In the Semantic Web, the Web Ontology Language (OWL) vocabulary is used for the representation of formal ontologies, while the Simple Knowledge Organisation System (SKOS) is a vocabulary designed for thesauri or concept taxonomies without formal semantics. Despite their di erent nature, on the Web these two vocabularies are often used together. Here, we try to explore and exploit the joint usage of OWL and SKOS. More precisely, we rst de ne usage patterns to detect problematic modeling from connections between SKOS and OWL. Next, we also investigate if additional information can be inferred from joint usage with SKOS in order to enrich semantic inferences through OWL alone { although SKOS was designed without formal semantics, we argue for this heretic approach by applicability \in the wild": the patterns for modeling errors and inference of new information are transformed to SPARQL queries and applied to real world data from the Billion Triple Challenge 2014; we manually evaluate this corpus and assess the quality of the de ned patterns empirically.
While each language, SKOS and OWL, and their fragments serve di erent use cases in theory, in the Linked Data world, ie., \in the wild", SKOS and OWL are often used together. Two of the questions arising from joint usage are: (i) how can the interplay of SKOS and OWL help to nd modeling problems and quality issues within the data, and (ii) can we nd recipes to infer additional formal information from joint usage patterns? In order to address these questions we de ne patterns { in the form of SPARQL queries { to detect cases of modeling problems and potential additional entailments. While the detection of modeling errors is rather straightforward, inferring new information may be seen as heretic, because the design of SKOS is very informal and open on purpose.
We would like to emphasize that, while our attempt to suggest inferences
from joint usage from OWL and SKOS may be viewed as contradicting the
often stressed informal nature of SKOS, previous works on OWL reasoning for
Linked Data have likewise shown that OWL reasoning on Linked Data itself
has to be taken \with a grain of salt", i.e. also OWL inferences themselves may
often be misleading when applied naively on Web data, cf.[
The reasons to visit the posters are: (i) Find out about SKOS-modeling errors appearing in the wild, (ii) learn about potential modeling errors which were observed in joint usage of SKOS and OWL, and (iii) discuss our ideas of inferring additional statements from joint usage of SKOS and OWL. 2
Previous work on the implications of combined usage of SKOS and OWL is
very limited. Related work can be grouped accordingly: (i) Methods to measure
and improve the quality of SKOS (which does not consider joint usage with
OWL), eg. Suominen and Mader [
Error Patterns: Error patterns detect potentially problematic modeling for (linked) data that uses both OWL and SKOS. The patterns are represented by generic SPARQL queries, which can be applied to any triple store / SPARQL endpoint. The patterns are kept simple, and combine SKOS vocabulary with OWL primitives. We applied a number of error patterns, of which ten patterns occurred in the data. Due to space restrictions we only list one example below, at our web site1 we present all patterns for error checking and inference.
EP4: skos:narrowerTransitive and rdfs:subClassOf
EP4 states that at the same time there is a rdfs:subClassOf and a
skos:narrowerTransitive between two resources. The SKOS Reference [
Connections Between SKOS and OWL Inference Patterns: Inferring new formal information from joint usage of OWL and SKOS is the most daring part of our research, because SKOS is deliberately informal by design. The SPARQL CONSTRUCT queries presented in the example below (and many other queries on our web site) suggest new triples to be added to the data set.
IP1: owl:equivalentClass when skos:closeMatch and superClass
?a skos:closeMatch ?b. ?a rdfs:subClassOf ?c.
?b rdfs:subClassOf ?c. } The motivation behind this pattern is that two owl:Class classes having the exact same super-class and being linked via skos:closeMatch are likely to be equivalent classes. This pattern adds a owl:equivalentClass relation between those classes to the triple store. We know that the added relation will not always be correct, all inference patterns are manually checked in the evaluation. 4
To evaluate the patterns, we aimed to use a real world data set published by a variety of di erent sources on the Internet. We decided to apply the patterns to the data from the Billion Triple Challenge 20142.
First, we applied the patterns for detecting modeling problems to the evaluation dataset. Some of the ten patterns (EP1{EP10) occurred more frequently in the dataset, others only a few times. We manually examined the data, all occurrences indeed pointed to modeling problems.
Figure 1 present a screenshot of the Web Frontend, which lets users not only see the SPARQL queries and evaluation results for all patterns, but also access the underlying data matching the patterns.
Secondly, the inference patterns were applied to the evaluation dataset. In summary, all of the presented patterns in the Web Frontend correctly inferred new relations. However, obviously the quality and the amount of the these relations depends heavily on the data set that the patterns are applied to. Some patterns work well with a particular data set while others do not. In the end, our goal was to nd out which connections between SKOS and OWL are used in practice and divide them into ones which arguably look like modeling errors and others which could suggest additional inferences: in cases where we could not detect a clear trend, it should be decided on a per-dataset basis which of these `recipes' to apply or not, which probably remains a manual task or, resp., in the control of the dataset owner.
Our contributions are as follows: (i) Presenting a number of usage patterns for OWL and SKOS usage which help to identify modeling errors in data sets, (ii) provide patterns to infer additional information, and (iii) evaluate the patterns de ned in (i) and (ii) with regards to the Billion Triple Challenge data set to determine their quality. There are many directions for future work, eg. increasing the number and complexity of patterns, and applying them to di erent datasets. Acknowledgments. This work has been supported by project uComp, which receives the funding support of EPSRC EP/K017896/1, FWF 1097-N23, and ANR-12-CHRI-0003-03, in the CHIST-ERA ERA-NET programme.