D-Lib Magazine
January/February 2015
Volume 21, Number 1/2
Table of Contents
A Methodology for Citing Linked Open Data Subsets
Gianmaria Silvello
University of Padua, Italy
silvello@dei.unipd.it
DOI: 10.1045/january2015-silvello
Printer-friendly Version
Abstract
In this paper we discuss the problem of data citation with a specific focus on Linked Open Data. We outline the main requirements a data citation methodology must fulfill: (i) uniquely identify the cited objects; (ii) provide descriptive metadata; (iii) enable variable granularity citations; and (iv) produce both human- and machine-readable references. We propose a methodology based on named graphs and RDF quad semantics that allows us to create citation meta-graphs respecting the outlined requirements. We also present a compelling use case based on search engines experimental evaluation data and possible applications of the citation methodology.
1 Introduction
One of the most relevant socio-economical and scientific changes in recent years has been the recognition of data as a valuable asset. The Economist magazine recently wrote that "data is the new raw material of business" and the European Commission stated that data-related "technology and services are expected to grow from EUR 2.4 billion in 2010 to EUR 12.7 billion in 2015" [H2020 WP, 20142015]. The principal driver of this evolution is the Web of Data, the size of which is estimated to have exceeded 100 billion facts (i.e. semantically connected entities). The actual paradigm realizing the Web of Data is the Linked Open Data (LOD), which by exploiting Web technologies, such as the Resource Framework Description (RDF), allows public data in machine-readable formats to be opened up ready for consumption and re-use. LOD is becoming the de-facto standard for data publishing, accessing and sharing because it allows for flexible manipulation, enrichment and discovery of data in addition to overcoming interoperability issues.
Nevertheless, LOD publishing is just the first step for revealing the ground-breaking potential of this approach residing in the semantic connections between data enabling new knowledge creation and discovery possibilities. Current efforts for disclosing this potential are being concentrated on the design of new methodologies for creating meaningful and possibly unexpected semantic links between data and for managing the knowledge created through these connections. This endeavor is shifting LOD from a publishing paradigm to a knowledge creation and sharing one.
Borgman in [Borgman, 2012b] outlined four rationales for sharing data that we think are gaining even more traction as the LOD paradigm extends its reach; Borgman pointed out that sharing and citing data is important for: (i) reproducing or verifying research, (ii) making results of publicly funded research available to the public; (iii) enabling others to ask new questions of extant data; and (iv) advancing the state of research and innovation. These rationales are to a varying extent rooted in the LOD paradigm, which makes data sharing a priority; we believe that along with data sharing, also data citation should be considered a prime concern of the research community. Indeed, together with data sharing, data citation is fundamental for giving credit to data creators and curators (attribution), to reference data in order to identify, discover and retrieve them [Borgman, 2012a] and for building and propagating knowledge [Buneman, 2006; Buneman and Silvello, 2010; Lawrence, et al., 2011].
In the context of LOD a dedicated methodology for citing a dataset or a data subset has not yet been defined or proposed. Recently, two EU projects i.e. PRELIDA and DIACHRON considered these aspects from the permanent preservation point-of-view [Auer, et al., 2012], but there are as yet no concrete solutions we can employ for data citation of LOD subsets.
In this paper we build on the newly defined "RDF Quad Semantics" [Klyne, et al., 2014] to pinpoint a methodology for automatically generating citations of LOD subsets, which are machine-readable, but at the same time are understandable to a human. This methodology allows for citing LOD subsets with variable granularity (i.e. we can cite a single entity, a single statement, a subset of statements and the whole dataset) and produce citations composed by a unique identifier (i.e. a reference) used to retrieve the cited data subset in a human- and machine-readable format and some human- and machine-readable descriptive metadata assessing the citation (i.e. its quality and currency) and enabling data attribution [Borgman, 2012a]. A further property of the methodology being proposed here is that it is defined within the boundaries of the LOD paradigm and related widely accepted and used technologies; this means that if a given organization already has in place an infrastructure for creating and exposing LOD on the Web, the very same infrastructure can be exploited as is for data citation purposes.
