The amount of Linked Data has been growing increasingly. However, the e cient use of that knowledge is hindered by the lack of information about the data structure. This is re ected by the di culty of writing SPARQL queries. In order to improve the user experience, we propose an auto-completion library1 for SPARQL that suggests possible RDF terms. In this work, we investigate the feasibility of providing recommendations by only querying the SPARQL endpoint directly.
Over the years, many contributions have been done towards facilitating the use of
SPARQL, either visually [
We propose a data-based auto-completion which retrieves possible items with
regards to the current state of the query. Recommended items can be predicates,
classes, or even named graphs. Firstly, we indicate the position in the SPARQL
query that is to be auto-completed, i.e., the Point Of Focus (POF), by inserting the
character `<'. Secondly, we reduce the query down to its recommendation scope [
Recommendation Scope. While building a SPARQL query, not all triple patterns are
relevant for the recommendation. Therefore, we de ne the scope as the connected
component that contains the POF. Figure 1a depicts a SPARQL query where the
POF is associated with the variable \?s": it seeks possible predicates that occur with
a \:Person" having the predicate \:name". Figure 1b depicts the previous SPARQL
query reduced to its recommendation scope. Indeed, the pattern on line 4 is removed
since it is not part of the connected component containing the POF.
Recommendation Capabilities. The scope may include content-speci c terms, e.g,
resources and lters, unlike to [
Systems. In this section, we evaluate the recommendations returned by the proposed
system, that we refer to as \S1", against the ones provided by the approach in [
Settings. We compare the recommendations with regards to (1) the response-time,
i.e., the time spent on retrieving the recommendations via a SPARQL query; and
(2) the quality of the recommendations. A run of the evaluation consists of the
following steps. First, we vary the amount of information retrieved via the \LIMIT"
clause. Then, we compare the ranked TOP-10 recommendations against a gold
standard. The ranking is based on the number of occurrences of a recommendation. The
gold standard consists in retrieving recommendations directly from the data without
the LIMIT clause, and retaining only the 10 most occurring terms. The TOP-10 of
the gold standard and the system are compared using the Jaccard similarity. We
consider that the higher the similarity, the higher the quality of recommendations.
Queries. We used the query logs of the DBpedia endpoint version 3.3 available
from the USEWOD20132 dataset. The queries3 were stripped of any pattern about
speci c resources, in order to keep only the structure of the query. In addition,
we removed queries that contain more than one connected component. Queries are
grouped according to their complexity, which depends on the number of triple
patterns and on the number of star graphs. A group is identi ed by a string that has
as many numbers as there are stars, with numbers separated by a dash '-' and
representing the number of triple patterns in a star. For example, a query with two
stars and one triple pattern each is then identi ed with 1-1. This de nition of query
complexity exhibits the potential errors, i.e., a recommendation having zero-result,
that a graph summary can have, as described in [
Graphs. We loaded into an endpoint the English part of the Dbpedia3.34 dataset, which consists of 167 199 852 triples. The graph summary consists of 29 706 051 triples, generated by grouping resources sharing the same set of classes. Endpoint. We used a Virtuoso5 SPARQL endpoint. The endpoint is deployed on a server with 32GB of RAM and with SSD drives.
Comparison. For each group of query complexity QC, we report in Table 1 the
results of the evaluation, with J 1 (resp., J 2) the average Jaccard similarity for the
system S1 (resp., S2); and T 1 (resp., T 2) the average response-time in ms for the
system S1 (resp., S2). The reported values are the averages over 5 runs. We can see
that as the LIMIT gets larger, the higher the Jaccard similarity becomes.Since the
graph summary used in S2 is a concise representation of the graph structure, the
data sample at a certain LIMIT value contains more terms than in S1. However,
this impacts negatively on the quality of S2 as re ected by the values of J2. This
shows the graph summary is subject to errors [
T 1 T 2
2 107 81 108 81 141 91
1-3 101 108 103 105 126 105
This material is based upon works supported by the European FP7 projects LOD2 (257943).