112
Matt Duckham and Mike Worboys
generated is not explicitly addressed by these systems. This chapter contains further
information on mediator systems, in particular on the distinction between local-asview (LAV) and global-as-view (GAV) approaches to mediation.
The theoretical foundations of most mediator systems are similarly focused more
on ontology integration than ontology alignment (e.g. [33, 62]). Formal concept analysis (FCA [29]) is a widely used technique, dating back more than 20 years, for
structuring and integrating ontologies based on concept lattices (a special type of ordering relation on categories). Although the integration of concept lattices has a precise formal definition and can be automated, human domain expertise is required to
identify salient attributes and categories within a domain and their interrelationships.
Another widely used technique for representing and reasoning about heterogeneous
ontologies is description logics. Description logics are decidable, tractable fragments
of first-order predicate calculus, and form the basis of many mediator systems and
components, including SIMS, OBSERVER, and OWL (Web ontology language, one
of the primary standards for the semantic Web). Description logics are especially
useful in the context of ontology integration because they provide several important reasoning services, most notably a subsumption service which can classify the
relationships between categories and derive a complete and consistent integrated ontology (see [11, 19] for more information on description logics and their reasoning
services). However, while description logics can offer efficient formally founded reasoning about ontologies, defining the relationships and rules that connect elements
of different ontologies remains a primarily human activity.
In summary, much of the existing research into mediators and ontology integration does not address the issue of ontology alignment directly. Instead, such research
typically assumes that the semantic relationships between ontological elements will
be established using user interaction, predefined mappings, preexisting top-level ontologies, or existing lexical correspondences [66]. Building these mappings is assumed to require an understanding of the underlying concepts, and so is at root a
human activity.
6.2.2 Automating Ontology Alignment
Some researchers have turned their attention to creating semi-automated or
fully automated ontology alignment systems (see [54] for an overview). Much of
this research adopts an intensional approach: it aims to analyze the definitions (intensions) of the concepts and categories used in the input information sources.
Intensional techniques usually analyze heterogeneous ontologies to identify lexical similarities (e.g. PROMPT [52], Active Atlas [61]), structural similarities (e.g.
DIKE [53], ONION [50]), or some combination of these (e.g. CUPID [45], FCAMerge [60]).
There are two main drawbacks of adopting a purely intensional approach to ontology alignment. First, how concepts are defined is not necessarily the same as how
they are used. As an analogy, people who learn to speak language from a dictionary
(definitions) often have very different speech patterns from native speakers, who also
learn from example. Only by looking at extensional information (specific instances
Matt Duckham and Mike Worboys
generated is not explicitly addressed by these systems. This chapter contains further
information on mediator systems, in particular on the distinction between local-asview (LAV) and global-as-view (GAV) approaches to mediation.
The theoretical foundations of most mediator systems are similarly focused more
on ontology integration than ontology alignment (e.g. [33, 62]). Formal concept analysis (FCA [29]) is a widely used technique, dating back more than 20 years, for
structuring and integrating ontologies based on concept lattices (a special type of ordering relation on categories). Although the integration of concept lattices has a precise formal definition and can be automated, human domain expertise is required to
identify salient attributes and categories within a domain and their interrelationships.
Another widely used technique for representing and reasoning about heterogeneous
ontologies is description logics. Description logics are decidable, tractable fragments
of first-order predicate calculus, and form the basis of many mediator systems and
components, including SIMS, OBSERVER, and OWL (Web ontology language, one
of the primary standards for the semantic Web). Description logics are especially
useful in the context of ontology integration because they provide several important reasoning services, most notably a subsumption service which can classify the
relationships between categories and derive a complete and consistent integrated ontology (see [11, 19] for more information on description logics and their reasoning
services). However, while description logics can offer efficient formally founded reasoning about ontologies, defining the relationships and rules that connect elements
of different ontologies remains a primarily human activity.
In summary, much of the existing research into mediators and ontology integration does not address the issue of ontology alignment directly. Instead, such research
typically assumes that the semantic relationships between ontological elements will
be established using user interaction, predefined mappings, preexisting top-level ontologies, or existing lexical correspondences [66]. Building these mappings is assumed to require an understanding of the underlying concepts, and so is at root a
human activity.
6.2.2 Automating Ontology Alignment
Some researchers have turned their attention to creating semi-automated or
fully automated ontology alignment systems (see [54] for an overview). Much of
this research adopts an intensional approach: it aims to analyze the definitions (intensions) of the concepts and categories used in the input information sources.
Intensional techniques usually analyze heterogeneous ontologies to identify lexical similarities (e.g. PROMPT [52], Active Atlas [61]), structural similarities (e.g.
DIKE [53], ONION [50]), or some combination of these (e.g. CUPID [45], FCAMerge [60]).
There are two main drawbacks of adopting a purely intensional approach to ontology alignment. First, how concepts are defined is not necessarily the same as how
they are used. As an analogy, people who learn to speak language from a dictionary
(definitions) often have very different speech patterns from native speakers, who also
learn from example. Only by looking at extensional information (specific instances