When the size of an ontology increases, it becomes hard to be managed. Ontology view extraction is an approach that can be used for overcoming the challenges that arise in this scenario. In this context, an ontology view is a subset of an ontology tailored to a specific set of user requirements. Well-founded ontology views were recently proposed as ontology views that follow well-founded ontological principles, which ensures some desirable ontological properties. In this paper, we propose a novel approach for extracting well-founded ontology views, which is more flexible than the previous approach. We also present a method for evaluating the quality of approaches for extracting ontology views. We apply this method for demonstrating that our novel approach produces ontology views that are more accurate than those produced by the previous approach. We illustrate our approaches using a domain ontology for Petrography.
Ontologies tend to evolve over time by incorporating new knowledge. The resulting ontology can lead to a scenario of information overload, where the information exceeds the cognitive capability of the users. Ontology views have been adopted as a solution for overcoming this scenario, since they are extracted from a base ontology according to specific user criteria, and provide only the knowledge that is relevant for a given task at hand.
The literature provides some approaches for extracting
ontology views
In this paper, we propose a new approach for
extracting WFOVs, which modifies the basic approach defined in
In Section , we provide an overview of the main approaches available in the literature for extracting portions of ontologies. In Section , we present a basic definition of the notion of well-founded ontology view and describe the basic approach for extracting WFOVs. Section presents our approach for extracting WFOVs. Section describes the method that we used for evaluating our approach. Section describes the application of the different approaches for extracting WFOVs in a case scenario with their corresponding evaluations. Finally, Section presents our conclusions.
In general, the literature provides two main approaches that
can be used for extracting manageable portions of
ontologies. The extraction of ontology modules
Some of these approaches
Algorithm 1 The basic approach for WFOV extraction. Require: Well-Founded Ontology procedure SEL(Ob, tConcepts, tRelations, So) So.C So.C [ tConcepts So.R So.R [ tRelations newC ; newR ; for all c 2 tConcepts do conservesT AX(Ob, c, newC, newR) conservesQUA(Ob, c, newC, newR) conservesIP (Ob, c, newC, newR) conservesED(Ob, c, newC, newR) conservesRD(Ob, c, newC, newR) conservesF R(Ob, c, newC, newR) conservesP R(Ob, c, newC, newR) newC newC So.C newR newR So.R end for if newC 6= ; then SEL(Ob, newC, newR, So) else if newR 6= ; then So.R So.R [ newR end if end if end procedure
In
In Table 1, we present a comparison of the approaches
discussed in this section. The approaches are identified as: 1
In this Section, we present the approach proposed by
The approach proposed by
One of the main categories of universals provided by UFO is Substantial Universal, whose instances are individuals that, in general, are existentially independent of all other individuals. Some of its instances can be existentially dependent when they are considered inseparable parts of their hosts. Sortal Universals are substantial universals that provide or carry some principle of identity (PI) for their instances. In this context, a PI is the principle that supports the judgment whether two instances of the universal are the same.
Another important ontological meta-property used by UFO is the rigidity. A certain universal is rigid when its extension (set of all particulars) is the same in all possible worlds. That is, an instance of a rigid universal cannot cease to be an instance of it without ceasing to exist. For example, Person can be viewed as a rigid universal, since persons cannot cease to be persons without ceasing to exist; meanwhile all instances of Student (which is an anti-rigid universal) can still exist (as persons) if they cease to be students.
Within the sortal universals, UFO includes three distinct types of substance sortals, which are rigid sortals that provide their own principle of identity: Kind, which represents functional complexes (Person, Dog, Chair, etc); Collective (Swarm, Forest, etc), which represents collectives; and Quantity, which represents objectified portions of matter (Wine, Water, Gold, etc). Besides that, Subkind is a rigid sortal that does not provide its own PI, but carries a principle of identity that is supplied by a given substance sortal.
UFO also defines two anti-rigid sortals: Roles and Phases. Phases are universals that constitute possible stages in the history of a substance sortal. Phases are relationally independent, since they depend solely on intrinsic properties. For example, Baby, Toddler, Kid, Teenager and Adult are considered phases of Human. On the other hand, Roles are relationally dependent, since they depend on extrinsic (relational) properties. This is the case, for example, when we say that for an instance of person to be considered a Student, she must be enrolled at an educational institution.
