In this paper we describe our experience with building the riese dataset, an interlinked, RDF-based version of the Eurostat data, containing statistical data about the European Union. The riese dataset (http://riese.joanneum.at), aims at serving roughly 3 billion RDF triples, along with millions of high-quality interlinks. Our contribution is twofold: Firstly, we suggest using RDFa as the main deployment mechanism, hence serving both humans and machines to e ectively and e ciently explore and use the dataset. Secondly, we introduce a new way of enriching the dataset with high-quality links: the User Contributed Interlinking, a Wiki-style way of adding semantic links to data pages.
2.
Statistical data on the (Semantic) Web. Looking at
related work reveals that there is actual demand for new
solutions to disseminate statistical data using semantic
technologies. As reported by Assini [
RDFa. As RDFa [
This section provides a short description of the Eurostat data, which served as the primary input for the riese dataset. Eurostat provides detailed statistics for the entire European Union as well as additional statistics for major nonEuropean countries. The Eurostat data is arranged along the following themes:
Three main data sources are being provided by Eurostat for public download5, namely (i) the statistical data itself, (ii) a table of content, and (iii) dictionaries.
The statistical data is provided as dump download of approximately 4,000 single tab-separated values (TSV) documents, having a total size of approximately 5GByte, and containing some 350 million data values. This data is updated twice a day. Only limited semantic exploitable information is contained in these TSV documents, hence it is inevitable to use other available information sources. A table of content (TOC) provides a hierarchical overview of the datasets|organised in so called themes|allowing to identify the structure and content of a dataset. 3http://www.biostat.harvard.edu/~carey/hbsfin.html 4http://rdfa.info/2007/08/23/ joost-using-rdfa-on-website/ 5http://europa.eu/estatref/download/everybody/ Dictionaries are especially valuable as they contain all information for resolving the nearly 100,000 data codes used in the statistical data. These data codes refer to dimensions such as time, location, currency, etc. The data codes also contain an implicit hierarchy, which can be used for further classi cation. However, various schemas have been used requiring individual processing for extracting classifying features. For example, in order to refer to locations, the Nomenclature of Territorial Units for Statistics (NUTS)6 is in use. This basically allows to extract information about the structure of administrative divisions of countries. For each of the dictionaries a di erent terminology is used. Most of the data is represented in time series with varying granularity, ranging from annual to daily data. Each single data item can be identi ed using the corresponding dataset and various dimensions as the following example illustrates: The population of the European Union can be seen as one single data item valued at 497,198,740 (contained in the dataset 'Total population'), having as time-dimension the year 2008, as indicator-dimension 'Population on 1. January', and as geo-dimension the 'European Union (27 countries)'. Additionally the data is agged as provisional and Eurostat estimate.
In a rst phase we have analysed the Eurostat data. We have identi ed the implicit semantics present in the TOC and the dictionaries and gathered a number of issues. Firstly, the Eurostat data set is highly heterogeneous; the data sources formats vary (TSV, HTML) and are not machine-processable per se. Another issue is the modelling of temporal data, more speci c how to represent time intervals. Further, the schemas in the dictionaries form a multidimensional space that somehow has to be linearised in order to be represented in a URI format. We have also identi ed data provenance (and trust) issues, which are currently only handled on a global level.
Based on the analyses given above we state the following requirements for a linked dataset that is designed to serve both humans and machines:
The system must serve both humans and machines in an adequate way by applying the don't-repeat-yourself (DRY)7 principle; To allow both humans and machines to reveal more information, the follow-your-nose8 principle must be applied.
To be a useful (real-world) Semantic Web application, the system must be able to scale to the size of the Web; Additionally we want to point out that we aim at providing high-quality interlinking. Hence, the sheer template-driven generation of global interlinks is certainly not su cient. 6http://ec.europa.eu/comm/eurostat/ramon/nuts/ 7http://skimstone.x-port.net/node/272 8http://www.inkdroid.org/journal/2008/01/04/ following-your-nose-to-the-web-of-data/
In order to demonstrate how to address the issues raised earlier in this paper, we have implemented the riese dataset (http://riese.joanneum.at) as a Semantic Web application. This section describes how the mapping|from the available, relational data into RDF form|has been done, explains the interlinking mechanisms applied, and nally introduces the riese system architecture.
