms. We generally find
a larger spectrum of the ways mapping between ontologies
are expressed. We will discuss several representations of 4. WE HAVE THE MAPPINGS. NOW WHAT?
mappings here: representing mappings as instances in an on- Naturally, defining the mappings between ontologies, ei-
tology of mappings; defining bridging axioms in first-order ther automatically, semi-automatically, or interactively, is
logic to represent transformations; and using views to de- not a goal in itself. The resulting mappings are used for vari-
scribe mappings from a global ontology to local ontologies. ous integration tasks: data transformation, query answering,
In the OntoMerge system [6] developed for semantic inte- or web-service composition, to name a few.
gration on the Semantic Web, authors use a general-purpose Given that ontologies are often used for reasoning, it is
inference engine to enable translation between mapped on- only natural that many of these integration tasks involve rea-
tologies. In OntoMerge the correspondence between two soning over the source ontologies and the mappings. For
ontologies is expressed as a set of bridging axioms relating example, the OntoMerge system mentioned earlier [6] uses
classes and properties of the two source ontologies. The vo- reasoning to perform several tasks related to ontology trans-
cabulary of the two ontologies are in different XML names- lation. The first task is translating instances that conform to
paces, so the bridging axioms are essentially translation rules one ontology (the source) to instances conforming to another
referring to concepts from source ontologies and specifying ontology (the target), given the mapping between the source
how to express for example a class in one ontology by col- and target. To perform this task, OntoMerge first creates a
lecting information from classes in another. The two source merged ontology that includes the source, the target, and the
ontologies, together with the bridging axioms are then treated mapping and performs inference on this merged ontology.

68 SIGMOD Record, Vol. 33, No. 4, December 2004

Afterwards, OntoMerge performs a projection step, where integration problem is also a common approach.
it retains only the new conclusions reached that exclusively The two communities can certainly share and reuse the
reference the target vocabulary. techniques that they have developed in their respective do-
The second task that OntoMerge deals with—generating mains. In fact, there has been a certain convergence trend
ontology extensions—is more specific to the area of ontolo- where schema-matching approaches for example employ more
gies. Consider for example, two ontologies describing Web expressive components of schema definitions in their tech-
services: OWL-S3 and WSDL4 . Suppose we have defined niques. On the other hand, ontology researchers are pay-
a mapping between these two ontologies. Suppose also that ing more attention to the experience of the database commu-
we have an ontology describing ticket-purchasing web servi- nity. We believe that such cross-fertilization will improve
ces—a domain-specific extension of the OWL-S ontology. semantic-integration solutions in both fields.
This ticket-purchasing ontology creates subclasses of some Finally, the emerging Semantic Web can prove to be an
of the classes in OWL-S, fills in some of the property val- excellent testbed for scalability of various approaches and a
ues, and so on. In other words, it extends the OWL-S on- common ground for experimenting with hybrid approaches.
tology. If we have a mapping between OWL-S and WSDL, Most researchers agree that semantic integration is one of the
OntoMerge can automatically generate a WSDL description most serious challenges for the Semantic Web today. On the
of ticket-purchasing—an extension of the WSDL ontology. one hand, the premise of the Semantic Web is that the use
Note that this case is different from data translation since of machine-interpretable ontologies defined in formal lan-
we are dealing with subontologies rather than instances con- guages amenable to reasoning will provide the next genera-
forming to ontologies. In both of these tasks, OntoMerge tion of services. On the other hand, the scale of the Semantic
uses forward-chaining reasoner to perform the translation. Web will certainly require well-tested approaches from the
Several tools process representation of mappings as in- database community.
stances of the mapping ontology by Crubézy and colleagues
discussed in the previous section [4] to perform various inte- 6. ACKNOWLEDGMENTS
gration tasks. First, a mapping interpreter uses the instances
in the mapping ontology to translate data from the source I would like to thank AnHai Doan, Michael Grüninger,
and Yannis Kalfoglou for thoughtful and helpful comments
ontology to the target ontology. Second, the P ROMPT tool
on earlier drafts at this paper.
for ontology merging [20], also mentioned earlier in the pa-
per, can take the mapping ontology as its input and merge
the ontologies based on the mapping. These tools are ex- REFERENCES
tensions to the Protégé ontology-development environment 5 [1] J. Barwise and J. Seligman. Information Flow: The
thus providing an integrated framework for ontology devel- Logic of Distributed Systems. Cambridge University
opment, knowledge acquisition, and semantic integration. Press, 1997.
In the OIS framework [2], ontologies are expressed in De- [2] D. Calvanese, G. Giacomo, and M. Lenzerini.
scription Logics and therefore it is natural that DL reason- Ontology of integration and integration of ontologies.
ers are used to answer queries in the data-integration frame- In Description Logic Workshop (DL 2001), pages
work. The authors address the general task of answering 10–19, 2001.
queries posed in terms of the global ontology using the data [3] D. Calvanese, G. D. Giacomo, M. Lenzerini,
in the local ontologies. However, while even in expressive R. Rosati, and G. Vetere. DL-lite: Practical Reasoning
Description Logics, computing certain answers to queries is for Rich DLs. In International Workshop on
decidable, it may often be intractable. In recent research, the Description Logics (DL2004), Whistler, Canada, 2004.
authors have explored less expressive subsets of Description
[4] M. Crubézy and M. A. Musen. Ontologies in support
Logics [3], making this type of query answering tractable.
of problem solving. In S. Staab and R. Studer, editors,
Handbook on Ontologies, pages 321–342. Sringer,
5. CONCLUDING REMARKS 2003.
As this brief survey shows, many issues that ontology re- [5] A. Doan, J. Madhavan, P. Domingos, and A. Halevy.
searchers in semantic integration grapple with are very simi- Learning to map between ontologies on the semantic
lar to the issues that database and information-integration re- web. In The Eleventh International WWW Conference,
searchers have been addressing. Some of the approaches are Hawaii, US, 2002.
also similar although the ontology community relies more [6] D. Dou, D. McDermott, and P. Qi. Ontology
heavily on the higher expressive power of ontology languages translation on the semantic web. In International
and on reasoning techniques. With ontologies, using a com- Conference on Ontologies, Databases and
mon upper ontology or reference ontology to alleviate the Applications of Semantics, 2003.

