Describing Linked Datasets
On the Design and Usage of voiD, the “Vocabulary Of Interlinked Datasets”
∗ †
Keith Alexander Richard Cyganiak
Talis Ltd. DERI, National University of
Ireland, Galway
‡ §
Michael Hausenblas Jun Zhao
DERI, National University of Department of Zoology,
Ireland, Galway University of Oxford
ABSTRACT linked data on the Web using bubbles for datasets and arcs
In this paper we discuss the design and implementation of between bubbles for the links between these datasets. The
voiD, the “Vocabulary Of Interlinked Datasets”, a vocabu- voiD vocabulary, the “Vocabulary Of Interlinked Datasets”,
lary that allows to formally describe linked RDF datasets. allows one to describe datasets (the bubbles) and linksets
We report on use cases for voiD, the current state of the (the arcs between the bubbles), and in turn enables a num-
specification and its potential applications in the context of ber of tasks to be automated in a scalable manner.
linked datasets. The remainder of this paper is as follows: in Section 2 we
discuss some use cases that provide the motivation for voiD.
Then, in Section 3 we report on the design of the voiD core
1. INTRODUCTION vocabulary and its usage along with other vocabularies such
With the growth of the number of linked datasets [12], as Dublin Core [25]. We describe the publication and the
automating certain tasks, such as discovery, selection and consumption of voiD descriptions in Section 4. In Section 5
optimisation, becomes more and more important. Now, one we discuss current and potential applications of voiD and
might argue that URIs and RDF [17] are all one needs to ex- report on related work in Section 6. We discuss future plans
plore the linked datasets; follow-your-nose1 , however bears and conclude in Section 7.
some inherent problems. The possible links that can be fol-
lowed from a starting URI raises both performance and trust 2. USE CASES
issues. The main reason for these issues lies in the granu-
larity level of the available descriptions. Additionally, the In the following we will describe our motivation use cases
dynamics of the data-sources [13] also has an impact on the for voiD. In the context of linked data, we basically differ-
performance of, say, a crawl over a collection of datasets; entiate between:
and the reliability of secondary data resources [26]. • on the one hand linked data publisher (a person or
In the early days of linked data (2006 and 2007) [5] the organisation exposing structured data as RDF on the
main focus of the community was on publishing data and Web and interlink it with other datasets), and
finding good practices. Now, in the second phase, other
issues such as usability, quality, performance, reliability of • linked data consumer on the other hand; these might
the infrastructure and the data in the linked data ecosystem be machines, for example using a semantic indexer or
are increasingly recognized to be important. a query engine or, as well, humans, e.g., when using a
How can we overcome the limitations of follow-your-nose Web of Data browser such as Tabulator [4].
while retaining the self-descriptive momentum and being
able to exploit available tools, methodologies, etc.? A sim- It is worth noting that the following use cases are not
ple yet effective approach is to decrease granularity. Rather necessarily restricted to the linked data domain.
than talking about single resources, we talk about something
which up to now only existed in drawings, such as in the
2.1 Efficient Discovery of Datasets
LOD cloud2 , which graphically represents the landscape of
2.1.1 Dataset Publisher
∗keith.alexander@talis.com A dataset publisher might not be identical with the party
†richard.cyganiak@cyganiak.de who created the raw datasets, but one who publishes them
‡michael.hausenblas@deri.org onto the Web in a more accessible format. A dataset pub-
§jun.zhao@zoo.ox.ac.uk lisher wants to be able to publish metadata about the dataset
1
such that:
http://esw.w3.org/topic/FollowLinksForMoreInformation
2
http://www4.wiwiss.fu-berlin.de/bizer/pub/ • The dataset can be found and aggregated by search
lod-datasets_2008-09-18.html engine applications, or discovered in relevant searches;
Copyright is held by the author/owner(s).
LDOW2009, April 20, 2009, Madrid, Spain. • The metadata provides clear licensing information so
. that consumers can know how they can use the data
and to whom they should attribute credits for creat- 2.3 Effective Dataset Selection
ing/publishing the dataset; A consumer may have discovered several datasets, for ex-
ample as a result of an indexer query. The question then
• That consumers can obtain information about access
arises how to select appropriate datasets from this list of
interfaces, such as APIs and SPARQL endpoints.
potential candidates. The consumer, either a human or a
It is in the best interest of a dataset publisher to provide query federation engine, might wish to define “appropriate-
potential users of the data with information that supports ness” along the following criteria:
them in accessing and using the dataset.
