rdf dataset quality assessment is currently performed primarily after data is published. Incorporating its results, by applying corresponding adjustments to the dataset, happens manually and occurs rarely. In the case of (semi-)structured data (e.g., csv, xml), the root of the violations often derives from the mappings that specify how the rdf dataset will be generated. Thus, we suggest shifting the quality assessment from the rdf dataset to the mapping de nitions that generate it. The proposed test-driven approach for assessing mappings relies on rdfunit test cases applied over mappings speci ed with rml. Our evaluation is applied to di erent cases, e.g., dbpedia, and indicates that the overall quality of an rdf dataset is quickly and signi cantly improved.
Although more and more data is published as Linked Data, there are signi cant
variations in quality [
Taking mappings of data to rdf as a software engineering task, a set of unit
test cases can be assigned to the mappings to ensure the correct generation of
rdf datasets from input data. Incorporating quality assessment as part of the
mapping is essential to prevent the same violations from appearing repeatedly
within the dataset and over distinct entities. After all, structural adjustments
can still be applied in this phase, as violations are identi ed at their root.
Moreover, if mappings are assessed, every other new data source also mapped using
them directly bene ts from the improvements. Therefore, we proposed a uniform
solution [
[R2]RML Mappings Quality Assessment with RDFUnit
Our solution relies on the rdf mapping language (rml) [
RDFUnit [
As [r]rml mappings can be processed as rdf documents, because of their native rdf representation and viewpoint (written as the generated triples), the same set of schema validation patterns normally applied on the rdf dataset is also applicable on the mappings that state how it is generated. Nevertheless, instead of validating the triple's predicate against its subject and object, the predicate is extracted from the Predicate Map and is validated against the Term Maps that generate the subject and object. To achieve this, the properties and classes are identi ed and their namespaces are used to retrieve the schemas and
Test-driven Assessment of [R2]RML Mappings generate the test cases as if they were the actual dataset. The expected value, as derived from the Predicate Map, is compared to the de ned one, as derived from the corresponding Subject Map and Object Map. For example, the extracted predicate is foaf:age normally expects an instance of foaf:Agent type for its domain and an integer datatype for its range, but the Term Map that generates the subject is de ned to be of foaf:Project type and the object is de ned to have a oat value. Its mapping document follows: <#Mapping> rr:subjectMap [rr:template "http://example.com/{id}"; rr:class foaf:Project]; rr:predicateObjectMap [rr:predicate foaf:age; rr:objectMap [rml:reference "age"]].
Corresponding rdfunit test cases and patterns were de ned to apply to the mappings, adjusting the assessment queries.6 The de ned test cases cover all possible alternative ways of de ning equivalent mappings that generate the same triples. rdfunit can annotate test cases by requesting additional variables and binding them to speci c result properties. The test case patterns applied to the aforementioned example and its instantiation are indicatively presented. On the left, the where clause of a sparql template query that assesses the datatype is presented. On the right it is presented how it is instantiated: ?resource %%P1%% ?c.
FILTER (DATATYPE(?c) != %%D1%%) ?resource foaf:age ?c.
FILTER (DATATYPE(?c) != xsd:int) The following is the where clause of the same test case applied to the mapping: ?resource rr:predicateObjectMap ?poMap. ?poMap rr:predicate %%P1%%;
rr:objectMap ?objM. ?objM rr:datatype ?c.
FILTER (?c != %%D1%%) ?resource rr:predicateObjectMap ?poMap. ?poMap rr:predicate foaf:age;
rr:objectMap ?objM. ?objM rr:datatype ?c.
FILTER (?c != xsd:int) 3
Evaluation and Discussion The assessed datasets and corresponding mappings, as well as the assessment results are summarized in Table 1: dbpedia mappings7, after the mappings were converted from wikitext markup to rml8, and its dataset were assessed. dblp 6 https://github.com/AKSW/RDFUnit/blob/master/data/tests/Manual/www.w3.org/ns/r2rml/rr.
tests.Manual.ttl 7 http://mappings.dbpedia.org/ 8 https://goo.gl/GPB1Ar dataset DBpEn DBpNL DBLP iLastic10 CDFLG11 CEUR-WS12 size 62M 21M 12M 150K 0.6K 2.4K dataset assessment time #fail.
16h 1,128 1.5h 683 12h 7 12s 23 7s 15 6s 7 #viol.
3.2M 815K 8.1M 37K 678 783 mappings, after the mappings were converted to rml9, and the corresponding dataset were assessed, too. The results show that the required quality assessment time is signi cantly reduced if the mappings are assessed instead of the rdf dataset, especially in the case of medium/large datasets. That happens because the dataset assessment requires examining each triple separately to identify, for instance, that 12M triples violated the predicate's range, whereas mapping assessment requires only 1 triple to be examined. The e ectiveness of mapping assessments is also high: the identi ed violations can be accurately indicated. Acknowledgements. This paper's research activities were funded by Ghent University, iMinds, the Institute for the Promotion of Innovation by Science and Technology in Flanders, the Fund for Scienti c Research-Flanders and by grants from the EU's 7th & H2020 Programmes for projects ALIGNED (GA 644055), GeoKnow (GA 318159) and LIDER (GA 610782).
10 http://explore.ilastic.be/ 11 http://ewi.mmlab.be/cd/all 12 http://rml.io/rml/data/SPC2015/