A Protege Plugin with Swift Linked Data Miner

Abstract

We present a Protege plugin implementing Swift Linked Data Miner, an anytime algorithm for extending an ontology with new subsumptions. The algorithm mines an RDF graph accessible via a SPARQL endpoint and proposes new SubClassOf axioms to the user.

Introduction

It is not uncommon for a Linked Data dataset to provide only a very shallow ontology [ 3 ], which does not cover more complex aspects of the underlying conceptual model. On the other hand, the data in the dataset follows the model and thus the model is re ected in the data in a form of patterns. Such patterns can be detected using pattern mining techniques and used to extend the ontology with new knowledge.

To address this use case, we developed Swift Linked Data Miner (SLDM) [ 8 ]. SLDM is a pattern mining algorithm, which can discover new partial de nitions, i.e. SubClassOf axioms, for a given class. The mined axioms are expressed in OWL 2 EL [ 6 ]. SLDM is an anytime algorithm, which means that it delivers patterns once they are mined and then re nes them. In other words, the longer the algorithm works, the more complex patterns are mined. The algorithm does not require an access to the whole RDF graph at the same time. Instead, it downloads on-demand necessary parts by querying the SPARQL endpoint. To avoid issues with high load put on the endpoint by a complex query, the queries are very simple, consisting only of a single triple pattern and a values clause. Direct usage of a SPARQL endpoint without overloading it is the main di erence between SLDM and former approaches to mining ontologies from RDF graphs (e.g. [ 9,2 ]), which require accessing the whole graph at the same time or posing complex SPARQL queries to the endpoint. Swift Linked Data Miner was implemented as a Java library using Apache Jena [ 5 ] to interact with a SPARQL endpoint and OWL API [ 4 ] to deliver mined axioms. On top of the library, we built a plugin for Protege [ 7 ] simplifying the use of SLDM to just few clicks. Both the source code and a JAR le with the plugin are available in a Git 1 repository at https://bitbucket.org/jpotoniec/sldm.

A typical work ow with the plugin is presented in Figure 1. First, the user con gures SLDM with the basic interface of the plugin (Figure 2) by choosing a class in the class hierarchy view and entering the address of the SPARQL endpoint, and starts SLDM with Run button. SLDM constructs a set of URIs belonging to the selected class, e.g. Astronaut, by posing a SPARQL query with a triple pattern ?x a Astronaut and then operates in two alternating phases of querying the endpoint and mining the obtained triples. In the rst phase, the endpoint is queried about all the triples having an URI from the set in a subject position, by using the following WHERE clause: ?s ?p ?o . VALUES ?s f<<URIs>>g. The triples are then organized into a three-level index, with predicates in the rst level, objects in the second and subjects in the third. Such an order enables e cient access e.g. to all URIs occurring in triples with predicate rdf:type and object Person. During the second phase, the index is scanned to discover the axioms, e.g. if for predicate rdf:type and object Person there are many subjects in the third level, an axiom Astronaut subClassOf Person is mined. If, for a given predicate, no pattern can be found, the corresponding subjects are used as an input to the rst phase of SLDM, to mine more complex axioms. The mined axioms are displayed using a standard Protege interface. An axiom is accompanied there by two buttons: the one with @ symbol to display additional information about the axiom (e.g. corresponding value of the measure used by SLDM) and the one with 3 to add the axiom to the ontology.

Figure 3 presents the Expert interface of the plugin, where the user can netune parameters of SLDM. Field Minimal support sets up a minimal value of support (i.e. the measure used during the mining) which an axiom must achieve in order to be presented to the user. Field Maximum level speci es a maximal number of nested some expressions in a mined pattern. Field Ignored properties enables the user to use regular expressions to specify a set of predicates to ignore, 1 https://git-scm.com/ 1 e.g. if a predicate conveys provenance information and is not directly relevant to the semantics of the selected class. If the graph in the SPARQL endpoint is big, it may be useful to use random sampling to decrease the load and increase the speed of the mining. Field Sample size allows the user to set how many di erent URIs from a given class will be considered. To ensure repeatability of sampling, eld Random seed enables the user to set a seed for a random number generator. Field Max VALUES size limits the number of URIs speci ed in a values clause of a single query. 3

Proposed Demo

During the demo we will present how to use the plugin using the DBpedia ontology [ 1 ] and a SPARQL endpoint with the DBpedia dataset loaded. To ensure smooth operation, we will provide our own SPARQL endpoint. We will also discuss how various settings of the parameters a ect the obtained axioms. The interested attendees will be able to use the plugin to mine an ontology of their choice. A short video similar to what will be presented during the demo is available at https://www.youtube.com/watch?v=ENdNQ8ESlEk. 4

Conclusion

In this paper, we presented a plugin to Protege enabling the user to discover new SubClassOf axioms directly from an on-line Linked Data dataset accessible via a SPARQL endpoint using Swif Linked Data Miner [ 8 ]. The basic con guration of the plugin is very simple and does not require prior expertise in data mining from the user. The plugin along with its source code is freely available at https: //bitbucket.org/jpotoniec/sldm.

Acknowledgement. Jedrzej Potoniec acknowledges the support from the Polish National Science Center (Grant No 2013/11/N/ST6/03065). This work was partially supported by the PARENT-BRIDGE program of Foundation for Polish Science, co- nanced from European Union, Regional Development Fund (Grant No POMOST/2013-7/8). Agnieszka Lawrynowicz acknowledges the support from the Polish National Science Center (Grant No 2014/13/D/ST6/02076).

References

1. Bizer , C. , Lehmann , J. , et al.: DBpedia - A crystallization point for the Web of Data . J. Web Sem . 7 ( 3 ), 154 { 165 ( 2009 )

2. Buhmann, L., Lehmann , J.: Pattern based knowledge base enrichment . In: Alani, H. , Kagal , L. , et al. (eds.) The Semantic Web - ISWC 2013. Lecture Notes in Computer Science , vol. 8218 , pp. 33 { 48 . Springer ( 2013 )

3. Glimm , B. , Hogan , A. , Krotzsch, M. , Polleres , A. : OWL: yet to arrive on the web of data? In: LDOW . CEUR Workshop Proceedings , vol. 937 . CEUR-WS.org ( 2012 )

4. Horridge , M. , Bechhofer , S.: The OWL API: A java API for OWL ontologies . Semantic Web 2 ( 1 ), 11 { 21 ( 2011 )

5. McBride , B. : Jena: A semantic web toolkit . IEEE Internet Computing 6 ( 6 ), 55 { 59 ( 2002 )

6. Motik , B. , Grau , B.C. , Horrocks , I. , Fokoue , A. , Wu , Z. : OWL 2 web ontology language pro les (second edition) . W3C recommendation , W3C (Dec 2012 ), http://www.w3.org/TR/2012/REC-owl2-pro les- 20121211 /

7. Noy , N.F. , Sintek , M. , et al.: Creating Semantic Web Contents with Protege-2000. IEEE Intelligent Systems 16 ( 2 ), 60 { 71 ( 2001 )

8. Potoniec , J. , Jakubowski , P. , Lawrynowicz , A. : Swift Linked Data Miner: Anytime Algorithm for Mining OWL 2 EL Class Expressions Directly from On-Line Linked Data, submitted to the J . of Web Semantics, available at https://goo.gl/HFghXp

9. Volker, J., Niepert , M. : Statistical schema induction . In: Antoniou, G. , Grobelnik , M. , et al. (eds.) The Semantic Web: Research and Applications. Lecture Notes in Computer Science , vol. 6643 , pp. 124 { 138 . Springer ( 2011 )