1. The content of the dataset, that is, what the dataset
2.1.2 Search Engine Provider mainly is about. Based on some kind of categorisation
scheme a selection could take place;
A search engine provider wants to discover detailed de-
scriptions about datasets efficiently. A crawler has stum- 2. The interlinking to other datasets, that is, to which
bled upon an individual RDF document on the Web. How other datasets and how the dataset is interlinked;
will it discover metadata that applies to the entire dataset
and cannot be repeated in every single document? The sim- 3. Vocabularies used in the dataset.
ple approach of just putting the voiD description online and
linking it from somewhere on the site does not meet our The criteria listed above can be understood in terms of
needs, as it could take the crawler a long time to find the quality and quantity. For example, one might be interested
description. It is important that the voiD description is only in datasets containing foaf:interest links to a certain
discoverable as soon as the crawler finds the first RDF doc- other dataset. Or, where the number of links are of inter-
ument, so that the crawler can use voiD metadata to guide est, one may specify that only datasets with more than one
its processing of the data. million links should be taken into account.
The Sindice search engine [23] already uses Semantic Site- 2.4 Query Optimisation
maps [7] to enable discovery and efficient processing of data-
sets. It seems natural to address the situations above by With many datasets now on the Web both connectable
building on Semantic Sitemaps. (through shared vocabulary terms) and connecting (by link-
ing to resources in other datasets), it is naturally desirable
2.2 Expressing Research Data to query across multiple datasets at once with SPARQL.
A developer working together with biologists wants to help Optimisation of SPARQL queries can be achieved in a
domain experts to find research data published by their peer static way. A set of logical rules [22] can be applied to a
colleagues. These are often produced for a particular exper- query engine, to calculate all equivalent query plans for a
iment, for a particular study or publication, or hosted by given query and then choose the most optimised query plan
a particular public database. Scientists might know whose to be executed. To optimise SPARQL queries dynamically,
datasets they would prefer to access because they often have i.e., deciding the best execution approach during the exe-
a clear idea about their content or they trust more on that cution phase [14], one can use the statistical information
data provider. When looking for new datasets, they might about datasets, such as how much information is provided
search for datasets that provide relevant content (such as in- about a particular entity or property. This information can
formation about genes, proteins, or micro-array gene expres- be used by the query mediator to optimise query plans by,
sion), that are produced in a right experiment environment, for example, modifying the order in which a query pattern is
or that provide additional information that will complement executed according to the estimated size of data results [8].
their local experiment results.
To find the right dataset and to make this dataset accessi- 3. VOCABULARY DESIGN
ble for biologists in a user-facing application, the developer The Vocabulary of Interlinked Datasets (voiD) [2] is a vo-
often has to go through the following process: cabulary and a set of instructions that enables the discovery
and usage of linked datasets. The principle of the voiD ef-
• Locate a dataset that contains information relevant
fort is to use real requirements to guide the scope of the
to biologists’ research interests, such as information
design, and to re-use existing vocabularies wherever possi-
about a specific organism; or more specifically, ge-
ble instead of creating our own. Therefore, we have kept
nomic information about a particular organism;
the creation of new classes and properties under the voiD
• Find out how this dataset can be programmatically ac- namespace (http://rdfs.org/ns/void#) to the minimum.
cessed, as an RDF dump, through SPARQL endpoint
or any other protocol;
3.1 Datasets
In the following, we will define and explain the basic con-
• Find out the licence associated with the dataset, mak- cepts voiD operates with. A fundamental entity in voiD is
ing sure that data are accessible under open-access li- a dataset.
cence or certain attribution;
Definition 1. A dataset is a set of RDF triples that are
• Understand the content of the dataset in order to per- published, maintained or aggregated by a single provider.
form an alignment with other datasets. Information
about URIs used in the dataset can help one with the We think of a dataset as a meaningful collection of triples,
data identity alignment, schema(s) used in the dataset that deal with a certain topic, originate from a certain source
for data schema alignment, and its links with other or process, are hosted on a certain server, or are aggregated
datasets for assisting data integration. by a certain custodian. The term thus has a social dimension
that is not easy to capture in a formal definition. This dif- The conceptual model of voiD links is depicted in Fig. 1.
ferentiates datasets from RDF graphs [17], which are purely Let us assume there are two datasets. One of them contains
mathematical constructs. Any arbitrary set of RDF triples links to the other, that is, it contains RDF triples that con-
is an RDF graph, by definition, regardless of the triples’ se- nect resources from both datasets. We model this in voiD
mantics. Also, typically a dataset is accessible on the Web, using two instances of void:Dataset, and another dataset
for example through resolvable HTTP URIs or through a :LS1 which is a subset of one of the datasets, and declared
SPARQL endpoint, and it contains sufficiently many triples to be of type void:Linkset. We define void:Linkset as:
that there is benefit in providing a concise summary.