http://www.daml.org/services/owl-s/1.0/ [7] J. Euzenat and P. Valtchev. Similarity-based ontology

http://www.w3.org/TR/wsdl alignment in OWL-Lite. In The 16th European

http://protege.stanford.edu Conference on Artificial Intelligence (ECAI-04),

SIGMOD Record, Vol. 33, No. 4, December 2004 69

 Valencia, Spain, 2004. Workshop on Applications of Ontologies and Problem
[8] A. Gangemi, N. Guarino, C. Masolo, and Solving Methods, ECAI’98, Brighton, England, 1998.
A. Oltramari. Sweetening wordnet with DOLCE. AI [22] E. Rahm and P. A. Bernstein. A survey of approaches
Magazine, 24(3):13–24, 2003. to automatic schema matching. VLDB Journal, 10(4),
[9] B. Ganter and R. Wille. Formal Concept Analysis: 2001.
Mathematical foundations. Springer, [23] G. Stumme and A. Mädche. FCA-Merge: Bottom-up
Berlin-Heidelberg, 1999. merging of ontologies. In 7th Intl. Conf. on Artificial
[10] F. Giunchiglia, P. Shvaiko, and M. Yatskevich. Intelligence (IJCAI ’01), pages 225–230, Seattle, WA,
Semantic matching. In 1st European semantic web 2001.
symposium (ESWS’04), pages 61–75, Heraklion, [24] M. Uschold and M. Grüninger. Ontologies and
Greece, 2004. semantics for seamless connectivity. SIGMOD
[11] M. Grüninger. A guide to the ontology of the process Record, 33(3), 2004.
specification language. In S. Staab and R. Studer, [25] A. Valente, T. Russ, R. MacGrecor, and W. Swartout.
editors, Handbook on Ontologies. Sringer, 2003. Building and (re)using an ontology for air campaign
[12] M. Grüninger and J. Kopena. Semantic integration planning. IEEE Intelligent Systems, 14(1):27–36,
through invariants. In A. Doan, A. Halevy, and 1999.
N. Noy, editors, Workshop on Semantic Integration at [26] C. Welty. Ontology research. AI Magazine, 24(3),
ISWC-2003, Sanibel Island, FL, 2003. 2003.
[13] E. Hovy. Combining and standardizing largescale,
practical ontologies for machine translation and other
uses. In The First International Conference on
Language Resources and Evaluation (LREC), pages
535–542, Granada, Spain, 1998.
[14] Y. Kalfoglou and M. Schorlemmer. IF-Map: an
ontology mapping method based on information flow
theory. Journal on Data Semantics, 1(1):98–127, Oct.
2003.
[15] Y. Kalfoglou and M. Schorlemmer. Ontology
mapping: the state of the art. The Knowledge
Engineering Review, 18(1):1–31, 2003.
[16] M. Klein. Combining and relating ontologies: an
analysis of problems and solutions. In IJCAI-2001
Workshop on Ontologies and Information Sharing,
pages 53–62, Seattle, WA, 2001.
[17] J. Madhavan, P. Bernstein, and E. Rahm. Generic
schema matching using Cupid. In The 27th
International Conf. on Very Large Data Bases (VLDB
’01), Rome, Italy, 2001.
[18] A. Maedche, B. Motik, N. Silva, and R. Volz.
MAFRA - a mapping framework for distributed
ontologies. In 13th European Conference on
Knowledge Engineering and Knowledge Management
EKAW, Madrid, Spain, 2002.
[19] I. Niles and A. Pease. Towards a standard upper
ontology. In The 2nd International Conference on
Formal Ontology in Information Systems
(FOIS-2001), Ogunquit, Maine, 2001.
[20] N. F. Noy and M. A. Musen. The PROMPT suite:
Interactive tools for ontology merging and mapping.
International Journal of Human-Computer Studies,
59(6):983–1024, 2003.
[21] S. Polyak, J. Lee, M. Gruninger, and C. Menzel.
Applying the process interchange format(PIF) to a
supply chain process interoperability scenario. In

70 SIGMOD Record, Vol. 33, No. 4, December 2004