This section explains the schemas utilised in riese and discusses the mapping to RDF.
The data used in riese is a snapshot of the data available for bulk-download taken on 9 Jan 2008. Depending on the type of data, three formats are used by Eurostat: HTML or plain text for the TOC, and TSV for the dictionary les and the actual data tables.
In Fig. 1 the riese core schema is depicted. Currently the
riese core schema is modelled using RDF-Schema [
1 data : e b 0 4 0 _ i n f l _ 2 0 0 6 _ a t a : Item ; 2 dc : title " Inflation rate Austria
2006 " ; 3 rdf : value " 1.7 " ; 4 : dimension dim : geo_at ; 5 : dimension dim : time_2006 ; 6 : dataset data : eb040 .
Additionally, the following schemas are used or have been extended:
Simple Knowledge Organisation Systems (SKOS) [
the event ontology [
Finally, it is possible to formulate very exible queries.
Other approaches, such as the U.S. Census data [
1 SELECT * 2 WHERE 3 { ? item riese : dimension dim : geo_at . 4 ? item riese : dataset ? dataset . 5 ? dataset dc : title ? ds_title 6 FILTER regex ( ? ds_title , " food " ,i ) }
The example in listing 2 demonstrates this. All items for Austria are returned that belong to a dataset with 'food' in the description9.
Leaving the mapping of the Eurostat data into RDF apart, it is equally important to apply the follow-your-nose principle, hence creating interlinks to other datatsets. For creating interlinks in riese we have basically used the following approach: 1. Restrict the source dataset to possible candidates for interlinking to the target dataset; 2. For each qualifying item in the source dataset look up the label or another identifying feature in the target dataset; 3. Restrict the results by appropriate classi cations or identi ers;
For example the interlinking between country descriptions in riese and Geonames is done using the ISO-3166 alpha2 country codes (AT) instead of the label (Austria) assuring that exactly the same resource is addressed in both datasets. Queries can be constructed with very little a-priori knowledge about the structure of the dataset; 9with default namespace http://riese.joanneum.at/ schema/core#
Note that using ISO-3166 codes for identifying country
descriptions in di erent datasets was already used by Voss [
In the practical implementation this means that rst of all the source dataset is restricted to only geographical features. According to the nomenclature used it is also possible to identify country descriptions in the source dataset. Then the Geonames search Webservice (i.e. the target dataset) is queried using the standardized codes. The result from the target dataset is then further restricted to return only countries, i.e. entries having a speci ed Geonames feature code (A.ADM1). Finally all the matches are being interlinked by inserting a new triple into the source dataset which relates the resources using owl:sameAs. In this case it is possible to create exact matching high-quality interlinks.
Further candidates for interlinking the Eurostat data are Geonames (more geographical features), DBpedia, CIA Factbook and Wikicompany. By introducing these interlinks users of riese will not only bene t from a larger interlinked dataspace but especially for the geographic features also by being able to produce even more exible and powerful queries.
As already mentioned above, the pure pattern-based
approach is believed to be not su cient for high-quality
interlinks. This is why we additionally allow users to add
their own links, a new feature called 'User Contributed
Interlinking' (UCI). The idea behind is applying the WikiWiki
approach to LOD: Users can add semantic links to other
datasets on their own. Currently three di erent types are
supported: rdfs:seeAlso, owl:sameAs and foaf:topic (cf.
also [
Based on the lessons learned from [
The riese Web application supports the following tasks:
For creating the RDF representation from the original Eurostat les, SWI-Prolog scripts are used. The SWI-Prolog Semantic Web Library provides an infrastructure for reading, querying and storing semantic web documents. Additionally the Prolog-2-RDF (p2r) modules11 and individually de ned mappings are used for translating the input data to RDF. The resulting RDF can be accessed via a SPARQL endpoint and it is further possible to consume a dump of the entire data. We have created one large dump containing all triples, 10http://sw.deri.org/2007/07/sitemapextension/ 11http://moustaki.org/p2r/ The latter approach is currently used for 'Rendering & Serving' where the PHP scripts looks up the les in the le system and renders a RDFa representation. Beside the data that originates from Eurostat (the o cial statistical data), the UCI module stores the user-contributed triples in a separate document. This physical separation is mainly due to being able to replace parts of the data without too much additional e ort.
