Content uploaded by Dieter Merkl
Author content
All content in this area was uploaded by Dieter Merkl on Dec 19, 2013
Content may be subject to copyright.
Content uploaded by Dieter Merkl
Author content
All content in this area was uploaded by Dieter Merkl
Content may be subject to copyright.
Content uploaded by Doron Goldfarb
Author content
All content in this area was uploaded by Doron Goldfarb
Content may be subject to copyright.
Art History on Wikipedia, a
Macroscopic Observation
Abstract
How are articles about art historical actors interlinked
within Wikipedia? Lead by this question, we seek an
overview on the link structure of a domain specific
subset of Wikipedia articles. We use an established
domain-specific person name authority, the Getty
Union List of Artist Names (ULAN), in order to
externally identify relevant actors. Besides containing
consistent biographical person data, this database also
provides associative relationships between its person
records, serving as a reference link structure for
comparison. As a first step, we use mappings between
the ULAN and English Dbpedia provided by the Virtual
Internet Authority File (VIAF). This way, we are able to
identify 18,002 relevant person articles. Examining the
link structure between these resources reveals
interesting insight about the high level structure of art
historical knowledge as it is represented on Wikipedia.
Keywords
Wikipedia; Art History; Visualization; Semantic Web
ACM Classification Keywords
J.5 Arts and Humanities: Fine arts
H.3.1 Content Analysis and Indexing: Thesauruses
H.3.4 Systems and Software: Semantic Web
Copyright is held by the author/owner(s).
WebSci 2012, June 22–24, 2012, Evanston, Illinois, USA.
ACM 978-1-4503-1228-8
Doron Goldfarb
Vienna University of
Technology
Favoritenstrasse 9-11
Vienna, 1040, Austria
doron@ec.tuwien.ac.at
Max Arends
Vienna University of
Technology
Favoritenstrasse 9-11
Vienna, 1040, Austria
max@ec.tuwien.ac.at
Josef Froschauer
Vienna University of
Technology
Favoritenstrasse 9-11
Vienna, 1040, Austria
josef@ec.tuwien.ac.at
Dieter Merkl
Vienna University of
Technology
Favoritenstrasse 9-11
Vienna, 1040, Austria
dieter@ec.tuwien.ac.at
2
Introduction
While the structure of Wikipedia as a whole has been
extensively studied, more focused observations of
domain specific subsets of its content are often limited
to qualitative comparisons on an individual article level,
or on the extraction/analysis of underlying category
schemes or taxonomies. With a growing crowd actively
participating in the Wikipedia project, the amount of
articles tied to specific knowledge domains has reached
an extent that makes it feasible to observe their large-
scale structure within defined domain borders.
Considering the domain of Cultural Heritage, a number
of institutions are increasingly becoming aware of
Wikipedia as a valuable tool to provide context for
museum artifacts [4]. Indeed, the number of art history
related resources on Wikipedia is impressive, e.g.,
searching only for distinct resources belonging to the
category yago:Painter1103916531 already returns
14,138 articles as of February 2012. There are,
however, lots of different types of Wikipedia articles
dealing with art history. There are articles about
famous artworks (e.g. Mona Lisa), about artistic styles
(e.g. Renaissance), about museums (e.g. Louvre) or
about art historical actors (e.g. Leonardo da Vinci).
Especially articles of the latter type are suited to a
large-scale structural observation, as biographies of art
history related people are highly interlinked due to
student/teacher, artist/patron or influential relations.
In this study, we therefore focus on Wikipedia articles
about art history related persons. We want to examine
how they are related to each other and if there are
interesting structural properties or groupings of inter-
1 http://www.mpi-inf.mpg.de/yago-naga/yago/
article links, such as if they reflect common attributes.
As a first step, we want to approach the analysis by
using visualization techniques.
