=Paper=
{{Paper
|id=Vol-1151/paper5
|storemode=property
|title=Towards Semantic Dataset Profiling
|pdfUrl=https://ceur-ws.org/Vol-1151/paper5.pdf
|volume=Vol-1151
|dblpUrl=https://dblp.org/rec/conf/esws/EllefiBST14
}}
==Towards Semantic Dataset Profiling==
https://ceur-ws.org/Vol-1151/paper5.pdf
Towards Semantic Dataset Profiling
Mohamed Ben Ellefi, Zohra Bellahsene, François Scharffe, Konstantin Todorov
{firstname.lastname@lirmm.fr}
LIRMM / University of Montpellier 2, France
Abstract. The web of data is growing constantly, both in terms of size
and impact. A potential data publisher needs to dispose with recapitu-
lative information on the datasets available on the web, so that she can
easily identify where to look for the resources to which her data relates.
This information will help discover candidate datasets for interlinking. In
that context, we investigate the problem of dataset profiling. We define
a dataset profile as a set of characteristics, both semantic and statistical,
that allow to describe in the best possible way a dataset by taking into
account the multiplicity of domains and vocabularies on the web of data.
Keywords: Linked data, Dataset discovery, Dataset profiling.
1 Introduction
”With linked data, when you have some, you can find other, related data”. This
is a simplified view of the fourth principle of linked data1 . Finding resources that
can be reused or linked requires a framework for comparison between datasets.
We propose to identify a dataset by its profile – a set of characteristics that
allow to describe this dataset in the best possible way. In our understanding,
a dataset profile should be based on three main criteria. (1) It should combine
a versatile set of features that describe a dataset, (2) it should be generated in
domain independent manner, i.e., the generation procedure should be applicable
to any dataset from any domain, and (3) It should be generated automatically. A
dataset usually relies on multiple different vocabularies to describe resources. A
profile should reflect these vocabularies allowing for the application of ontology
matching techniques in the dataset discovery and interlinking task.
There has been relatively little research dedicated to this task. Fetahu et al.
[1] propose an approach for creating structured dataset profiles, where a profile
describes the topic coverage of a particular dataset. In the topic extraction, the
full textual content of a resource is analyzed from all its literals. Then, DBpedia
Spotlight2 is used as a named entity recognition and disambiguation tool.
Böhm et al. [2] introduce the notion of k-similarity, where two resources
are k-similar, if k of their property/value combinations are exact matches. The
intuition is that two resources are similar to some degree if they share a common
set of attributes, and could therefore be related. The k-similarity approach can
be seen as similar to a dataset profiling technique based on k property/value
combinations.
1
http://www.w3.org/DesignIssues/LinkedData
2
http://spotlight.dbpedia.org
2 Ben Ellefi et al.
Atencia et al. [3] introduce a method for analyzing datasets based on key
dependencies. This approach is inspired by the notion of a key in relational
databases. A key in a dataset is a set of property/value pairs indicating that any
resource in this dataset will have a unique set of values for a given set of proper-
ties. This definition of a key tolerates a few instances having the same values for
the properties, what the authors have named a ”pseudo-key” – a relaxed version
of a key on the basis of a discriminability threshold. The pseudo-keys can be
used to select sets of properties/values with which to compare resources issued
from different datasets. Such a set of properties/values can be seen as a set of
dataset features forming a profile.
While Fetahu et al. propose an automatic and domain independent profiling
technique based on topics, the extracted profile is not intended to the comparison
task. The other methods do not define the dataset profiling problem explicitly.
Böhm et al. adopt an automatic selection of properties and the dataset profile is
fixed after the comparison task. In particular, cross-vocabulary properties map-
ping is not performed which limits the space of semantic comparison. Atencia et
al. appears to have less of the flaws of the other two methods, the set of prop-
erties being selected automatically and by taking into account the cross-domain
context. However, the problem of cross-vocabularies pseudo-keys comparison is
not discussed. Consequently, none of these techniques satisfies all the criteria and
none of them addresses the main cross-vocabularies comparison problem. In next
section, we give our proposition for a dataset profile based on these comparison
criteria.
2 A Generic Framework for Dataset Profiling
A variety of characteristics can be included in a definition of a profile, such as a
set of property/values pairs, types, topics or statistics. Several questions arise:
What are the most representative features for a dataset? Are these features
sufficient with respect to the dataset comparison task and where to look for
them? We suggest that there are two main types of information that are relevant
when constructing a dataset profile.
The first type is based on the declarative description of statistical dataset
characteristics, such as property usage, vocabulary usage, datatypes used and
average length of string literals, size, coverage, discriminability, frequency, etc.
Thus, a framework like the LODStats [4] can be used as a large-scale dataset
analytics which allows this kind of dataset profiling.
