Ontology alignment is the process of finding related entities in different ontologies. In this context, precise and explicit representation of conceptualizations is essential for reaching semantic integration to ensure that only data related to the same (or sufficiently similar) real-world entity are merged. Foundational ontologies describe general concepts independent of a domain and precisely define meta-properties so as to make the semantics of each concept in the ontology explicit. In this paper we show how the use of OntoUML, a conceptual modeling language based on Unified Foundational Ontology, allows the application of alignment patterns and exemplify how this approach may improve precision, recall and refine the type of the alignment.
An essential issue for reaching semantic integration is the precision of an explicit
conceptualization representation.
In this paper we show how the use of OntoUML, based on some design patterns
explored in
Another contribution of this paper is a review in the classification of alignment
approaches proposed by
This paper is structured as follows. Section 2 presents some OntoUML design patterns that explore constraints underlying UFO. In section 3 we discuss the ontology alignment process and present a review in the classification of ontology alignment approaches. Section 4 introduces the alignment patterns based on OntoUML design patterns. In section 5 we exemplify the application of these alignment patterns. Section 6 reviews related works, followed by the conclusions in the section 7.
Foundational ontologies (also called upper-level or top-level ontologies) describe very
general concepts, which are independent of a particular problem or domain
UFO is one example of foundational ontology that has been developed based on
a number of theories from Formal Ontology, Philosophical Logics, Philosophy of
Language, Linguistics and Cognitive Psychology
OntoUML is a conceptual modeling language designed to comply with the ontological distinctions and axiomatic theories put forth by UFO that results from a redesign process of the Unified Modeling Language (UML). The OntoUML classes, for example, make explicit the distinctions between an object and a process, types of things from their roles, among others.
A fundamental distinction in UFO is between particulars and universals. Particulars are entities that exist in reality possessing a unique identity, while universals are patterns of features, which can be realized in a number of different particulars.
UML class diagrams are intended to represent the static structure of a domain, in
which classes typically represent endurant universals. The UML profile proposed by
Substantials are entities that persist in time while keeping their identity (as opposed to events such as a business process or a birthday party). Constructs that represent Sortal Universals can provide a principle of identity and individuation for its instances. Mixin Universal is an abstract metaclass that represents the general properties of all mixins, i.e., non-sortals (or dispersive universals). A type is rigid iff for every instance of that type, it is necessarily an instance of that type. In contrast, a type is antirigid iff for every instance of the type, there is always a possible world in which it is not an instance of this type.
A kind (and subkinds) represent rigid sortals that applies necessarily to its instances, i.e., in every possible world (such as a Person, Man or Woman). A phase represents an anti-rigid sortal instantiated in a specific world or time period, but not necessarily in all of them (such as Child, Adolescent and Adult phases of a Person). A role defines an anti-rigid sortal which may be assumed in a world, but not necessarily in all possible worlds (such as a Student or a Professor role played by a Person), but once it is, this depends on its participation in a specific relation or event. Due to space restrictions, we will not define all other OntoUML categories. The design patterns presented in the next section are limited to these primitives: kind/subkind, phases and roles.
The design patterns presented in this section were explored by
Subkinds can be manifested as a part of a generalization set which has as a common superclass a Kind S. In this case, the subkind classes are disjoint and complete. The Subkind Design Pattern is illustrated in figure 2(a).
Phases are always manifested as part of a phase partition (PP). In a PP there is always one unique root common supertype which is necessarily a Kind S. As well as subkinds, phases are manifested as a part of a generalization set of type S. The Phase Design Pattern is illustrated in figure 3(a).
Roles represent (possibly successive) specializations of a Kind S by using a relational specialization condition R with another type T of the model. The Role Design Pattern is illustrated in figure 4(a).
Ontology alignment is the process of finding corresponding entities (concept, relation,
or instance) in two ontologies describing the same domain. A general ontology
alignment function based on the vocabulary, E, of all terms e Є E, based on the set of
possible ontologies, O, and based on possible alignment relations, M, is a partial
function: align: E x O x O → E x M. Apart from one-to-one equality alignments, mostly
investigated in existing work, one entity often has to be aligned not only to equal
entities, but based on another relation (e.g., subsumption). Further, there are complex
composites such as a concatenation of terms (e.g., name equals first plus last name)
Precision and recall are commonplace measures in information retrieval and are also applied to evaluate alignment results. Precision measures the correctness of the method by the ratio of correctly found correspondences over the total number of returned correspondences. Recall is a completeness measure and considers the ratio of correctly found correspondences over the total number of expected correspondences.
The classification of
If the classification is read in ascending it focus on the kinds of manipulated objects: strings (terminological), structure (structural), models (semantics) or data instances (extensional).
The approach proposed in this paper fits the classification described as "Upper level, domain specific ontologies", which is an element-level technique based on external semantic input. This classification groups approaches based on domain specific ontologies used as external sources of background knowledge of the particular domain being aligned and those ones that actually exploit foundational ontologies as external sources of common knowledge. Although both cases involve the use of an external ontology, their role in the alignment process is very different. In this paper we propose a dissociation of these inputs in two techniques, as illustrated in figure 6.
Considering this new classification, our approach is instantiated in the “Upper level ontologies” technique. In section 6 we will present the related work instantiated in this classification.
Considering the design patterns presented in section 2 it is possible to derive some alignments patterns given below.
