A correspondence antipattern is a set of generic correspondences between two ontologies that represents an incorrect alignment. It is useful to help identify incorrect correspondences between two ontologies, thus improving the Ontology Matching process. The specification of a correspondence antipattern requires the identification and correct understanding of a relevant alignment problem, and its representation in a proper modeling language. In this work we investigate the last three editions of OAEI challenge datasets so as to identify correspondence antipatterns from frequent and recurring errors; some of the the resulting antipatterns are presented and discussed.
As the research and practice on Ontology become more popular and evolve, several
ontology artifacts arise for the same universe of discourse. However, they differ
among each other in several perspectives, such as distinct representation languages
(syntactic heterogeneity), variations in names referring to the same entity
(terminological heterogeneity), different conceptualizations for the same domain (conceptual
heterogeneity) and entities being perceived differently (semiotic heterogeneity) [
The Ontology Matching task consists in identifying the correct correspondences
among entities of multiple ontologies, which it is a necessary condition for
establishing the interoperability among them [
On the other hand, in the context of software development, antipatterns are considered a valuable tool for the identification of bad or incorrect practices in the software development process. Antipatterns prevent or hamper a good execution of the software development or maintenance process. In the context of ontology matching, bad solutions consist of incorrect (including missing) or problematic correspondences. A correspondence antipattern is a matching model for identifying problematic correspondences that may occur repeatedly in ontology matching processes. A correspondence antipattern may be useful in several scenarios in which Ontology Matching is applied (such as in ontology merging, ontology comparison, query translation), since it helps refining an alignment produced by an ontology matching tool.
Looking for correspondence antipatterns, we “dig” the alignments available by
OAEI and apply a methodology previously proposed in [
This work is divided as follows: Section 2 shows an overview about ontology correspondence antipatterns, Section 3 presents how we “dig” some correspondence antipatterns from the data published by OAEI, Section 4 presents related works and, finally, Section 5 points final considerations of this work. 2
Ontology matching identifies correspondences between the entities of multiple
ontologies, and it is a necessary condition to establish interoperability between them [
Due to possible precision errors that every ontology alignment tool is subject to, it may be the case that a correspondence included in an ontology alignment is not correct. Take, for example, a real problem illustrated in Figure 1, showing an alignment problem that occurs in the last three OAEI1 editions, between ConfOf and Edas ontologies. The Ontology Alignment Evaluation Initiative (OAEI) is a coordinated international initiative whose goal is to evaluate the strengths and weaknesses of the ontology alignment tools. OAEI organizes annual campaigns addressing several domains, and publishes the results of the evaluated tools. The correspondence between the ConfOf.Conference and Edas.Conference classes is a problematic one. Let’s analyze this case: suppose that x is an instance of Edas.Conference. Since an equivalent relationship between the entities Edas.Conference and ConfOf.Conference has been established, we may deduce that there is a possible world w in which x is an instance of ConfOf.Conference as well. Since ConfOf.Conference is a specialization of Con1 http://oaei.ontologymatching.org/ fOf.Event, x is necessarily an instance of ConfOf.Event in w. We also notice that there is an equivalence correspondence established between ConfOf.Event and Edas.Conferece_Event. Thus, x is also an instance of Edas.Conference_Event in w. However, considering that Edas.Conference_Event and Edas.Conference are disjoint classes, there should be no possible world in which x instantiates both Edas.Conference and Edas.Conference_Event simultaneously, which leads to a contradiction, thus evidencing an alignment problem.
Patterns assist in building a collective experience based on the skills of domain specialists. On the other hand, an antipattern is a description of a given solution to a common problem that generates, definitely, negative consequences.
Given two ontologies O and O’ to be aligned, a correspondence antipattern is a set of generic, domain-independent correspondences and/or non-correspondences between the entities of O and O´ that lead to a contradiction. The purpose of a correspondence antipattern is, then, to help domain specialists in identifying a mismatch (a wrong correspondence) within an alignment.
We may generalize the example scenario illustrated in Figure 1 as follows: Consider a class e1 in an ontology o1 that is a subclass of a class e2, which in turn is subclass of a class e3 in o1. If class e3 in the ontology o1 is equivalently correspondent to class e2 in ontology o2, and classes e1 (from ontology o2) and e2 (from ontology o2) are disjoint, then class e1 from ontology o1 cannot equivalently match class e1 from ontology o2. As shown in [28], this correspondence antipattern can be represented as follows: {(?o1:?e1 ≡ ?o2:?e1) ⊓ (?o1:?e1 ⊑ ?o1:?e2) ⊓ (?o1:?e2 ⊑ ?o1:?e3) ⊓ (?o1:?e3 ≡ ?o2:e2) ⊓ (?o2:?e1 ⊓ ?o2:?e2 ⊑ ⊥)} (1)
As shown in [
The methodology was applied on the results provided by the OAEI in the last three editions (2011.5, 2012 and 2013). The identification of correspondence antipatterns considered recurring incorrect correspondences generated by the evaluated tools. We identified incorrect correspondences by comparing tool results with the reference alignment published by OAEI. Each step of this process will now be briefly explained and illustrated in the OAEI scenario.
