Automated Exploration of Ontology Repositories OndřejZamazal ondrej.zamazal|svatek@vse.cz University of Economics
W. Churchill Sq.4 130 67 Prague 3 Czech Republic
VojtěchSvátek University of Economics
W. Churchill Sq.4 130 67 Prague 3 Czech Republic
Automated Exploration of Ontology Repositories 3D7535711E152E6E348957186B3FBE8E GROBID - A machine learning software for extracting information from scholarly documents

Motivation The choice of adequate ontology repository is an important prerequisite to finding an ontology to be reused or adapted for a concrete use case. As the repositories are mostly affiliated to particular communities within the semantic web, understanding the typical features of ontologies in each of them is also helpful for designers of ontology management tools.

Overall Process, Metrics and Results Our ontology exploration process includes ontology collection, materialization and then metrics computation; finally, the resulting metrics are explored using the R language1 to automatically get a summary report in the form of tables. To automate the collection phase, we partly employed Ontohub, 2 which is an open ontology repository mirroring several other repositories. The materialization includes ontology storing (into the database) in order to decompose them into entities, names, relations, imported ontologies and head nouns. We use the OWL-API3 to manipulate the ontologies.

We considered metrics related to four aspects of ontologies. 4 Logical and structural metrics include, e.g., the numbers of different types of entities and axioms. We also categorize the complexity of ontologies into bins (as in [2, 3]). The naming aspect reflects some basic information regarding the length of class name (local fragments of URI or labels), capitalization and usage of concatenation symbol/technique, i.e. a hyphen, underscore, camel-case or dot (as in [1]). For the annotation aspect we compute the proportions of RDFS annotations.

We explored ontologies from five prominent ontology repositories (Table 1 contains just a few selected metrics). Due to parsing problems, unavailability of ontologies or their imports we however did not collect all ontologies from the repository. BioPortal5 is a web portal providing access to a library of wellcurated biomedical ontologies via REST-ful services. It contains ontologies from another ontology repository, the OBO Foundry. 6 We collected ontologies using the Ontohub mirror where only ontologies with size below 5MB (thus only 342 of the total) are available. 7 The Dumontier lab ontologies 8 are biological ontologies aimed at knowledge representation and reasoning. Their ontologies are quite interconnected (many mutual imports). LOV 9 is a well-curated collection of linked open vocabularies used in the Linked Data Cloud. The Protégé ontology library mostly contains ontologies developed within the Protégé editor. As there is no programmatic access to the library, we manually downloaded them. It turned up that out of 93 ontologies (except Dumontier ontologies on which there is also a link) 43% ontologies were not available. Finally, the TONES repository (using its Ontohub mirror of 207 ontologies -collected 88%) contains ontologies of various domains, many of them however designed for testing purposes.

Future Work

We plan to run such an analysis repeatedly, include more repositories (preferably via Ontohub) and more metrics. We also want to keep the ontology exploration services available via a web interface 10 where the users could ask, on the one hand, for the latest summaries of particular repositories, and on the other hand for particular ontologies or ontologies meeting some criteria.

Ondřej Zamazal has been supported by the CSF grant no. 14-14076P and this research is also supported by UEP IGA project F4/34/2014 (IG407024).

Table 1 .Selected metrics. Average (Avg) is either mean or median, according to better representativeness. The min statistics is omitted since it is always zero. The max statistics is omitted for ratios because it is nearly always 100%. The larges value across all repositories is in bold.Metrics(June 2014 snapshot)BioPortal Dumontier LOV Protégé TONESOntologies processed2547035341183Percentage of all74%95% 83%44%88%Complex class using existential restr. Avg57%28%7%14%43%Complex class as superclassAvg74%35%69%57%67%BranchingAvg0.880.550.480.610.79Max2.391.091.771.431.78Multiple inheritanceAvg3201612Max187725432149724800Annotation as labelAvg38%51% 32%13%38%Annotation as commentAvg1%37% 25%49%37%Camel techniqueAvg15%61% 39%28%29%Underscore techniqueAvg54%0%0%23%36%
http://www.r-project.org/ https://ontohub.org/ http://owlapi.sourceforge.net/ Due to the space limitation full list of metrics and complete results are at the supplementary web page: http://owl.vse.cz:8080/MetricsExploration/ 5 http://bioportal.bioontology.org / 6 http://obofoundry.org/
A Survey of Identifiers and Labels in OWL Ontologies NA A MManaf SBechhofer RStevens OWLED 2010 A Snapshot of the OWL Web NMatentzoglu SBail BParsia ISWC 2013 A Survey of the Web Ontology Landscape TDWang BParsia JHendler ISWC-2006 To overcome 5MB limitation we gathered BioPortal ontologies directly by RESTful services Corresponding ontology metrics are available via the supplementary web A sample service, providing metrics for a given ontology