But who are art historical actors? While it may be clear
that a significant proportion are (groups of) artists like
painters, sculptors, architects etc., there are also many
other persons that have played major roles in the
domain, such as important scholars, collectors, patrons,
politicians or monarchs. This variety of different person
roles makes it difficult to identify relevant persons. We
therefore make use of a well-established external
authority control, the Getty Union List of Artist names
(ULAN)2, in order to find art history related person
biographies in extracted Wikipedia articles on Dbpedia.
Besides supporting identification, this vocabulary also
offers consistent biographical attributes such as
birth/death-date, role or nationality, which is not
always the case with Dbpedia [1]. Moreover, the ULAN
provides associative relationships between its person
records, which serve as a reference for comparison with
the inter-article links between Wikipedia articles.
We make use of another authority, the Virtual Internet
Authority File (VIAF)3, for mapping ULAN records to
Dbpedia resources. As of February 2012, there exist
18,002 of such mappings in VIAF. Since we use the
Resource Description Framework (RDF) for the mapping
between ULAN, VIAF and Dbpedia, this project also
serves as an example on how Semantic Web
technologies can be effectively used for combining
different data sources, enhancing the quality of the
data corpus and thus leading to interesting results.
2 http://www.getty.edu/research/tools/vocabularies/ulan
3 http://www.oclc.org/research/activities/viaf
3
Related Work
The large-scale structure of Wikipedia has been
described in [6], pointing out that Wikipedia inter-
article links form a scale-free network. A study focused
on Wikipedia link structure is presented in [2]. The
suitability of Wikipedia for deriving domain-specific
thesauri is demonstrated in [3] by comparing its
coverage with a professional thesaurus in the domain of
agriculture. In [5], person related Wikipedia articles are
analyzed for the domains philosophy and literature,
with articles chosen by the people's occupation. The
gathered information is used to show for the two
domains which countries/cities have been important
centers during the centuries. A comparison between
ULAN and Dbpedia is presented in [1]. The authors
were able to match 14,972 Dbpedia resources to their
ULAN counterparts. Regarding Dbpedia coverage of
biographical attributes, they were able to identify 8,077
(53.95%) attributes of type birth date, 9,628 (64.31%)
of type nationality and 5,442 (36.35%) of type role.
These findings encouraged us to rely on the ULAN as
source for biographical attributes for the analysis of the
link structure of the observed Wikipedia articles.
Sytem Setup
We use RDF as data format, as it is the native format of
the Dbpedia dataset, representing the core data model
underlying the Semantic Web Vision. Like Dbpedia, we
use Openlink Virtuoso4 as storage platform. Using a tool
provided by the MultimediaN N9C Eculture5 project, we
first converted the ULAN dataset to RDF. As the ULAN
identifiers are by now completely integrated into VIAF,
we were able to retrieve the VIAF RDF data for each of
4 http://virtuoso.openlinksw.com
5 http://e-culture.multimedian.nl
the ULAN records. Using the provided mappings, we
downloaded the 18,002 RDF datasets from the public
English Dbpedia Sparql endpoint. As this endpoint
currently does not provide the internal Wikipedia links,
we downloaded the complete internal wiki link dump
from the Dbpedia website. We extracted all relevant
wiki links between the 18,002 entities and loaded them
into our triple store.
A First Look at the Data
After gathering all the necessary data sources, we had
a first look at the resulting structure by making a
visualization using the Gephi6 Platform. Figure 1 shows
a force based layout of the network of all 18,002
Dbpedia resources and their 54,590 mutual links,
where a node's size reflect its degree (in&out). The blob
in the center suggests the existence of a giant
connected component. Indeed, treating all links as
undirected yields 4,187 different weakly connected
components (CC), with a giant component - colored
green in Fig. 1 - consisting of 75.44% (~13,581) of all
articles, the rest being distributed across the remaining
4,186 CCs whose sizes range between 0.06% (~11)
and single unconnected articles. Treating all links as
directed yields 10,071 strongly connected components
(SCC), with a giant SCC consisting of 40.55% (~7,300)
of all articles, the remaining 10,070 SCCs counting
between 0.05% (~9) and single unconnected articles.
