Analysis of the naming of entities across ontological structures can help reveal both naming issues and underlying conceptualization issues. Cross-entity naming analysis thus extends the standard logical satis ability checking by an extra, less rigorous and reliable but often farther reaching layer. We show how such naming patterns can be applied within the transformation pattern paradigm used by the PatOMat transformation framework. We describe how the PatOMat tool has been integrated into the (logic-oriented) Ontology Repair and Enrichment tool (ORE), and present the results of application of a prominent naming pattern, `non-matching child', on a collection of linked data vocabularies.
During the decades of knowledge engineering research, there has been recurrent
dispute on how the natural language structure in uences the structure of formal
knowledge bases and vice versa. A large part of the community seems to
recognise that the content expressed in formal representation languages, such as the
semantic web ones, should be accessible not only to logical reasoning machines
but also to humans and NLP procedures, and thus resemble the natural language
as much as possible.3 We build upon the assumption that naming in ontologies
matters, can be sensibly designed, and to some degree even automatically
identi ed in existing ontologies, with the help of naming patterns. Current ontology
debugging methods, mostly dealing with the logical structure of the ontology
only, can thus be extended by debugging of naming issues. Detecting improper
or awkward naming should be ideally followed by repairing suggestions. While
in the biomedical eld there have already been e orts in naming analysis, e.g.,
in [
presented approach thus contributes to lling in a missing piece: domain-neutral cross-entity analysis.
The paper follows up on earlier research described in [
The paper is structured as follows. Section 2 brie y surveys the PatOMat transformation framework. Section 3 explains the NMC (non-matching child) pattern and describes its implementation via PatOMat transformation patterns. Section 4 describes the integration of the whole functionality into ORE. Section 5 reports on an experiment in NMC pattern detection over 16 ontologies (namely, popular linked data vocabularies). Section 6 then wraps up the paper. 2
PatOMat framework principles The PatOMat Framework4 has been originally designed with the goal of transforming ontologies between `structural' modelling styles, e.g., via de/reifying properties, metamodelling classes by individuals, switching between object and data properties, and the like. However, the entity naming aspect has been considered from the beginning.
The central notion in PatOMat is that of transformation pattern (TP). A TP
contains two ontology patterns (source OP and target OP) and the description
of the transformation betweem them, called pattern transformation (PT). For
instance, we can specify a very simple TP such that a subsumption relation
between two classes (as source, OP1) should be transformed to a SKOS5 taxonomic
relationship between two individuals (as target, OP2). A schematic description
follows.6
OP1: Class: ?OP1_A subClassOf ?OP1_B
OP2: Class: ?OP2_A skos:broader ?OP2_B
PT: ?OP1_A = ?OP2_A ?OP1_B = ?OP2_B.
4 [
The representation of OPs is based on the OWL 2 DL pro le. However, while an OWL ontology refers to particular entities, e.g. to class Person, in the patterns we generally use placeholders, e.g. ?OP 1 A. Entities are speci ed (i.e. placeholders are instantiated) at the time of instantiation of a pattern. An OP consists of entity declarations (referring to placeholders or concrete entities), axioms and naming detection patterns (NDPs); the last capture the naming aspect of the OP important for its detection, see below. A PT consists of a set of transformation links and a set of naming transformation patterns (NTPs). Transformation links are either logical equivalence relationships or extralogical relationships holding between pairs of entities of di erent type (such as class vs. individual, as in our example above). NTPs serve for generating new names for original or newly created entities.
PatOMat currently supports naming operations at the level of short URIs; its extension to rdfs:label values would however be straightforward. Various token separators, such as underscore, hyphen or camel-case, are supported. PatOMat implementation The framework prototype implementation is available either as a Java library or as three core services.7 The whole transformation is divided into three steps, which correspond to the three services: { The OntologyPatternDetection service takes the transformation pattern and a particular original ontology on input, and returns the binding of entity placeholders on output, in XML. The structural/logical aspect is captured in the structure of an automatically generated SPARQL query;8 the naming aspect is dealt with based on its description within the source pattern. { The InstructionGenerator service takes the particular binding of placeholders and the transformation pattern on input, and returns particular transformation instructions on output, also in XML. Transformation instructions are generated according to the transformation pattern and the pattern instance. { The OntologyTransformation service takes the particular transformation instructions and the particular original ontology on input, and returns the transformed ontology on output. The service is based on our speci c implementation over OWL-API,9 and enables operations on axioms, entities and adding OWL annotations.