Other substantial universals do not have the properties of sortals; they are dispersive universals. This is the case, for example, of Categories, which are rigid universals that do not provide or carry a PI for their instances. Categories represent essential properties that are common to all instances of many disjoint universals that provide distinct PIs. Rational agent is an example of Category, since it abstracts an essential property (namely, the rationality) of instances of Person and Artificial Agent, which are disjoint universals, with distinct PIs. Role Mixins, on the other hand, are anti-rigid universals that do not provide and do not carry a PI for their instances. They can be viewed as generalizations of roles of different substance sortals. For example, Customer is a role mixin that generalizes Personal Customer, which is a role of Person; and Corporate Customer, which is a role of Organization. Finally, Mixins are universals that do not provide and do not carry a PI for their instances and that are semi-rigid; that is, they have some instances that are necessarily their instances, but they also have some instances that are only contingently their instances. They usually generalize rigid and anti-rigid universals. For example, Seatable Object is a mixin that generalizes Chair, which is a rigid universal; and Solid Crate, which is an anti-rigid universal (actually, it is a phase of a Crate, which can also be a Broken Crate).
On the other hand, Moment Universals are Universals whose instances are existentially dependent individuals that inhere in other individuals. Some moment universals depend existentially on a single entity. This is the case of Quality Universals and Modes. Quality Universals represents the properties in the conceptual models. A Quality Universal characterizes other Universals and is related to Quality Structures, that is, a structure that represents a set of all values that a quality can assume. Thus, considering the property Color as a Quality Universal, a given instance of Car could be characterized by an instance of Color, which is associated with a value (called quale) in the ColorStructure, which represents all the possible values that the property Color can assume. On the other hand, Modes are universals whose instances are existentially dependent individuals, and that are not associated to Quality Structures. Examples of modes are Skill, Belief, Headache, etc. Both Quality universals and Modes are related to the entities that they characterize through a relation of characterization. Besides that, Relators are moments that depend existentially on two or more entities. Examples of relators are Enrollment, Contract, etc. Relators are related to entities that it relates through a relation of mediation. The relators also represent the relational dependency of roles and role mixins. Due to this, roles and role mixins must be related to some relator, through a relation of mediation.
UFO proposes four types of parthood relations: componentOf, memberOf, subCollectionOf and subQuantityOf. Each parthood relation can only be established between individuals of specific UFO meta-types, respecting some ontological constraints embedded in UFO. These relations can be characterized by five meronymic meta-properties that indicate: essential part, inseparable part, immutable part, immutable whole and shareable part.
As important as the characterization of the metaproperties and meta-types are, UFO also provides some postulates that a model should follow: • Postulate 1: Every individual in a conceptual model of the domain must be an instance of a sortal. • Postulate 2: An individual represented in a conceptual model of the domain must instantiate exactly one ultimate Substance Sortal (kind, quantity or collective). • Postulate 3: A rigid universal cannot specialize (restrict) an anti-rigid one. • Postulate 4: A dispersive universal cannot specialize a Sortal.
Furthermore, it is important to notice that every sortal that does not provide its own principle of identity (Role, Phase and SubKind) must be subsumed by exactly one concept that provides its own identity (one of the Substance Sortals).
The notion of ontologically well-founded ontology view
is defined considering certain principles of conservation
proposed by
In order to illustrate the proposed conservation principles, we present portions of two WFOVs generated from a base ontology for the domain of Petrography (a field of Geology). These WFOVs were generated for meeting the interests of two different communities of users within the domain of Petrography: Diagenesis (Figure 1) and Microstructural Analysis (Figure 2).
The principles of conservation proposed by
that does not provide its own principle of identity, then v should also include all the supertypes of c from which c inhered its principle of identity, as well as, all the subsumption relations that are held between these concepts. For example, if zeolite is included in v, mineral should also be included, since mineral provides the identity to zeolite.
