This paper overviews the participation of Scalewelis in the QALD-3 open challenge. Scalewelis is a Faceted Search system. Faceted Search systems re ne the result set at each navigation step. In Scalewelis, re nements are syntactic operations that modify the user query. Scalewelis uses the Semantic Web standards (URI, RDF, SPARQL) and connects to SPARQL endpoints.
SELECT DISTINCT ?class WHERE {
[] a ?class . }
Query-based Faceted Search allows for more expressive search than traditional
set-based Faceted Search. In Sewelis and Scalewelis, facets are query elements
and selecting a facet adds it to the search query. Only valid facets are provided,
so that navigation is safe: there are no dead ends. Facets are computed from the
results of the previous query. Scalewelis uses only a subset of those results. In
experiments on datasets of up to 15 millions triples, automatically generated with
the Berlin SPARQL Benchmark [
At the initial step, Scalewelis retrieves all classes in the dataset (Figure 1). Properties can be computed in a similar manner but are not in the current implementation for e ciency reasons. In the next steps, partial results and facets are computed with four successive SPARQL queries (Figure 2). Query 1 retrieves the rst 1,000 results matching <Pattern>, which is the SPARQL graph pattern translated from the user query. Queries 1, 2, and 3 incorporate those results in order to retrieve the facets for the current query, using the VALUES construct of SPARQL. Query 2 retrieves the types of the results, i.e. classes. Query 3 retrieves the properties linking from the results. Query 4 retrieves the property linking to the results, i.e. inverse properties. Finally, the selection of a facet by the user modi es the query so that it is more constrained but is ensured to return results. 3
Regarding the QALD-3 open challenge, Scalewelis was connected to the standard DBpedia SPARQL endpoint because response times of the provided endpoint were too long for e ective use. The process is intrinsically interactive and we acted as the user. We answered 70 questions out of 99, requiring a few minutes per question (human and machine time). Two kinds of questions were not addressed. Firstly, questions involving quantities (such as Which German cities have more than 250000 inhabitants?) because lters and aggregations are not yet implemented in Scalewelis. Secondly, questions for which facets were not found by the user. In general, it was unclear whether those facets were not listed (due to partial results) or whether the user was not knowledgeable enough to identify the right facet (for example Give me all B-sides of the Ramones.).
One of the question in the challenge is Which lms starring Clint Eastwood did he direct himself ?. We explain how we answered this question with Scalewelis. At the initialization step, the user ltered the list of classes with the word lm. The selection of the class a dbo:Film led to the query What is a dbo:Film?. At this step, there are property facets and the selection of has dbo:starring and has dbo:director led to the query What is a dbo:Film and has dbo:starring a thing and has dbo:director a thing?. Finally, the selection of the result Clint Eastwood at the two unde ned element represented by a thing led to the nal user query: What is a dbo:Film and has dbo:starring dbr:Clint Eastwood and has dbo:director dbr:Clint Eastwood?. 4
Scalewelis ranked third out of six candidates with 32 correct answers plus 1 partial answer. Scalewelis was theoretically able to correctly answer the 70 questions the user chose to answer, but he failed to answer 37 of them because incorrect facets were chosen. 1. Firstly, English is not the mother tongue of the user so that incorrect facets were used because of misunderstandings. 2. Secondly, the user had sometimes the choice between two facets di ering only by their namespace, for example DBpedia ontology and infobox properties. 3. Thirdly, the standard DBpedia SPARQL endpoint to which Scalewelis was connected sometimes gave di erent results compared to the provided SPARQL endpoint.
Scalewelis can be improved in several ways. First, there are missing functionalities from SPARQL such as comparisons or aggregations. Second, the partial computation of results and facets entails a bias in the search, because the rst results returned by SPARQL may not form a representative sample of all results. Third, Scalewelis could take advantage of the DBpedia ontology to further improve scalability, and to display facets as class/property hierarchies.