The process of transformation is decomposed into parts in order to enable
an user intervention within the whole work ow. User intervention can be carried
out using a generic graphical tool, GUIPOT [
8 http://www.w3.org/TR/rdf-sparql-query/ 9 http://owlapi.sourceforge.net/
It is quite common in ontologies that a subclass has the same head noun as
its parent class.10 By an earlier study [
Owned' (subclass of `Company').
While the former requires complex refactoring of the ontology fragment, the latter can be healed by propagation of the parent name down to the child name. NMC pattern in PatOMat Let us now show how to capture the NMC pattern within a PatOMat TP. The source OP is here just a subClassOf relationship between two classes. There are however two variants of this source OP: one using subClassOf and one using directSubClassOf, the latter operating on class pairs identi ed by a reasoner.11
An NDP within the source OP can consist of several naming operations such as detection of a head noun; their results can then be compared using di erent methods. We designed two variants of such an NDP: 1. comparing whether ?OP 1 A has the same head noun as ?OP 1 P (e-variant, for `equality') or 2. comparing whether head noun of ?OP 1 P is a hypernym of head noun of ?OP 1 A (t-variant, for `thesaurus').
The target OP has only one variant, which is structurewise identical to the source OP, i.e., Class: ?OP2 A SubClassOf: ?OP2 P (and there is no NDP part). Finally, the PT contains transformation links specifying equality of ?OP 1 A and ?OP 2 A and analogously for ?OP 1 P and ?OP 2 P . More interestingly, it also includes an NTP, which represents the naming repair step. It speci es that ?OP A should be extended by the head noun of ?OP 1 P .
In combination, we can have four variants of the NMC pattern, of which we
consider three: (1) Se, St and Dt.12 The Se variant simply matches the head
10 The head noun is typically the last token, but not always, in particular due to
possible prepositional constructions, as, e.g., in `HeadOfDepartment'.
11 We used Pellet, http://pellet.owldl.com/.
12 All variants available at http://nb.vse.cz/~svabo/patomat/tp/np/
nouns, the Dt variant needs to employ a reasoner, and both t variants employ
a thesaurus in order to verify the hypernym relationships. In our case we use
WordNet [
In the experiment described in Section 5, we used the St variant of the pattern, but, referring to the conjecture formulated at the beginning of this section, distinguished between single- and multi-token child. 4
The ORE13 (Ontology Repair and Enrichment) tool was designed so as to allow
knowledge engineers to improve knowledge bases in the form of SPARQL
endpoints and OWL ontologies. ORE helps xing several kinds of problems such as
logical errors, i.e., unsatis able classes and inconsistencies, by applying
state-ofthe-art methods [
The PatOMat framework is integrated into ORE by means of a separate task and visualized in a single view as shown in Figure 1. Here the user can select a naming pattern, as for example non-matching child1 (corresponding to the St variant of the non-matching child pattern), in the leftmost list ( 1 ). PatOMat then detects instances of the selected pattern in the currently loaded ontology, e.g., [?OP1 P=Contribution;?OP1 A=Poster](see 2 ). For the selected pattern instances the user will be provided a list of renaming instructions (see 3 ), for example to rename the class Poster to PosterContribution, which can then be used to transform the ontology and solve the detected naming issues. 5
We carried out a small experiment on 16 randomly selected vocabularies from the Linked Open Vocabularies (LOV) catalog.14 They belong to four of the eight major clusters suggested by the LOV curators: City (related to personal, social and governmental data), Library, Media and Market. Four of the vocabularies, ontopic, swc, wi and gc, imported other vocabularies; we considered them together with these imports. For completeness we also include two vocabularies that did not contain any subclass axioms. 13 http://aksw.org/Projects/ORE.html 14 http://lov.okfn.org
The summary of the results is in Table 1. The rst two columns display the nickname of the vocabulary within LOV,15 its catalogued name and cluster. The next three columns show the total number of asserted subclass axioms, the number of axioms matching the NMC pattern (the ordering in the table is in the descending order of this eld), and the respective ratio. The following three columns are analogous, but only refer to subclass axioms where the subclass has a multi-token name. The last two columns show the time spent (in seconds) and the average time corresponding to one subclass axiom.