The ultimate purpose of creating a void:Dataset instance Definition 3. A linkset LS is a set of RDF triples where
is that this single resource represents the entire RDF dataset, for all triples ti = hsi , pi , oi i ∈ LS, the subject is in one
and thus allows us to make statements about the entire dataset, i.e. all si are described in DSsrc , and the object is
dataset within the standard RDF model. The relationship in another dataset, i.e. all oi are described in DSsink .
between a void:Dataset instance and the concrete triples The natural expectation is that both DSsrc and DSsink
contained in the dataset is established only in an operational are themselves described in voiD. We note that the triples
manner: A voiD description usually contains access informa- ti are often referred to as “interlinking triples”.
tion, such as the address of a SPARQL endpoint where the
triples can be accessed. 3.2.1 Inline links vs. 3rd-party links
We find that most datasets describe a well-defined set of In voiD we are able to model two different situations: the
resources. Hence, a dataset can also be seen as a set of de- classic LOD3 case vs. the 3rd-party case (Fig. 2). In the
scriptions for certain resources, which often share a common classic LOD case, the linkset is a subset of one of the two
URI prefix (such as http://dbpedia.org/resource/). involved datasets, while in the 3rd-party case a third dataset
HTTP URIs have “owners”, due to their use of DNS do- is involved that actually contains the linkset.
main names. URI ownership is defined as “a relation be- Though the 3rd-party cases is not yet widely implemented
tween a URI and a social entity, such as a person, organisa- in the context of linked data, this pattern of keeping links
tion, or specification.”[15] Information about a URI that is separate from interlinked datasets has been well argued in
provided by the URI owner is called authoritative informa- existing research such as found in the Hypertext commu-
tion. We use this notion to define authoritative datasets: nity [6]. In LOD, there are already first applications (RKB
explorer, see section 5.1), and it is very likely that such sys-
Definition 2. A dataset is authoritative with respect to tems will evolve over time and grow considerably.
a certain URI namespace if it contains information about
resources named by URIs in this namespace, and is published 3.2.2 Interlinking Regarding Directionality
by the URI owner. Independent of the former situation, voiD distinguishes
between the non-directed vs. directed cases. In some
A straightforward method of publishing authoritative data- cases one is interested in stating the direction of the in-
sets is by using resolvable HTTP URIs in the linked data terlinking (for example with foaf:interest), and in other
style. The URI owner also configures the server that re- situations the direction is of no interest (e.g., owl:sameAs),
sponds when the URI is resolved. Therefore, if resolving as shown in Fig. 3.
yields a description of the resource named by the URI, then In order to express the interlinking as outlined above, voiD
the data is authoritative. offers the following RDF properties:
The notion of authoritative information supports the so-
cial convention that a URI owner gets to decide what a URI • void:subset to state where the interlinking triples re-
identifies. Providing authoritative information is how the side (read: a dataset :DS has a subset :LS);
URI owner communicates this decision to the world. Even
if third parties disagree with that information, they can still • void:target to declare an interlinking target (for the
agree that they are talking about the same thing, which non-directed case); in the directed case, one can use
would be much harder without the grounding provided by void:subjectsTarget and void:objectsTarget to de-
the existence of an authoritative source. termine the direction (both being sub-properties of
void:target);
3.2 Linksets • void:linkPredicate to express the RDF property (type)
Besides datasets, voiD also deals with interlinking be- of the interlinking in a linkset.
tween datasets. Interlinking in voiD is a first-class citizens,
hence modelled as a class. We note that it is expected that per RDF predicate a re-
spective instance of void:Linkset could be created, depend-
ing on the needs of an application. Further, one may take
into consideration that due to the modelling of void:target
and its sub-properties, light-weight subsumption inferencing
may be necessary to apply generic queries that will not dis-
tinguish between the directed and non-directed case.
In listing 1 a sample voiD description is depicted describ-
ing the interlinking from DBpedia to DBLP. It is an example
3
LOD ... Linking Open Data datasets, see http://esw.
w3.org/topic/SweoIG/TaskForces/CommunityProjects/
Figure 1: Interlinking modelling in voiD. LinkingOpenData
(a) Non-directed case.
(a) Classic LOD case: Describing the :DBpedia
dataset and its contained :DBpedia2DBLP linkset
(b) Directed case.
Figure 3: Interlinking regarding direction.
• foaf:homepage of the dataset’s homepage should be
used, to allow one to connect different descriptions of
the same dataset provided in different places on the
Web. The recommended process in voiD is IFP smush-
(b) 3rd-party case: Describing the stand-alone ing on foaf:homepage property;
:DBpedia2DBLP linkset
• dcterms:subject should be used to categorise a dataset.
