B-Annot: Supplying Background Model Annotations for Ontology Coherence Testing

Abstract

The demo paper presents B-Annot, a Protege plugin for annotation of ontologies and linked data vocabularies by background model distinctions. In addition, it brie y demonstrates the subsequent use of the annotations created by B-Annot, for verifying the ontological coherence of the ontologies/vocabularies at the level of meta-models. Finally, possible further extensions of the tool and its role in the (backgroundmodel-driven) ontological engineering work ow are brie y discussed.

-

remaining two instance-level resources in the diagram (2 and 4) are the `o ering' itself and the value `90' (minutes) understood as `typical' and thus modeled as a resource rather than literal.5 In the lower part of the diagram we approximate the ontological background of this fragment (omitting the entities that would be types in both diagrams, for easier readability). Among other things we see that notion of `album', originally being the value of the object property mo:release type, now becomes an additional type of the product o ered, and that the `O ering' object becomes absorbed by the `o ers' relationship (now with arity >2).

Obviously, modelling the ontological background for each individual data fragment is infeasible. The mapping between the `foreground view' of the domain (as contained in the vocabulary) and the corresponding `background view' thus has to be established at the level of entity types, which means, indirectly (note that especially less expressive vocabularies are just collections of unlinked entities whose connection is only established at the level of instance data). On the one side of the mapping is an ontological foreground model (OFM), i.e., the structure of an RDFS/OWL ontology; on the other side is an analogous ontological background model (OBM). OBM models should be represented in a suitable OBM language of modelling primitives (OBML). Two such languages are { OntoClean [ 3 ], which labels OFM classes with the ontological notions of essentiality, rigidity (e.g., in the rst example mentioned, `permanent' classes Human and Robot would be rigid while the `temporary' class Student would be anti-rigid), identity and unity. { The recently designed PURO OBML [ 9 ],6 aiming to capture the background distinctions of OFM entities as in the bottom part of Fig. 1: that between objects (`particulars') and their types (`universals') and that between relationships (or `valuations' by a quantitative value) and self-standing objects. OntoClean has proven useful for taxonomy-centric ontologies that dominate, e.g., in bioinformatics. On the other hand, PURO has been speci cally designed for `relation-centric' ontologies/vocabularies [ 8 ], which are prominent in LD. Another important phenomenon in LD is that an existing entity might be systematically used with a di erent background distinction than foreseen in the vocabulary speci cation; for example, a property that is assumed to have categories of objects in its range might refer to individual objects in some dataset. Therefore, `generic' annotation of vocabularies might not be su cient; we should also be able to annotate vocabularies `as they are used' in a speci c dataset.

By their capacity of underlying the entities from various operational (typically, domain-restricted) knowledge models with background ontological distinctions, OBMLs are analogous to foundational ontologies. The di erence is in the way the `surface' and `deep' model are interconnected. A foundational ontology provides root concepts upon which the `surface model' concepts are grafted; both models thus share the same space. In contrast, OBMs reside in their own `layer'; 5 Such kind of modeling is not common in MO, but, rather, in GR-compliant ontologies, cf. http://www.ebusiness-unibw.org/ontologies/opdm/#ontologies. 6 A more extensive description is in [ 7 ]. when connecting an OFM with an OBM, we thus need to `inject' a `proxy' of one model to the other model, in order not to let the one interfere with the formal semantics of the other. Two alternatives for creating such a `proxy', assuming both layers are to be expressed in OWL-DL, are as follows: { OBM entities could become values of speci c OWL annotation properties, and be saved as unobtrusive part of (a copy of) the OFM. { OFM entities (classes, properties and individuals) could be uniformly metamodelled as syntactical instances to be inserted as an A-Box into a metamodelling ontology, where their mapping to OBM can be captured. The rst alternative is favourable for visibility of the OBM distinctions to a human when working with the OFM. The second alternative, in turn, allows to carry out conceptual coherence checking according to constraints de ned in the meta-modelling ontology, via a generic OWL DL reasoning mechanism. This approach has been previously tested for the OntoClean OBML in [ 2, 10 ], and later for the PURO OBML by us [ 9 ].