Conservation of the existential dependence: If a concept c1 is included in the view v, and instances of c1 are existentially dependent on instances of c2, then it is necessary to include in v also the concept c2 and the relation held between c1 and c2. For example, if Porosity is included in v, the concept Rock must be included because the porosity is existentially dependent of Rock.
Conservation of relational dependence: If a concept c1 is included in the view v, and c1 is relationally dependent on a relation (materialized through a given relator) with the concepts in {c2, ..., cn}, then it is necessary to include in v also: the relator r, all the concepts in {c2, ..., cn} and all relations that are held between the concepts in {c2, ..., cn}, r and c1 that are necessary for the conservation of the relational dependence. For example, if the concept is Cement is included in v, the concepts pore and filling should also be included in v, because a mineral is considered cement when the mineral is filling pore. Thus, cement is relational dependent of filling and pore.
cept c1, it should also include all the concepts that are subsumed by c1. For example, if concept Silicate Mineral is included in v, all the concepts that it subsumes are included in the v.
cept c1, every attribute1 of c1 must also be included in v. For example, if the concept Diagenetic Constituent is included in v, the concept habit should also be included, because it is a quality of Diagenetic Constituent.
cludes a concept c1, every concept that is related to c1 in a formal relation is added. For example, if the concept Rock Unit is included in v, the concept Rock is also included because there is the formal relation constituted by between Rock Unit and Rock.
c1, all the concepts whose instances are parts of instances 1Adopting the UFO, attributes are considered Quality Universals of c1 should be included. For example, if the concept Rock Unit is included in v, the concepts deformation zone and sedimentary facies should also be included.
The basic approach for extracting WFOVs is formalized in
the Algorithm 1. It takes as input the following
parameters: the ontology base (Ob), a set of user required concepts
(targets), a set of relations (relations), and the resulting
extracted sub-ontology (So). At the beginning, relations
and So are empty. The algorithm analyses each concept in
targets. For each concept, the conservation principles are
applied for ensuring that the result will be an ontologically
well-founded ontology view. The conservation principles
are applied through the following functions: conservesTAX,
for the conservation of taxonomy; conservesQUA, for the
conservation of attributes; conservesIP, for conservation of
identity principle; conservesED, for the conservation of
existential dependence; conservesRD, for the conservation of
relational dependence, conservesFR, for the conservation of
formally related concepts; and conservesPR, for the
conservation of partonomy. In the main loop, these functions
accumulate concepts (in newC) and relations (in newR) that are
necessary for ensuring the defined principles for a given
concept c in tConcepts. More details regarding this approach
can be found in
Our new approach aims at reducing the number of concepts included in the views by the basic approach, and covering more precisely the requirements of a task at hand. In an overview, our approach relaxes some criteria adopted by the basic approach and provides more flexibility to the user.
The main difference in our approach regarding the basic approach concerns the application of the principle of conservation of taxonomy. While the basic approach includes in the resulting view the taxonomy of every concept that is already included in the view, our novel approach includes only the immediate taxonomy of the original target concepts. This modification was motivated by the fact that, in general, only the taxonomies of the target concepts are useful for the task for which the view was built. Besides that, the inclusion of the taxonomies of every concept in the view leads to a rapid increase in the size of the view. In this way, the inclusion of irrelevant taxonomies can be considered as a source of information overload.
Moreover, our approach also provides more flexibility to the user, by allowing the setting of three parameters. These parameters are the variables wP (with Partonomy), wRT (only Rigid Taxonomy) and wFR (with formal relation). In this way, the ontology engineer can specify if the desired WFOV should include the partonomies of every concept or not; if it should include only the rigid concepts in the taxonomies or if non-rigid concepts should be included as well; and if it should include all the concepts that are related (through formal relations) to concepts already included in the view.