Looking at the node degree distribution, the link
structure of the selected subset of articles clearly shows
a Power Law distribution, as it has been identified with
Wikipedia in general [6] - see Fig. 2. Observing the
distribution of node in- and out-degree, we find
comparable values as in [2].
6 www.gephi.org
Figure 1: Visualization of 18,002
Dbpedia resources
Figure 2: Power Law distribution
of domain specific Wiki links
1
10
100
1000
10000
1
10
100
1000
Number of Nodes with degree d
Degree d
Out-Degree
In-Degree
4
Temporal Person-Person Link Properties
The network visualization in Fig. 1 is not very
expressive, as the giant connected component visually
resembles an unstructured, but highly connected “blob”
of nodes. Therefore, we considered a closer look at the
link properties, having a certain assumption in mind: As
we observe only person articles, we can expect their
mutual links to bear strong temporal characteristics,
i.e., the timespan that a link covers ranges from zero
years up to millenniums for persons referencing people
from, e.g., ancient Greece. We approximate the
interval for each person-person link by the difference
between their birth dates as provided by the ULAN. Fig.
3 shows the distribution of link timespans on a linear
scale for all the 54,590 Wiki links compared to all the
9,406 ULAN links for the same subset. What comes to
attention first is that the number of ''short'' links
covering 0-37.5 years clearly outnumbers the sum of
all the other (Dbpedia: 75.3%, ULAN: 81.2%). This can
be interpreted as such that contemporaries are much
more likely to be interlinked than persons who are
generations apart. We measure the link distance in
both possible directions: into the future and into the
past. This reveals an interesting difference between the
associative links from ULAN and those from Wikipedia:
While ULAN links are highly reciprocal, i.e., for almost
every link from person A to B, there is also one from B
to A, this is not the case in Dbpedia, where twice as
many links point into the past (66.8%) than into the
future (33.2%). Of the 54,590 Wiki links, 15,606
(28.59%) are mirrored and the remaining 38,984 are
''unique''. From a historical viewpoint, this seems quite
logical, as artists can always draw from influences from
previous generations, which are asymmetric by nature.
From an encyclopedic viewpoint, however, this is not
the case: While the Getty editorial guidelines for the
inclusion of associative relationships in ULAN7 (chapter
3.5.2.3.1) clearly state that each link has to be
mirrored, this is not the case with the Wikipedia link
style guidelines8. Nevertheless, the observed amount of
mirrored Wiki links still differs significantly from [2],
reporting about 8.7% reciprocal links for a much larger
cross-domain subset of Wikipedia. This, however, might
likely be related to our domain-specific set of person
articles. Another difference between ULAN and Dbpedia
links is that the latter do generally cover much larger
timespans than the former - see Fig. 4 with link
distances plotted against a log scale. This might also be
explained through the ULAN editorial guidelines
(chapter 3.5.1.4), stating that ''Relationships should be
made only between records that are directly related''.
Refining the Visualization
The temporal properties of the Dbpedia links indicate
that we might see more structure in the visualization by
omitting links that extend beyond a certain time
interval. As we use a force-based algorithm for the
layout, we expect the node link structure to “unfold“
into an approximately chronological sequence. Fig. 5
shows the result of applying the same layout algorithm
as for Fig. 1, but without links whose endpoints are
more than 75 years apart. We chose this distance, as
89.13% of all Dbpedia links are “shorter” (ULAN:
98.87%) and it is roughly suited to separate
contemporary from inter-generational relationships.
The resulting layout now has a chronological alignment,
representing the flow of art history from the
Renaissance (left) until today (right), solely based on
the relations between its numerous historical actors.