The second type is based on a set of types (schema concepts) that repre-
sent the topic and the covered domain. A set of properties/values describes the
semantics of a dataset and enables the semantic comparison on an instance level.
The first type is appropriate to obtain information with regard to the struc-
ture, coverage and coherence of data. These statistical characteristics can be
helpful to evaluate the dynamicity of a dataset in order to optimise reuse, re-
vision or query tasks. In the current study, we are interested in characteristics
of the second type only, since they provide explicit semantic information, useful
for the dataset discovery task.
Towards Semantic Dataset Profiling 3
Fig. 1. Dataset Profiling Workflow
Our proposal is described in Figure 1. First, we select the k-most frequent
concepts (types). We are inspired by [2] where the dataset content is summarized
by the top ten discovered types. Then, we extract all instances associated to the
selected set of concepts. Thereafter, the pseudo-keys generation technique [3]
is applied to these instances, resulting in a set of property/values pairs. Thus,
we can surpass the complexity problem of the pseudo-keys and we take advan-
tage of the automaticity and the cross-domain features of the approach, assuring
that the profile fulfills criteria 2 and 4. Note that for selecting a set of proper-
ties/values pairs, we can apply statistical measures like the property entropy [5],
discriminability and frequency, thus including statistical information of the first
type in the profiling process.
Finally, we address the problem of integrating cross-vocabulary datasets
based on their profiles. The generated profiles are described by cross-vocabulary
features. Basca et al. [6] defines a vocabulary as a simple ”lightweight” ontology.
Hence, in order to be able to compare profiles and measure similarities, a sys-
tem needs to ”know” the correspondences between the types of resources. For
example, we consider two datasets, one described using FOAF, the other using
VCard. When comparing resources of these types, the properties foaf:givenname
should be compared to vcard:fn, as well as the property foaf:familyname - to
the property vcard:ln. The proposition is to establish correspondences between
the different features described by different vocabularies allowing for more pre-
cise semantic comparison and, consequently – for a more representative seman-
4 Ben Ellefi et al.
tic dataset profile. D’Aquin et al. [7] propose the manual alignments between
schemas as that seems less costly than an automatic alignment and they consider
only the most frequent types and properties. Our tendency is converging towards
the automatic alignment for only popular vocabularies, e.g., http://schema.org/.
3 General Discussion
Due to the inherent heterogeneity of linked data, an efficient profiling technique
is necessary in the context of dataset comparison and candidate dataset iden-
tification for the interlinking process. Our study of the existing techniques has
helped identify several problems that have to be addressed in future. Most im-
portantly, the extraction of profile features has to be automated as much as
possible and made applicable for all domains and all vocabularies. In order to
deal with a dataset comparison based on the semantic content, the profile fea-
tures have to be semantically descriptive. The very question of the actual set of
features that best describe a dataset is a subject of ongoing analysis.
Our proposal potentially meets all criteria listed in the introductory section.
We repose on both semantic and statistical features and show their connexion.
The process of semantic features generation is totally independent of the dataset
domain and can be performed automatically on the basis of discriminability,
frequency and support. Finally, the use of ontology matching techniques, ensures
the compatibility of profiles defined for cross-vocabulary datasets comparison.
In future, we plan to implement our proposal and test it on existing datasets
on the web of data and in the context of the Datalyse project3 .
References
1. B. Fetahu, S. Dietze, B. Pereira Nunes, M. Antonio Casanova, D. Taibi, and W. Ne-
jdl, “A scalable approach for efficiently generating structured dataset topic profiles,”
in In Proceedings of the 11th Extended Semantic Web Conference, Springer, 2014.
2. C. Böhm, J. Lorey, and F. Naumann, “Creating void descriptions for web-scale
data,” J. Web Sem., vol. 9, no. 3, pp. 339–345, 2011.
3. M. Atencia, J. David, and F. Scharffe, “Keys and pseudo-keys detection for web
datasets cleansing and interlinking,” in EKAW, pp. 144–153, Springer, 2012.
4. S. Auer, J. Demter, M. Martin, and J. Lehmann, “Lodstats – an extensible frame-
work for high-performance dataset analytics,” EKAW, p. 353, 2012.
5. T. Gottron, M. Knauf, S. Scheglmann, and A. Scherp, “A systematic investigation
of explicit and implicit schema information on the linked open data cloud,” The
Semantic Web: Semantics and Big Data, pp. 228–242, 2013.
6. C. Basca, S. Corlosquet, R. Cyganiak, S. Fernndez, and T. Schandl, “Neologism:
Easy vocabulary publishing,” in SFSW, pp. CEUR–WS.org/Vol–368, 2008.
7. M. D’Aquin, A. Adamou, and S. Dietze, “Assessing the educational linked data
landscape,” in (WebSci), Paris, France, pp. 43–46, ACM, 2013.
3
This research is funded under the Datalyse project (http://www.datalyse.fr/)