Algorithm 2 Novel algorithm for extracting well-founded ontology views Require: Well-Founded Ontology procedure EXTRACTOR(Ob, tConcepts, tRelations, So, wP, wRT, wF R) So.C So.C [ tConcepts newC ; newR ; for all c 2 tConcepts do if wRT then conservesT AXR(Ob, c, newC, newR) else conservesT AX(Ob, c, newC, newR) end if end for newC newC [ tConcepts newR newR [ tRelations selection(Ob, newC, newR, So, wP, wF R) end procedure
Our approach is formalized in the algorithm 2. Firstly, it
applies the conservation of taxonomy only to the original
target concepts. At this point, the parameter wRT controls if
the taxonomy takes all the concepts or only the rigid ones
(Algorithm 4 presents how to recover only rigid concepts in
the taxonomy). Then, the algorithm calls the selection
algorithm (algorithm 3), which applies the other principles of
conservation, according to the parameters. Notice that the
algorithm 3 is a variation of the basic algorithm proposed by
We assume that the quality of the ontology view extraction approach can be measured by the degree to which the extracted ontology views fit to the required conceptualizations. In our work, we adopt an approach for evaluating this fitness in an indirect way.
We assume that the required conceptualization (of a
community, of some task) is properly represented, in natural
language, in the relevant literature. Thus, our evaluation is
based on measuring the correspondences between a given
ontology view (built for some specific task or some
community) and the set of terms extracted from the relevant
literature (related the correspondent task or community for which
the ontology view was built). We measure these
correspondences through well-known measures used in Information
Retrieval: Precision, Recall and F-measure
Considering this, we assume that in a useful approach for extracting ontology views, an ontology view generated for a community A should have a value of f-measure that is greater than the f-measure value of any well-founded ontology view generated for community B, when compared to a set of terms extracted from the literature of the community A. In other words, a view generated for the community X should fit better the conceptualization of the community X rather than the ontology view generated for another community.
Considering LT as the set of terms extracted from the literature, the precision (P) of the generated ontology view O is given by , the recall (R) is given by
P (OT, LT ) = |OT \ LT |
|OT | R(OT, LT ) = |OT \ LT | |LT | , and the f-measure (F) is given by
P (OT, LT ) ⇤ R(OT, LT )
F (OT, LT ) = 2 ⇤ P (OT, LT ) + R(OT, LT ) where |S| indicates the cardinality of the set S and OT is the set of terms that identify the set of ontology elements (concepts, relations and properties) of the ontology view O.
This section describes the application of the evaluation
method described in Section for comparing the performance
of our approach (A2) for extracting ontology views with the
performance of the basic approach proposed by
Our approach also requires the extraction of sets of terms
that are representative for the considered communities. For
this step, a domain expert selected six peer-reviewed papers
about Diagenesis, such as
Algorithm 4 Conserve Taxonomy only Rigid procedure CONSERVESTAXR(Ob, c, newC, newR) for all v 2 Ob.C|9 r = Rel(subsumption, c, v) do if metaT ype(v) 2 { SubKind, Collective, Kind, Quantity, Category} then newR newR [ r newC newC [ v conservesT AXR(Ob, v, newC, newR) end if end for end procedure
This extraction was done manually, for ensuring the quality of the extraction. We excluded the terms that were not exclusive to the communities of Diagenesis and Microstructural analysis. We also excluded the terms that were common for both communities. The result was two lists of geological terms; one (DT ) for the community of Diagenesis and other (M T ) for the community of Microstructural analysis.
The next step consists of generating the two well-founded ontology views (one for each community), from a set of key terms that are representative of the community. These key terms were provided by domain experts. The WFOVs that were considered in this evaluation were extracted using Detrital Constituent, Diagenetic Constituent and Pore for the WFOV of Diagenesis and Deformational Band, Fault, Breccia and Microfracture for the community of Microstructural analysis.
The result of this step is an ontology view for Diagenesis (DO) and ontology view for Microstructural analysis (M O). For our proposed approach, we extracted one WFOV for each community, considering each combination of parameters.
The last step is to calculate and compare the f-measures, considering the WFOVs and the sets of selected terms. As depicted in Figure 4, we expect that the f-measure (F DD = F M easure(DO, DT )) between ontology view for Diagenesis (DO) and the terms for the community of Diageneis (DT) is greater than the f-measure (F DM = F M easure(M O, DT )) between ontology view for Microstructural analysis MO and the terms for the community of Diageneis DT. And, in the same way, it is also expected that the f-measure (F M M = F M easure(M O, M T )) between MO and MT is greater than the f-measure (F M D = F M easure(DO, M T )) between DO and MT. In the following subsections, we describe the evaluation of these two cases. In Case 1, we evaluate the ontology view for Diagenesis, comparing it with the ontology view for Microstructural analysis, considering the terms for Diagenesis. In Case 2, we evaluate the ontology view for Microstructural analysis, comparing it with the ontology view for Diagenesis, based on the microstructural terms.