Figure 2: Interlinking regarding authoritative datasets. For the general case, we recommend the use of a DB-
pedia resource URI (http://dbpedia.org/resource/
XXX) to categorise a dataset, where XXX stands for
the thing which best describes the main topic of what
for a directed case. The description defines nothing about
the dataset is about. However, DBpedia might not
who published this voiD description about DBpedia, which
contain concepts for describing some domain specific
means that it could also be an example for a 3rd-party case.
datasets. For example, there are no exact DBpedia re-
Further, the listing 2 shows a SPARQL query that is exe-
source URIs for describing that a dataset is about “in
cuted against listing 1 to search for a dataset that is about
situ hybridisation image”. We hence encourage pub-
“computer science” and which is linked from DBpedia. The
lishers to describe such datasets using concepts widely
result yields the dataset :DBLP.
adopted in their own communities, so that they can
not only capture precisely the categorisation of their
3.3 Reuse of Other Vocabularies datasets but also ensure that these datasets could be
In the voiD guide [1] we describe the reuse of other vocab- connected with other relevant data from their domains;
ulary terms not directly defined in the core voiD vocabulary
alongside with voiD. Some important properties from other • Statistical information represented using the “Statis-
vocabularies are listed in the following. We note that there tical Core Vocabulary” (SCOVO)4 [11].
are many other aspects one may want to choose to describe
in a dataset. A complete description of recommended usage 3.4 Dataset Licensing
can be found from the voiD user guide [1].
As stated in Section 2, it is crucial for a data publisher to
associate appropriate licensing information with their pub-
• Properties from the dcterms namespace for general lished data, so that potential users of the dataset would
metadata, such as the publishing organization and pub-
4
lishing date of a dataset; http://purl.org/NET/scovo
1 @prefix owl : < http :// www . w3 . org / 2 0 0 2 / 0 7 / owl # > . SCOVO for representing statistics. The main class in SCOVO
2 @prefix foaf : < http :// xmlns . com / foaf /0.1/ > . is the scovo:Item, which records a single number or statis-
3 @prefix dc : < http :// purl . org / dc / terms / > .
4 @prefix void : < http :// rdfs . org / ns / void # > .
tical value along with so called dimensions. We provide two
5 @prefix dbp : < http :// dbpedia . org / r e s o u r c e / > . types of information for describing statistics:
6
7 : DBpedia a void : Dataset ; • Statistics concerning the whole dataset or linkset, such
8 foaf : homepage < http :// dbpedia . org / > ;
9 void : subset : DBpedia2DBLP . as overall triple count or fine-grained statistics, ex-
10 pressing the number of instances of a class or prop-
11 : DBLP a void : Dataset ; erty by using different pre-defined dimensions, includ-
12 foaf : homepage < http :// dblp . l3s . de / d2r / > ;
13 dc : subject dbp : C o m p u t e r _ s c ie n c e ;
ing void:numberOfResources, etc.;
14 dc : subject dbp : Journal ;
15 dc : subject dbp : Proceedings . • Attributing statistics to a source, recording where a
16
17 : DBpedia2DBLP a void : Linkset ; statistical datum stems from.
18 void : subjectsT arget : DBpedia ;
19 void : objectsTarget : DBLP ; Listing 4 demonstrates possible statistic information one can
20 void : linkPredicate owl : sameAs .
publish for their dataset. The current modelling of statistics
Listing 1: An exemplary voiD description. 1 : DBpedia a void : Dataset ;
2 void : statItem [
3 rdf : value 20000;
1 SELECT DISTINCT ? dataset 4 scovo : dimension void : n u m b e r O f R e s o u r c es ;
2 WHERE { 5 scovo : dimension foaf : Person ;
3 ? dataset a void : Dataset ; 6 dcterms : source < http :// wiki . dbpedia . org / > ;
4 dcterms : subject dbp : C o m p u t e r _s c i e n c e . 7 ] .
5 ? linkset void : s ubjectsT arget : DBpedia ;
6 void : objectsTarget ? dataset .
7 } Listing 4: Expressing statistics about a dataset in
voiD.
Listing 2: An exemplary query on voiD.
in voiD is still experimental. We had to make choices be-
tween (i) a precise usage of scovo through a rather verbose
know under which terms they can use it and what attri- expression and (ii) creations of shortcuts to express statistics
bution they should apply. The dcterms:license property needed for describing linked datasets:
should be used to to point to the license under which a
dataset has been published. Further, to allow automatic • Scovo has an implicit assumption that all scovo:Items
analysis of datasets, voiD also recommends a set of canoni- associated with the dataset they describe share the
cal identifiers for well-known licenses [1]. The example be- same dimensions. This does not fit well with our re-
low states that the DBpedia dataset is published under the quirements for being able to mix items of different di-
terms of the GNU Free Documentation License. mensions for a dataset. On the other hand, the correct
Scovo modelling would lead to awkwardly complex and
verbose notation for simple statistics.