7 http://www.getty.edu/research/tools/vocabularies/guidelines
8 http://en.wikipedia.org/wiki/Wikipedia:Manual_of_Style/Linking
0
5000
10000
15000
20000
25000
30000
-4800 < -2400
-2400 < -1200
-1200 < -600
-600 < -300
-300 < -150
-150 < -75
-75 < -38
-38 < 0
0 < 38
38 < 75
75 < 150
150 < 300
300 < 600
600 < 1200
1200 < 2400
2400 < 4800
Dbpedia
Ulan
1
10
100
1000
10000
100000
-4800 < -2400
-2400 < -1200
-1200 < -600
-600 < -300
-300 < -150
-150 < -75
-75 < -38
-38 < 0
0 < 38
38 < 75
75 < 150
150 < 300
300 < 600
600 < 1200
1200 < 2400
2400 < 4800
Dbpedia
Ulan
Figure 3: Distribution of link
distances (linear scale)
Figure 4: Distribution of link
distances (logarithmic scale)
5
Figure 5: Visualization of Wiki links between art-historical actors with max. distance 75 years. A high-res version is available at http://www.wikiarthistory.info.
6
Coloring the nodes according to each person’s
nationality, as provided by the ULAN, reveals
interesting patterns of cultural interaction within the
network, as they are perceived by the English speaking
Wikipedia community: The left side of Fig. 5 is
dominated by Italians (green). This cluster spans
Renaissance and Baroque times, fading out by the end
of the 17th century. A small cluster on the lower left
represents German Renaissance around Albrecht
Duerer (black). Further right, a cluster of Dutch
Renaissance and Baroque (orange) persons marks the
beginning of an interesting development. Two distinct
groups of British (cyan) and French (red) historical
actors dominate the center, until Americans (blue)
unite them into a big melting pot around the late 19th
century. The rightmost part represents Post-Modernist
Americans, with a nationality-independent cluster of
Architects beneath.
Discussion & Future Work
By visualizing the reduced link set, we were able to
identify structural characteristics of the inter-article
links. There exist link based article clusters that
represent the nationality of the respective persons and
possess strong temporal characteristics that represent
art historical periods and schools. A thorough
discussion of these observations is, however, left for
future work. As we have limited our observation to
English Dbpedia and used a mainly Anglo-American
authority file for the identification of articles, we expect
to obtain different results for using authorities of
different origin and for other language versions of
Dbpedia. Comparing the results could reveal differences
in cultural perception between the different Wikipedia
communities. Furthermore, it would be interesting to
apply the described method to other domains.
Conclusion
We have effectively integrated heterogeneous data
sources for the analysis of 54,590 links between 18,002
domain specific Wikipedia person articles. Our findings
include indications that properties such as node degree
distribution are comparable to entire Wikipedia, while
others like reciprocity of links seem different. The
observed links tend to cover timespans comparable to a
human lifetime, and are more likely to point into the
past than vice-versa. By removing links covering more
than 75 years, we visually identified article clusters
with interesting cultural and temporal properties.
Acknowledgements
This work was funded by the FWF Austrian Science
Fund, Project No. L602, The Virtual 3D Social
Experience Museum
References
[1] P. D. Clough, N. Ireson, and J. Marlow. Extending
domain-specific resources to enable semantic access to
cultural heritage data. JoDI 10, 6 (2009).
[2] J. Kamps and M. Koolen. Is Wikipedia link structure
different? In Proc. WSDM, ACM (2009), 232-241.
[3] D. Milne, O. Medelyan, and I. H. Witten. Mining
domain-specific thesauri from Wikipedia: A case study.
In Proc. WI, IEEE CS (2006), 442-448.
[4] J. Oomen and L. Aroyo. Crowdsourcing in the
cultural heritage domain: opportunities and challenges.
In Proc. C&T, ACM (2011), 138-149.
[5] A. Popescu and G. Grefenstette. Spatiotemporal
mapping of Wikipedia concepts. In Proc. JCDL, ACM
(2010), 129-138.
[6] J. Voss. Measuring Wikipedia. In Proc. ISSI, (2005).