In the following we show how riese can both satisfy the human user, as well as the machine (Semantic Web agents). Please note that the alpha version of the riese system is available at http://riese.joanneum.at/.
Both human and machine users would presumably start at the top-level page in order to get an overview of the available data. In Fig. 3 the hierarchical rendering of a selected Eurostat theme (the 'Economy' theme) is depicted. A machine accessing the same page would have another view, namely focusing on the embedded RDF, exemplary shown in example 3.
Note that although both humans and machines access the same resource, di erent parts are relevant. This is made possible through the deployment in XHTML+RDFa. The Listing 3: The Eurostat theme 'Economy' viewed by a machine. browser will render a nice GUI, the machine gets what it deserves: triples.
Further, a single table may be explored; this is depicted in Fig. 4.
However, till now the user was passively consuming the information. But riese o ers more: Users can provide their own links using the UCI (cf. Fig. 5).
The UCI module enables the user to add (and remove for
that matter) additional links to a certain data page. As
the user must specify the type (cf. the drop-down box in
Fig. 5) it is ensured that only valid triples are introduced
to the system|the subject of the RDF statement is always
the page where the 'Related' box is on; the predicate is
determined through the type selection. The object (named
target in our context) is the only variable we are not able to
control. However, we rely on the community e ect, i.e. we
expect that 'wrong' links will be removed. A REST-based
interface for adding UCI-triples automatically is available
as well. Regarding the acceptance of the UCI, i.e. enabling
users to contribute semantically typed links, we refer to the
success story of Wikipedia [
In this paper we have presented the riese dataset containing statistical data from Eurostat. We have shown how to RDFise and interlink this data, hence making it possible to expose it onto the Semantic Web. The bene ts of supplying data for both humans and machines have been explicated and a WikiWiki approach for adding user contributed interlinks has been introduced.
We have also identi ed some issues and bottlenecks when
deploying datasets of such enormous size. Generating a static
le-structure with small RDF les requires quite a lot of
time. This is due to our current way of storing the data
items in the le system. Because in riese several hundred
millions of folders and les have to be created, the bottleneck
is somehow obvious. Moreover, when accessing datasets
(tables) containing thousands of items (cells) in individual les
this yields thousands of le access operations for simply
parsing them. Regarding the le system we came across another
limitation: reserved names on the MS Windows operating
systems (as it turned out, it is not possible to create les or
folders named 'con', 'aux', etc. [
When modelling the representation of time related to a certain statistical information we encountered some challenges as the raw data from Eurostat is sometimes ambiguous and can only be resolved by analysing the corresponding document. For example the statement time\2007 can stand for the value over a period of time (e.g. entire year) or at the end of the reporting period (e.g. 31 Dec). In our future work we will focus on resolving these issues.
The future work roughly comprises a thorough analysis of the current bottlenecks, as well as gathering feedback from end-users of the system. We are planning to use a solution based on an triple-store (such as SESAME or Virtuoso) allowing us to generate triples at a faster pace|currently it would take us several weeks to RDFise the entire Eurostat data set. Using a dedicated store will likely improve the performance serving the data to both human and machine users.
Finally, as Eurostat updates their data twice a day, we aim at updating the data on riese continuously. One of the issues to be solved in this respect is how to deprecate the data when updating the items. From a UI point-of-view we also want to address navigational issues (using maps and timelines12) to further enhance the user experience. 12http://simile.mit.edu/timeline/
The research leading to this paper was carried out in the \Understanding Advertising" (UAd) project13, funded by the Austrian FIT-IT Programme, and was partially supported by the European Commission under contract FP6027026-K-SPACE.
The authors would like to thank the Linking Open Data community and the RDFa folks. Additionally we would like to credit all the people that made available the following magni cent technologies: SWI-Prolog, Apache, PHP, RAP - Rdf API and YUI. We wish to thank Danny Ayers for his early comments on modelling issues, Giovanni Tummarello for feeding sindice (allowing advanced queries), and Jan Wielemaker for his superb SWI-Prolog support.