1 : DBpedia a void : Dataset ;
2 dcterms : license
< http :// www . gnu . org / c o p y l e f t / fdl . html > . • We encourage the use of classes and properties in places
where scovo requires an instance of scovo:Dimension.
This breaks the symmetry of the scovo model. scovo
Listing 3: An exemplary voiD description about
would require us to create a scovo:Dimension for each
data license.
class or property. This would be quite verbose.
Licensing of datasets is a complex issue. Datasets are col- Because of the issues above, queries for statistics informa-
lections of facts rather than creative works, and different tion using SPARQL can be awkward. It will often require
laws apply. Scientists are most cautious about publishing a verbose check to make sure that an item has only certain
their datasets onto the Web and they might request very dimensions and no others.
specific or strict policies for sharing their data. Most li-
censes such as Creative Commons or the GPL are based on 3.6 Additional Terms in voiD
copyright and are designed to protect creative works, but RDF datasets use one or more RDF-Schema vocabularies
not databases, and applying them to datasets might not or OWL ontologies, hence we provide the void:vocabulary
have the desired legal result. Meanwhile, efforts such as to list vocabularies used in a dataset. To express technical
Open Data Commons [19] and Science Commons [16] are features of a dataset, such as formats in which the data is
developing dedicated licenses for data. available, one can use void:feature. Further, a SPARQL
endpoint that provides access to a dataset via the SPARQL
3.5 Statistics protocol can be announced using the void:sparqlEndpoint
Of special interest to distributed SPARQL agents will property. Listing 5 shows the usage of the terms described
be the statistics about the triples available in the dataset, above. We note that a complete list of the terms is available
described with the void:statItem predicate. We adopt from the voiD user guide [1].
1 : DBpedia void : sparqlEn dpoint
< http :// dbpedia . org / sparql > ;
2 void : feature [ dcterms : format
" application / rdf + xml " ; ] ;
3 void : vocabulary
< http :// xmlns . com / foaf /0.1/ > .
Listing 5: Additional voiD terms usage.
4. PUBLICATION AND CONSUMPTION
We envision dataset publisher to offer a voiD description
along with their dataset. A voiD description typically has
two parts, (i) manually created part (categorisation, vocab-
ulary, license, etc.), and (ii) automatically generated part,
mainly regarding statistics.
Figure 4: The voiD discovery-via-sitemaps process.
In the following we will discuss the publication process of
voiD descriptions and their discovery in order to consume
them. 5. VOID IN THE WILD
4.1 Publication After releasing the first edition of voiD in early 2009, we
have seen a certain community uptake. People and organi-
Publishing a voiD file means to physically deploy it on
sations would start using in different areas and for different
the Web in an RDF serialisation. We have detailed out the
purposes, potentially far beyond what we have envisioned in
options in the voiD guide [1].
the realm of our own use cases. We report on known usages
For dataset that are published as a collection RDF docu-
of voiD in the following and point out potential application
ments, as commonly seen in the linked data publishing style,
areas.
one can use a dcterms:isPartOf triple in each document to
link back to the URI identifying the voiD dataset, as shown 5.1 Existing Applications
in listing 6. Resolving the dataset URI will answer a voiD
descriptions about the entire dataset, allowing agents to dis- 5.1.1 Tools for Creating voiD Descriptions
cover the voiD description when encountering an individual
To boot-strap the process of creating voiD descriptions,
document from the collection. The intuition behind using
several tools are available: ve, the voiD editor (Fig. 5),
the dcterms:isPartOf property is that the RDF document
liftSSM5 , an XSLT script able to boot-strap from a Semantic
contains an RDF graph whose triples are part of the dataset.
Sitemap, and, for creating the quantitative, statistical data,
a new release of the NX parser6 , offering a voiD export for
1 < http :// dbpedia . org / data / Berlin > dcterms : isPartOf statistics.
: DBpedia .
5.1.2 “Linked Datasets Explorer” (LDE)
Listing 6: Use backlinks publish voiD description of To let user browse and explore a collection of voiD de-
a dataset. scriptions, we have developed the LDE demonstrator. Fig. 6
shows the current state of LDE7 which operates on a man-
ually created, so called “seed” set of voiD descriptions.