The proportion of axioms satisfying the NMC pattern (`Ratio all' column) ranges from 0% to 87%. The results however seem to con rm our conjecture that subclasses with multi-token names are more likely to follow (at least via thesaurus correspondence) the head noun of their parent class name. For 11 of the 16 vocabularies, the frequency of occurrence of the NMC pattern was reduced by focusing on multi-token subclasses, and only for 2 (swc and gr ) it increased.
Manual analysis of the `alerts' revealed several interesting cases: { Although possibly with some `philosophical' excuse, the class-subclass pair `Topic'{`TopicSignature' in ontopic is suspect for tacit partonomy, as the latter is commented as `the subcollection of terms populating a Topic' (a set of terms is likely not same as the topic it populates). { Head noun of the superclass is sometimes a meta-level term. For example, in swc the class `ProgrammeCommitteeMember' is a subclass of `Role'; obviously, if this subclass is populated by `Person' entities, they would become instances of `Role', which would be undesirable. In gnd, which is a taxonomy of keyword types rather than a true data vocabulary, class `PlaceOrGeographicName' has subclasses such as `MemberState'; however, the latter could be, in the linked data setting, populated by true `state' entities (rather than just by their names viewed as keywords). 15 The URI of the catalog item page is http://lov.okfn.org/dataset/lov/details/ vocabulary_<nickname>.html. Nick Name/topic name ontopic Ontopic Library swc SemWeb Conference City gnd GND Library mvco Media Value Chain Media pattern Pattern Library gr GR - GoodRelations Market wi Weighted Interests City bibo Bibliographic Library frbr Core FRBR Concepts Library gc oeGOV Governm.Core City foaf FOAF City sioc SIOC City pna Press.net Asset Media chord OMRAS2 Chord Media comm Incident communication City part Participation Schema City Total { The previous is a special case of name shorthanding. A more typical case is such that no head noun can be detected at all, e.g., with pair such as `PoliticalSystem'{`Tribal' (in gc) or `Publication'{`Unpublished' (in swc). { Some vocabularies rede ne common terms in their speci c manner, which generates false alerts. For example, in bibo, `CourtReporter' is subclass of `Periodical' (while a `reporter' would not primarily be viewed as periodical) and `LegalCaseDecision' is a subclass of `LegalCaseDocument' (while a `decision', in the general sense, is not a document). Similarly, ontopic has `TimeInterval' as subclass of `Region' (here the unusual choice of term `region' follows from the upper-level nature of the ontology). { Sometimes the head noun detection fails due to non-intuitive agglutination of tokens. A grammatically sound one is `SubjectHeadingSensoStricto' (in gnd ), unmatched to its parent `SubjectHeading'. Less sound seem to be, e.g., `PaymentMethodCreditCard or `QuantitativeValueFloat' (in gr ). { An unusual case setting may make the tokenizer fail and thus lead to a false alert, e.g., for a class named `Vevent' (a kind of `event' in swc).
The relatively high computation times are partly owing to SPARQL query evaluation and partly to WordNet traversal. However, given the o ine nature of the task, they are not prohibitive. Many ontologies, including linked data vocabularies, have recently been created, often with little concern for naming coherence, and sometimes even with conceptualization aws (also re ected by naming incoherence). Visual analysis of the naming aspect of their taxonomic structures is typically feasible, as most ontologies/vocabularies are not extremely large. However, allowing the user to focus on `suspicious' cases (and ignoring those apparently sound) can be helpful.
We present a solution for such machine-supported analysis, which combines the functionality of an ontology debugging (and enrichment) tool, ORE, with that of a pattern-based ontology transformation framework, PatOMat, and with online access to WordNet. Empirical results of naming analysis (regarding the NMR pattern) on 16 linked data vocabularies have been presented.
In future we plan to come up with more sophisticated naming/transformation patterns (e.g., indicating a taxonomy/partonomy mismatchh), and involve a larger number of vocabularies in the experiment.
The research is supported by the EU ICT FP7 under No.257943, LOD2 project.