In Table 2, we present the results of the evaluation process of approaches A1 (basic approach) and A2 (novel approach), in the two considered cases. Notice that the table presents the ratios between the considered measures of the two ontology views, evaluated according to a set of terms. Thus, in the row Case 1, the table presents, for both approaches A1 and A2, the ratios between the measures (precision, recall and f-measure) of the ontology view for Diagenesis and the measures of the ontology view of Microstructural analysis, considering the terms of Diagenesis. Notice that it is expected that the resulting ratios are greater than 1. In a similar way, the row Case 2, the table presents, for both approaches A1 and A2, the ratios between the measures of the ontology view for Microstructural analysis and the measures of the ontology view of Diagenesis, considering the terms of Microstructural analysis.
In the approach A1 (basic approach), the ontology view generated for the Diagenesis community obtained a ratio of fmeasure smaller than 1. This means that the basic approach does not satisfy the expectation. This happens because the approach A1 includes many ontology elements that are not necessary for the community. However, the ratio of recall is greater than 1, as expected. This means that the ontology view generated for the Diagenesis community contains more relevant terms than the ontology views obtained for the Microstructural analysis community.
For the approach A2, proposed in this paper, in all cases, the ratio of P , R and F was greater than 1. This occurs because it applies the principle of conservation of taxonomy just once to the target concepts, eliminating taxonomies that are useless for the community in focus. In general, approach A2 results satisfy our expectations about the generated ontology view for all set of representative terms given in this case study.
In the column Case 1 of Table 2, it is possible to see that the new approach (A2) achieve better results than the basic approach (A1), for the ontology views of the community of Diagenesis. The best results of approach A2 for Case 1 were achieved with wRT and with all the parameters as false.
In Case 2, the ratio of f-measure achieved by the approach A1 is greater than 1. Also, the precision obtained by this approach achieved the highest value, in comparison with the approach A2, considering all parameter combinations.
The approach A2 also achieved high quality results in Case 2. Each combination of parameters of approach A2 satisfies our expectations about the generated ontology views, considering the set of representative terms given in this case study. However, the precision in all combination of parameters for Microstructural analysis ontology view has a ratio smaller than 1. This means that the ontology view generated for the Microstructural analysis community contains few terms to cover the terminology used in the literature of Microstructural analysis, than the ontology view for Diagenesis.
The row Case2 of Table 2 presents the ratio between the measures of the ontology view of Microstructural analysis and the measures of the ontology view of Diagenesis. In this row, it can be seen that approach A2 achieved its best result using the parameter wP.
The approach A2 achieved results of low quality in some settings because, for this domain, the key terms could be related by formal relations with other concepts that do not belong to the Microstructural analysis ontology view.
In this work, we propose a novel approach for extracting
well-founded ontology views, by improving the approach
proposed in
It is important to notice that, although this work adopts the
meta-properties defined by UFO, it can be viewed as a
specific implementation of a more general idea. The notion of
WFOV is defined according to a set of principles of
conservation that should be followed by the extraction algorithm.
The set of principles of conservation can be changed, by
including, excluding or modifying the principles proposed
in this work. In this way, the general approach proposed in
In this work, we also propose a method for evaluating
approaches for extracting ontology views. This method uses
well-known measures used in information retrieval
(precision, recall and f-measure), for evaluating the fitness of the
resulting ontology views to the target conceptualization (of a
community or task). According to this method, our novel
approach outperforms the basic approach
In future works, we also plan to improve the proposed approach by identifying and eliminating other sources of information overload in the resulting ontology views. Besides that, we also intend to investigate if the ontological metaproperties considered in this work can also be applied for guiding the extraction of ontology modules.