As discussed in [10], we can imagine that voiD descriptions
are crawled and indexed by semantic search engines (such 5.1.3 RKB explorer
as Sindice [23] or Yahoo’s search monkey [18]) in order to
The RKB explorer has a voiD site8 which enables querying
provide a central point of lookup.
and browsing for CRS datasets. Further, the interlinking of
4.2 Discovery via Sitemaps the RKB sites can be visualised using the underlying voiD
descriptions (Fig. 7).
A discovery mechanism for use by RDF-harvesting web
crawlers (Fig. 4) has been defined as follows: 5.1.4 Query Federation
1. Given a domain name, the client gets the file robots.txt Only recently Clarck-Parsia announced their voiD sup-
and searches for a line that starts with Sitemap:; the port9 :
rest of that line is the URI of a sitemap; “There is a touch point with the linked data ef-
2. The semantic sitemaps extension to the sitemap pro- fort, which meant that the new voiD vocabulary
tocol defines a element that can have a 5
http://rdfs.org/ns/void-guide#sec_4_3_Publishing_
child element. If present, the value tools
6
of that element is a URI that identifies the dataset http://sw.deri.org/2006/08/nxparser/release/
datasetURI; nxparser-1.1.jar
7
http://ld2sd.deri.org/lde/
8
3. The dataset URI datasetURI is dereferenced which http://void.rkbexplorer.com/
9
yields the voiD description of the dataset. http://clarkparsia.com/weblog/2009/02/04/
distributed-query-pellet-into-the-void/
for describing datasets turns out to be very useful about a resource. The discovery protocol for XRDS doc-
for describing the distributed data sources that uments given a URI was defined in 2006 as part of Yadis,
we query over, including their interrelations.” focusing on services such as OpenID and OAuth. In early
2008, XRDS-Simple was proposed12 , but is now obsolete.
Further, OpenLink plans to release its “Smart SPARQL Only very recently, the latest draft of “/host-meta” [21]
Federation capabilities”, based on voiD, soon. was proposed. The core of this proposal is a single “well-
5.1.5 Middleware known location”, /host-meta, acting as a directory of the
interesting metadata about a Web site. The format allows
OpenLink’s Sponger Middleware uses voiD for generat- different types of site metadata to be referenced by an URI
ing linked data from non-RDF data sources such as HTML or included inline.
pages. An example10 from http://linkeddata.uriburner. One could understand the “HTTP Link: header” [20] pro-
com/ with the voiD description deployed as XHTML+RDFa posal related to voiD, as it also supports discovery, offering
is shown in Fig. 8. Further, the statistics maintenance in metadata about resources by resurrecting a (currently dep-
their Virtuoso Quad Store is performed based on voiD. recated) feature of HTTP. This proposal is at the time of
5.2 Potential Applications writing still under vivid discussion and not yet seen stable.
Regarding federated SPARQL queries, DARQ (Distributed
We envision voiD to be applied in many scenarios, some ARQ) [22] proposes so called “service descriptions” that
of which we have identified earlier in section 2. Only re- are able to specify capabilities of a SPARQL endpoints.
cently, for example, we have started to develop a dataset The service descriptions enable the DARQ query engine
ranking algorithm based on voiD descriptions; this is subject to decompose a query into sub-queries, each of which can
to more research. One could further apply voiD to DARQ be answered by an individual service using query rewrit-
(Distributed ARQ) [22]. ing and cost-based query optimisation to speed-up query
A totally different application domain is visualisation: for execution. Further, we note an attempt called “SPARQL
example, “The Map of Data”11 in Sindice can be generated Endpoint Description”13 that aimed to allow the announce-
automatically thanks to voiD. ment of endpoint capabilities and contents, support discov-
Ultimately, to be of use, one wants applications that ben- ery through service directories, and supply browsing and
efit from voiD. Put in other words, this means that, given federation hints. Both proposals seem to be not further
there are applications that consume voiD and offer some maintained and/or have not reached wide-spread adoption.
added value, the incentive for publishers to provide voiD Finally, we note that the W3C Technical Architecture
descriptions is self-evident. One such application could be Group (TAG) started to contemplate about “Uniform Ac-
a sort of dynamic dataset selector which, configured with a cess to Metadata”14 , basically being a survey regarding the
specification of the dataset (topics, license, interlinking with problem of specifying a uniform method for obtaining infor-
certain other datasets) would at run-time of an application mation pertaining to a resource without necessarily having
discover and select appropriate datasets according to the to parse a representation of the resource.
search specification.