In this system/demonstration paper we present B-Annot : a Protege plugin7 that allows to create and save meta-models of a selected vocabulary with respect to either OntoClean or PURO, and (especially for the latter) in two modalities, `generic' and `dataset-speci c'. (Storage of OBM distinctions in annotation property values, as well as other enhancements, is forthcoming.) We also brie y demonstrate how the annotations can be used for conceptual coherence checking; in contrast to PURO-only coherence checking described in [ 9 ], we nowadays rely on a modular set of ontologies that also includes an OntoClean module. 2

B-Annot Functionality

In summary, the tool allows the user, for the vocabulary to be annotated already loaded into the Protege editor, { to select the meta-ontology (either OntoClean or PURO) and decide whether generic or dataset-level annotation is going to take place; { for dataset-speci c annotation, to inspect the statistics of presence of entities from the given vocabulary in di erent datasets, fetched online from LODStats [ 1 ], to select an appropriate dataset, and to view the list of entities from the vocabulary that occur in this dataset; { for dataset-speci c annotation, to browse a pre-computed summary of the dataset (inspired by [ 4 ]), with entities from the vocabulary highlighted; { select an entity (in one of the Protege tabs) and annotate it with a background model distinction; { save the whole annotation set to an RDF le, and load it back. 7 Available from http://patomat.vse.cz/cz.vse.bannotation.plugin.view.jar. mo:Musical Manifestation 7351...3537 gr:Quantitative

ValueFloat rdf:type mo:release rdf:type

type mo:ean 1

ex:recordedLength rdf:type 4 gVra:hluaes gr:includes 90 mo:album gr:SomeItems gr:Offering rdf:type

2 gr:offers 3 gr:eligible

regions rdf:type “US-CA” gr:Business

Entity Album type 1 recorded length of instances 90 min

Territory of USA and Canada

offers where? offers instances of what?

2 offers who offers?

3

We will now describe the scenario of dataset-speci c annotation, since generic annotation is essentially a subset of it. Furthermore, we will use the PURO metaontology as more relevant in the dataset-speci c mode. (OntoClean would be applied in the same way.) Fig. 2 shows the B-Annot interface after the choice of PURO and dataset-speci c annotation mode (FOAF has been previously loaded into Protege as ontology to be annotated). The user can see that of the 14 datasets for which the statistics has been fetched, 10 use some number of FOAF entities, ranging from 1 to 23; these are relevant to the annotation session. After clicking at the `summary' button for the Geospecies dataset, an ordered listing of frequent `class-property-class' is displayed, a part of which is in Fig. 3.8 FOAF entities, here the properties depiction, isPrimaryTopicOf, primaryTopic and topic, are displayed in red. Finally, the actual annotation takes place. In Fig. 4 we see that the user, based on the observation that foaf:topic is usually valued by biological taxa9 in this dataset, assigns this property the PURO label `PrT' (`property whose range is a type') from the pull-down menu (with items picked from the meta-ontology depending on the entity type to be annotated: class, property or individual). Entity annotations are subsequently listed in the bottom part of the window, and can, eventually, be saved (and reloaded) in bulk, as a set of hasLabel 10 triples. 8 We also experiment with `class-property-class-property-class' paths, but they are not implemented in the current version of the system. 9 For the sake of this example, we omit the philosophical discussion whether and for what purpose a taxon should indeed be understood as a universal. 10 Every meta-modelling ontology has its own hasLabel property; here it is the one from the PURO ontology.

Coherence Checking Examples

For each OBML considered, the distinctions underlying a particular OFM can be compared to a prede ned set of coherence rules. For OntoClean there are four standard coherence rules [ 3 ]: Given two properties, p and q, when q subsumes p then: a) if q is anti-rigid then p must be anti-rigid, b) if q carries an identity criterion then p must carry the same criterion, c) if q carries a unity criterion then p must carry the same criterion, and d) if q has anti-unity then p must also have anti-unity. The PURO OBML, in turn, speci es three constraints: for a) entity coherence, b) type coherence, and c) relation coherence (for details see [ 9 ]). We demonstrate the coherence checking on two example annotations.

The rst is a fragment of the GR ontology annotated with PURO OBML,11 containing class ProductOrService with subclasses Individual and ProductOrSer11 http://patomat.vse.cz/gr_mm.owl viceModel. Using DL consistency checking over the PURO meta-ontology12 and this fragment leads to inferred membership of class ProductOrService to a special `diagnostic' class of the PURO ontology: Incoherent-TPU. This class which contains meta-models of classes that `do not have homogeneous instances' in terms of PURO, speci cally, whose instances can be both particulars and universals.

The second example is annotation of a fragment of the ontology used to demonstrate OntoClean inconsistencies in [ 3 ]. This fragment13 includes metaentities representing six classes of the original ontology annotated with OntoClean labels. The OntoClean meta-ontology used for coherence checking14 allows for validation of all four coherence rules. The ontology contains four classes (Incoherence-Antiunity, Incoherent-Identity, Incoherent-Rigidity, Incoherent-Unity) that are { as result of inference { lled with individuals that represent classes in meta-model that are incoherent with regard to respective OntoClean rules. For example, the class AmountOfWater was annotated with OntoClean labels +O U +R. Its subclass LivingBeing was annotated with OntoClean labels +O +U +R. The defect of the model is that a class with anti-unity label (simply said, class of objects whose arbitrary `section' is again an instance of the same class) cannot subsume a class with unity label (i.e., containing objects that have `strict boundaries around themselves'). Therefore it is inferred that the individual meta-modelling the class LivingBeing belongs to the diagnostic class Incoherent-Antiunity. 4