First step: Show problematic solution. The first step towards the construction of correspondence antipatterns is the correct understanding of the problem being treated. To start the search for correspondence antipatterns, the first step was the identification of incorrect correspondences, or false positives, in the set of selected alignments. False positives are the correspondences found by the evaluated tools that are not in the reference alignments. Within the universe of identified incorrect correspondences, we selected those that most frequently occurred (i.e., that were identified by many of the evaluated tools). We selected 40 incorrect correspondences, which were the ones that occurred over 50% of the analyzed alignments, as shown in Table 1. The columns Ontology 1 and Ontology 2 denotes the ontologies being aligned and the columns Entity 1 and Entity 2 denotes the entities involved in the incorrect correspondences found. The Total Problems column shows the quantity of alignments analyzed in which the incorrect correspondence was found. The Total Alignments column shows the quantity of alignments analyzed. Percent is calculated as Total Problems / Total Alignment.
Ontology 1
Ontology 2
Entity 1 Conference
Edas Edas
Call for paper has an email
CallForPapers hasEmail 27 56 48%
Second Step: Evidentiate problematic solution. For a solution to be considered
problematic, this should in fact occur [
Third Step: Demonstrate Implications. For each incorrect correspondence, the
error and its implications are analyzed according to the classification of types of
inferences examined in [
Error Number 16: In the set of alignments analyzed, the correspondence
conference.paper, ekaw.paper, ≡, _ occurs 47 times. By analyzing the correspondence
together with the aligned ontologies we identified the following problem: let e1 be a
class in an ontology o1 which is subclass of a class e2, which in turn is a disjoint class
of a class e3, also in ontology o1. If class e1 in ontology o1 equivalently corresponds
to class e1’ in ontology o2, class e2 in ontology o1 corresponds to class e2 in
ontology o1 and class e2’ in o2 is a subclass of e1 in ontology o1, then there is a
contradiction (more specifically, a disjointness-subsumption contradiction alignment problem
[
Error Number 20: In the set of alignment analyzed, the correspondence
confof.paper, ekaw.paper, ≡, _ occurs 46 times. By analyzing the correspondence
together with the aligned ontologies we established the follow problem: let e1 be a class
in ontology o1 that is disjoint with class e2 in the same ontology o1, and a class e1’ in
ontology o2 that specializes class e2’ in the same ontology o2. If class e1 in o1
equivalently corresponds to class e1’ in o2 and class e2 in o1 equivalently corresponds to
class e2’ in o2, then there is a contradiction (a disjointness-subsumption contradiction
alignment problem [
Error Number 25: In the set of alignment analyzed, the correspondence cmt.writtenBy, confof.writtenBy, ≡, _ occurs 41 times. By analyzing the correspondence together with the aligned ontologies we established the following problem: let p1 be a property in ontology o1 that has class e1 as its domain and class e2 as its range, both in ontology o1, and a property p1’ in an ontology o2 that has class e1’ as its domain class e2’ as its range, both in ontology o2. If p1 in o1 equally corresponds to the property p1’ in o2, but class e1 in o1 does not correspond to class e1’ in o2 or class e2 in o1 does not correspond to class e2’ in o2, then there is a domain and range incompleteness alignment problem. Figure 5 shows the case identified on the correspondence number 25, where the above problem occurs.
Error Number 27: In the set of alignment analyzed, the correspondence confof.hasStreet, edas.hasStreet, ≡, _ occurs 40 times. By analyzing the correspondence together with the aligned ontologies we established the following problem: let p1 be a property in an ontology o1 that has classes e1 and e2 as its domain, both in ontology o1, and a property p1’ in an ontology o2 that has as its domain a class e1’ in ontology o2. If p1 in o1 equally corresponds to the property p1’ in o2 and class e1’ in o2 does not correspond to any domain class of p1 in o1, then there is a domain and range incompleteness alignment problem. Figure 6 shows the case identified on the correspondence number 27, where the above problem occurs.
Fourth Step: Identification of the Problematic Solution. The formal representation of how to identify an alignment problem is what gives life to correspondence antipattern. For each problem analyzed was created one correspondence antipattern, as summarized in Table 2.
For the construction and computational representation of a correspondence
antipattern, we adopt EDOAL (Expressive Declarative Ontology Alignment Language), an
open and agnostic language [
Fifth Step: Refactored Solution. Refactoring in this case means repairing the alignment to eliminate logical inconsistencies. This is not a trivial activity, since there may exist many solutions for a specific scenario. Morevover, the best solution may also depend on the task, or even point some problem in the semantics of the correspondence. Therefore, this task is currently carried out by the specialist, with no automatic support. Further evolution of this approach will investigate automatic approaches for alignment refactoring. 4
In ontology research, Ontology Design Patterns (ODPs) are an emerging approach
that favors the reuse of encoded experiences and good practices. ODPs are modeling
solutions to solve recurrent ontology development problems [
Correspondence patterns, proposed by [
Ontology matching is a very active research field in the scientific community, where various techniques and approaches have been proposed. However, existing tools are still likely to identify incorrect correspondences between the entities of the ontologies that are being aligned. The identification of recurrent errors may serve as input for the construction of correspondence antipatterns. A correspondence antipattern is a set of generic correspondences between two ontologies that represents an incorrect alignment. They assist in identifying incorrect correspondences in a given alignment, and should be computationally representated in an open and agnostic language.
OAEI is an important initiative that provides the community with the results os evaluations of several ontology matching techniques and tools. This published data constitutes a rich environment for analyzing recurrent errors in practical alignments.
In this work, the results provided by OAEI in three evaluation editions (2011.5,
2012 and 2013) were analyzed. The identified recurring alignment problems were
considered and some correspondence antipatterns were specified and codified and
EDOAL, following the methodology proposed in [
Future works include the exhaustive analysis and identification of correspondence antipatterns from other OAEI datasets, and the construction of a framework to make use of these antipatterns in refining ontology alignments.
Fernanda Baião is partially funded by CNPq brazilian research council under the project 309069/2013-0.