6. RELATED WORK 7. OUTLOOK
To the best of our knowledge, no comparable approach We have released the voiD vocabulary and voiD user guide
to voiD exists. That is, in the context of the Web of Data, to linked data communities in January this year. In this re-
we are not aware of any specification that allows the de- lease, we have used the use cases presented in section 2 to
scription of datasets and their interlinking the way voiD guide the design scope of the voiD vocabulary. Supports
does. However, we acknowledge previous work of Semantic for describing the quality, provenance and versions of linked
Sitemaps [7] and build upon it. datasets are to be addressed in the next release of voiD.
In the scope of the Web of Documents, we note that at the Also, the statistics modelling in the current voiD model is
time of writing a W3C Working Draft of POWDER (Proto- still experimental. We are communicating with user com-
col for Web Description Resources) [3] is available. POW- munities and the SCOVO team in order to propose a more
DER aims at providing information about Web resources, stable modelling in the coming release15 . Additionally we
such as scope, authoritative information, etc., without re- will liaison with initiatives such as the “Ontology Metadata
trieving the resources themselves. POWDER comes in two Vocabulary” [9] sharing similar goals.
flavours, (i) as human-legible XML, and (ii) in an RDF ver- To test and evaluate the usefulness of voiD, we need tools
sion. It also provides a GRDDL transformation to turn the that use voiD to support the discovery of datasets or the
former into the latter. The descriptions can be applied to SPARQL query federation. Fortunately, semantic query en-
groups of resources defined via listing of URIs, regular ex- gines like Sindice and SPARQL query processing systems
pressions, etc. Several publishing methods are suggested (like OpenLink) are adopting voiD in their implementations.
(via HTML in the header, HTTP Link: header or It is challenging to completely automate the creation of voiD
using XHTML+RDFa). Especially in the Web of Trust, descriptions. We need tools like the NX parser to take as
POWDER is expected to play a vital role, though imple-
12
mentation complexity might hinder wide-spread adoption. http://www.hueniverse.com/hueniverse/2008/03/
Further, OASIS’s XRDS (eXtensible Resource Descriptor putting-xrds-si.html
13
Sequence) [24] is an XML format for metadata discovery http://esw.w3.org/topic/SparqlEndpointDescription
14
http://www.w3.org/2001/tag/doc/
10
http://linkeddata.uriburner.com/about/html/http: uniform-access-20090205.html
//twitter.com/mhausenblas#Dataset 15
See http://code.google.com/p/void-impl/issues/
11
http://sindice.com/map list?can=2&q=milestone:Release2.0forplannedissues.
much as possible of the heavy lifting for non-technical data [11] M. Hausenblas, W. Halb, Y. Raimond, L. Feigenbaum,
publishers as possible. and D. Ayers. SCOVO: Using statistics on the Web of
data. In 6th European Semantic Web Conference
(ESWC2009), Semantic Web in Use Track, 2009.
Acknowledgements [12] M. Hausenblas, W. Halb, Y. Raimond, and T. Heath.
Our work has partly been supported by the European Com- What is the size of the Semantic Web. In Proceedings
mission under Grant No. 217031, FP7/ICT-2007.1.2, project of I-Semantics 2008, Graz, Austria, 2008.
“Domain Driven Design and Mashup Oriented Development [13] M. Hausenblas, W. Slany, and D. Ayers. A
based on Open Source Java Metaframework for Pragmatic, performance and scalability metric for virtual RDF
Reliable and Secure Web Development” (Romulus)16 , and graphs. In 3rd Workshop on Scripting for the
the Joint Information Systems Committee [Project “Fly- Semantic Web (SFSW07), Innsbruck, Austria, 2007.
Web”]. The authors would further like to thank (alphabet- [14] HP Lab. TDB/Optimizer.
ically): Orri Erling, Hugh Glaser, Olaf Hartig, Tom Heath, http://jena.hpl.hp.com/wiki/TDB/Optimizer, 25
Andreas Langegger, Ian Millard, Marc-Alexandre Nolin, Yves October, 2008. Accessed in March 2009.
Raimond, Yrjänä Rankka, Francois Scharffe, and Giovanni
[15] I. Jacobs and N. Walsh. Architecture of the World
Tummarello.
Wide Web, Volume One. W3C Recommendation 15
December 2004, W3C Technical Architecture Group
8. REFERENCES (TAG), 2004.
[1] K. Alexander, R. Cyganiak, M. Hausenblas, and [16] J. Klump, R. Bertelmann, J. Brase, M. Diepenbroek,
J. Zhao. voiD guide—Using the Vocabulary of H. Grobe, H. Höck, M. Lautenschlager, U. Schindler,
Interlinked Datasets. Community Draft, voiD working I. Sens, and J. Wächter. Data publication in the open
group, 2009. http://rdfs.org/ns/void-guide/. access initiative. Data Science Journal, 5:79–83, 2006.