Conclusions and Future Work

The B-Annot plugin represents the rst proof-of-concept implementation of annotation technology for ontologies and vocabularies that is (1) not restricted to a single theoretical framework but supports multiple OBMLs, and (2) interconnects the browsing/editing of ontologies (as supported by common ontological editors) with LD summaries. It is a part of a prospective eco-system of tools (other existing ones include, e.g., pattern-based ontology transformation tools [ 6 ]) supporting (informed rather than merely intuitive) reuse and design of ontologies on the semantic web.

Serious usability tests and requirement collection for B-Annot is only planned after some of the envisaged enhancements will have taken place.

A straightforward extension of B-Annot will be the possibility to also store annotations in OWL annotation properties of a copy of the annotated ontology. This will allow for easy browsing of the annotations in their original context.

Background annotation by distinctions referring to notions like `rigid' (in OntoClean) or `particular' (in PURO) risks to discourage even reasonably experienced ontological engineers without philosophical background. The threshold should thus be set as low as possible in the future, via operationalized annotation guidelines. For OntoClean's rigidity alternatives, a promising approach has 12 http://patomat.vse.cz/puro_v1.1.owl 13 http://patomat.vse.cz/ontoclean-coherence-check-1.owl 14 http://patomat.vse.cz/ontoclean-v.1.0.owl already been shown by Seyed, who designed a wizard relying on common-sense verbalization of the meaning of these alternatives [ 5 ]. For the PURO OBML distinctions, textual guidelines with examples have already been designed and tested in an classroom assignment; the experience gained will be used to design verbalisation templates similar to those from [ 5 ].

As the amount of mature vocabularies and their stable entities is still low15 their purely manual annotation via B-Annot is feasible. In long term, however, partial automation could be achieved by leveraging on two di erent sources: (1) via linguistic parsing of associated texts, especially the values of rdfs:comment, and, (2) via logically inferring the most likely annotations based on previously assigned annotations of interrelated entities, e.g., from superclasses to subclasses. This work has been supported from the EU ICT FP7 under no. 257943 (LOD2 project), from the VSE IGA project no. 34/2014, from the Slovak VEGA project no. 1/1333/12, and from project APVV-0513-10.

References

1. Auer , S. , Demter , J. , Martin , M. , Lehmann , J.: LODStats An Extensible Framework for High-Performance Dataset Analytics . In: EKAW 2012 , Galway, Springer LNCS 7603.

2. Glimm , B. , Rudolph , S. , Volker, J.: Integrated metamodeling and diagnosis in OWL 2 . In: Proc. ISWC 2010 .

3. Guarino , N. , Welty , C. : An Overview of OntoClean . In: Staab, S. , Studer , R., eds.: The Handbook on Ontologies , pp. 151 { 172 , Springer-Verlag, 2009 .

4. Presutti , V. , et al.: Extracting core knowledge from Linked Data . In: Proceedings of the Second Workshop on Consuming Linked Data , COLD2011 . ( 2011 )

5. Seyed , P. : A Method for Evaluating Ontologies { Introducing the BFO-Rigidity Decision Tree Wizard . In: FOIS 2012 : 191 { 204 .

6. Svab-Zamazal , O. , Dudas , M. , Svatek , V. : User-Friendly Pattern-Based Transformation of OWL Ontologies . In: Proc. EKAW'12 , Galway .

7. Svatek , V. , Homola , M. , Kluka , J. , Vacura , M. : Ontological Distinctions for Linked Data Vocabularies . Technical Report TR-2013-039 . Comenius University, Bratislava, 2013 . Available online: http://kedrigern.dcs.fmph.uniba.sk/ reports/display.php?id= 54

8. Svatek , V. , Homola , M. , Kluka , J. , Vacura , M. : Mapping Structural Design Patterns in OWL to Ontological Background Models . In: Proc. K-CAP 2013 , ACM.

9. Svatek , V. , Homola , M. , Kluka , J. , Vacura , M. : Metamodeling-Based Coherence Checking of OWL Vocabulary Background Models . In: Proc. OWLED 2013 , online http://ceur-ws. org/ Vol- 1080 /owled2013_6.pdf.

10. Welty , C. : OntOWLClean: Cleaning OWL ontologies with OWL . In: Proc. FOIS 2006 .

15 The statistics at http://lov.okfn.org/dataset/lov/stats/ reveals that out of the several thousand entities referenced in LD, there are only about 150 that are at the same time reused by more than one other vocabulary and instantiated by at least 100 LOD instances .