Rule 1: The alignment (equivalence) of a <<kind>> class C1 of an ontology A to a <<kind>> class C2 of an ontology B is possible if all <<subkind>> classes of C1 have a corresponding class that is also a <<subkind>> of C2 in ontology B (a one-to-one equivalence is not required). Rule 2: The alignment (specialization) of a <<kind>> class C2 of an ontology B to a <<kind>> class C1 of an ontology A is possible if some of the <<subkind>> classes of C1 have equivalent classes in ontology B, and these equivalences cover all the <<subkind>> classes of C2 in ontology B. Rule 3: The alignment of a <<kind>> class C1 of an ontology A to a <<kind>> class C2 of an ontology B is not possible if at least one <<subkind>> class of C1 in ontology A is not aligned to a <<subkind>> class of C2 in ontology B.
Besides the distinct semantics of Subkind and Phase classes, since both are manifested as a part of a disjoint and complete generalization set which has as a common superclass a Kind S, the rules of the Phase Design Pattern are analogous to the rules of the Subkind alignment pattern that were previously presented.
Rule 4: The alignment (equivalence) of a <<role>> class C1 of an ontology A to a <<role>> class C2 of an ontology B is only possible if the <<kind>> rigid class that the <<role>> classes specialize, and the relation its instantiation depends on, are aligned to each other.
Is this section we will exemplify the application of the alignment patterns presented in section 4. The ontologies describe the domain of organizing conferences, which corresponds to one track of the Ontology Alignment Evaluation Initiative (OAEI) 20111. The ontologies and the reference alignments indicated by the initiative are available on the Conference Track2.
A prerequisite for application of the alignment patterns is that the ontologies to
be aligned must comply with the UFO constraints set out by
We have analyzed the alignment of two ontologies identified in Table 1 by considering the submitted alignments results of tools evaluated in group 1, which consists of best evaluated matchers of the track. We will explore two common errors committed by the four matchers of this group, one affecting the precision and other the recall measure. * Ontologies have been based upon actual conference (series) and its web pages
All four matchers have identified a correspondence between the classes Iasted::Document and SigKdd::Document that is not indicated by the reference alignment (which harms the precision). The fragments are illustrated in figure11.
In this case, both Document classes are of the type kind and correspond to a generalization set of subkind classes, disjoint and complete.
140 49 Dotted lines indicate the expected alignment between the classes of these ontologies fragments. The set of matchers considered have identified the correspondence between Iasted::Review and SigKdd::Review, besides the wrong correspondence between the Document classes. In this context, the application of Rule 3 would reject the correspondence between Iasted::Document and SigKdd::Document since some of their subkind classes are not aligned.
The other common error is that all four matchers could not identify the reference alignment between Iasted::Sponsor and SigKdd::Sponzor indicated by the reference alignment (thus reducing recall). The fragments are illustrated in figure 12. In this example, both Sponsor and Sponzor classes are stereotyped as roles. In Iasted ontology a Sponsor is a role played by a person that gives some Sponzorship. In SigKdd ontology a Sponzor is a role characterized by the payment of a Sponzor_fee. However, the kind class specialized by this role is not explicit in this ontology. Based on Role Design Pattern, we define this role as a specialization of the class Person, already defined in this ontology and aligned to the class Iasted::Person. With this redesign, the rigid kind classes that these roles specialize are aligned to each other, which itself brings additional information to allow the identification of the alignment between Sponsor and Sponzor classes. However, by Rule 4, to guarantee the alignment between Iasted::Sponsor and SigKdd::Sponzor, Iasted::Sponzorship and SigKdd::Sponzor_fee must be aligned, which suggests an update in the reference alignment to include this equivalence. Otherwise, this correspondence would be rejected.
In figure 13 there are fragments of other two ontologies of the same conference domain, identified in Table 2. * Ontologies have been based upon actual software tool for conference organisation support
The question identified here is not an error that affects the precision or recall, but a review of the type of the correspondence identified.
In this case, both Document classes are of the type kind and correspond to a generalization set of disjoint and complete subkind classes, as the first example. The dotted lines indicate the reference alignments that mean equivalence relation. Since part of <<subkind>> classes of Edas ontology have equivalent classes in Cmt ontology, and these equivalences cover all the <<subkind>> classes of Cmt ontology, the alignment between the classes Document could be refined considering that Cmt::Document is a specialization of Edas::Document.
One main point that has guided the development of the approach presented in
The approach presented in this paper, in turn, proposes a directly use of foundational ontologies to improve semantic integration by considering some alignment patterns based on meta-properties of the OntoUML constructs, with the determination of rules to be applied during the alignment process. Considering the suggested dissociation of the “Upper level” and “Domain specific ontologies” it is instantiated in the “Upper level ontologies” technique.
Other works address foundational ontologies in the context of ontology
alignment but they are more directly related to the use of domain ontologies to support
the alignment of other ontologies on the same domain. In
Ontology alignment is an active research area and some challenges consider semantic
issues. In this paper we discussed how the use of OntoUML oriented by some design
patterns improves the ontological adequacy of the ontologies being aligned and allows
the application of some rules based on alignment patterns. We have used some
ontologies from the main initiative for evaluation of ontology alignment to demonstrate
how the design patterns and the alignment patterns may improve precision, recall and
refine the type of the alignment, with a manually performed example. However, the
process of annotation the classes with the correct OntoUML stereotypes can be assisted
by a software tool
Another contribution of the paper is a review in the classification of
Future work includes formalization of indicative and restrictive rules based on the meta-properties of a larger set of constructs of OntoUML to be applied during the alignment process. Moreover, the automatization of the proposal and its application in complete scenarios will also be considered.