[2] K. Alexander, R. Cyganiak, M. Hausenblas, and [17] G. Klyne, J. J. Carroll, and B. McBride. RDF/XML
J. Zhao. voiD, the “Vocabulary of Interlinked Syntax Specification (Revised). W3C
Datasets”. Community Draft, voiD working group, Recommendation, RDF Core Working Group, 2004.
2009. http://rdfs.org/ns/void/. [18] P. Mika. Microsearch: An interface for semantic
[3] P. Archer, K. Smith, and A. Perego. Protocol for Web search. In Semantic Search, International Workshop
Description Resources (POWDER): Description located at the 5th European Semamntic Web
Resources. W3C Working Draft 14 November 2008, Conference (ESWC 2008), volume 334 of CEUR
POWDER Working Group, 2008. Workshop Proceedings, pages 79–88. CEUR-WS.org,
2008.
[4] T. Berners-Lee, Y. Chen, L. Chilton, D. Connolly,
R. Dhanaraj, J. Hollenbach, A. Lerer, and D. Sheets. [19] P. Miller, R. Styles, and T. Heath. Open data
Tabulator: Exploring and analyzing linked data on commons, a license for open data. In Proceedings of
the Semantic Web. In In Proceedings of the 3rd the Workshop on Linked Data on the Web
International Semantic Web User Interaction (WWW2008), 2008.
Workshop (SWUI06), Athens, Georgia, USA, 2006. [20] M. Nottingham. Link relations and HTTP header
[5] C. Bizer, T. Heath, K. Idehen, and T. Berners-Lee. linking. Internet-Draft, 1 December 2008, IETF
Linked Data on the Web (LDOW2008). In Linked Network Working Group, 2008.
Data on the Web Workshop(WWW2008), 2008. [21] M. Nottingham and E. Hammer-Lahav. Host
[6] L. A. Carr, D. C. DeRoure, W. Hall, and G. J. Hill. metadata for the Web. Internet-Draft, 10 February
The Distributed Link Service: A tool for publishers, 2009, IETF Network Working Group, 2009.
authors and readers). In Proceedings of the 4th [22] B. Quilitz and U. Leser. Querying distributed RDF
International World Wide Web Conference: The Web data sources with SPARQL. In Proceedings of the 5th
Revolution), pages 647–656, Boston, USA, 1995. European Semantic Web Conference 2008), pages
[7] R. Cyganiak, H. Stenzhorn, R. Delbru, S. Decker, and 524–538. Springer, 2008.
G. Tummarello. Semantic Sitemaps: Efficient and [23] G. Tummarello, R. Delbru, and E. Oren. Sindice. com:
flexible access to datasets on the Semantic Web. In Weaving the open linked data. Proceedings of the 6th
Proceedings of the 5th European Semantic Web International Semantic Web Conference 2007
Conference, volume 5021, pages 690–704, Tenerife, (ISWC2007), 4825:552–565, 2007.
Spain, 2008. [24] G. Wachob, D. Reed, L. Chasen, W. Tan, and
[8] O. Hartig and R. Heese. The SPARQL query graph S. Churchill. Extensible Resource Identifier (XRI)
model for query optimization. In Proceedings of the Resolution Version 2.0. Committee Draft 03 28
4th European Semantic Web Conference 2007, pages February 2008, OASIS eXtensible Resource Identifier
564–578, Innsbruck, Austria, 2007. (XRI) TC, 2008.
[9] J. Hartmann, Y. Sure, P. Haase, R. Palma, and M. del [25] S. Weibel, A. S. for Information Science, and
Carmen Suárez-Figueroa. OMV – Ontology Metadata Technology. The Dublin Core: A simple content
Vocabulary. In C. Welty, editor, ISWC 2005 - In description model for electronic resources. Bulletin of
Ontology Patterns for the Semantic Web, 2005. the American Society for Information Science and
[10] M. Hausenblas. Discovery and usage of linked datasets Technology, 24(1):9–11, 1997.
on the Web of data. In Talis NodMag 4, 2008. [26] J. Zhao, A. Miles, G. Klyne, and D. Shotton. Linked
data and provenance in biological data webs. Briefings
16
http://www.ict-romulus.eu/ in Bioinformatics, 2008.
Figure 5: Manual creation of voiD descriptions with ve.
(a) Looking-up datasets. (b) Browsing datasets.
Figure 6: Linked Datasets Explorer (LDE).
Figure 7: Visualisation of RKB interlinking.
Figure 8: OpenLink’s instant-voiD-generator for structured HTML pages.