The rest of the paper is organized as follows: in Section 2 we report on the LOD paradigm and RDF model highlighting the role of named graphs and quad semantics; furthermore, we outline the main requirements that a data citation methodology must fulfill and discuss some existing data citation systems. In Section 3 we present a use case based on search engine experimental data discussing why a data citation methodology for LOD is required. In Section 4 we describe the data citation methodology for LOD and in Section 5 we relate it to the presented use case reporting some possible applications. In Section 6 we draw some final remarks.
2 Background
2.1 Linked Open Data and RDF
The LOD paradigm [Heath and Bizer, 2011] refers to a set of best practices for publishing data on the Web 1 and it is based on a standardized data model, the Resource Description Framework (RDF). RDF is designed to represent information in a minimally constraining way and it is based on the following building blocks: graph data model, IRI-based vocabulary2, data types, literals, and several serialization syntaxes.
The basic structural construct of RDF is a triple (subject, property, and object), which can be represented in a graph; the nodes of this graph are subjects and objects and the arcs are properties. IRIs identify nodes and arcs. RDF adopts a property-centric approach allowing anyone to extend the description of existing resources; properties represent relationships between resources, but they may also be thought of as attributes of resources, like traditional attribute-value pairs. RDF graphs are defined as mathematical sets; adding or removing triples from an RDF graph yields a different RDF graph.
RDF 1.1 [Klyne, et al., 2014] specifications introduced the concept of RDF dataset, which is a collection of RDF graphs composed of: (i) a default RDF graph which may be empty and (ii) a set of named graphs which is a pair consisting of an IRI (i.e. the name of the graph) and an RDF graph.
The semantics as well as the formal definition of such named graphs are still debated by the research community, but a consensus about a limited number of options has been reached as described by Zimmermann in [Zimmermann, 2014]; in the following we consider two definitions: named graph and quad semantics.
Named graph: The graph name denotes an RDF graph or a particular occurrence of that graph. An example of named graphs is:
ex:g1 { | ex:a | ex:b | ex:c |
| ex:d | ex:e | ex:f } |
In the example above there is an RDF graph named "ex:g1" composed of two triples.
Quad semantics: The named graph is considered as a set of quadruples where the first three elements are subject, property and object as usual and the fourth is the name of the graph as shown in the example below where "ex:x" is the name of the graph:
ex:a | rdf:type | ex:c | ex:x | . |
ex:c | rdfs:subClassOf | ex:d | ex:x | . |
In general the fourth element can also be used as a statement identifier, a model identifier, or to refer to the "context" of a statement. In the literature, the fourth element has been used to denote a time frame in [Gutiérrez, et al., 2007], to deal with uncertainty in [Straccia, 2009] and to handle provenance in [Carroll, et al., 2005]. In all these cases the fourth element is used with a semantics tailored to the specific need of the application under exam.
In the following we use the fourth element as a triple identifier in order to label statements and use them for building citation graphs.
2.2 Requirements and Existing Systems for Data Citation
The "Joint Declaration of Data Citation Principles" produced by the Data Citation Synthesis Group outlined the main principles of data citation. Leveraging on the insights and considerations outlined by [Altman and Crosas, 2013] and [Ball and Duke, 2012] we point-out four main requirements a data citation methodology must fulfill:
- provide a description of the data in order to give scholarly credit and normative and legal attribution to data creators and curators. Description metadata that would make up a complete citation are open to debate, but there is a certain agreement about the minimum required set, which must contain: author, title, date and location (i.e. a persistent reference to the cited data);
- uniquely identify the cited object and the associated metadata;
- enable variable granularity data citation (i.e. to cite a dataset as a whole, a single unit or a subset of data);
- produce references that are both human- and machine-readable.
Many of the existing approaches to data citation allow us to reference datasets as a single unit having textual data serving as metadata source. As pointed out by [Proll and Rauber, 2013] most data citations "can often not be generated automatically and they are often not machine interpretable."
The rule-based citation system proposed by [Buneman and Silvello, 2010] meets the desired features for data citation because it allows for citing data with variable granularity, creates both human- and machine-readable citations and associates description metadata with the cited data. On the other hand, this system works under the assumption that data are hierarchically structured (e.g. XML files) and thus it cannot be straightforwardly adopted in the context of LOD where we deal with RDF graphs.
[Proll and Rauber, 2013] proposed an approach based on assigning persistent identifiers to time-stamped queries, which are executed against time-stamped and versioned relational databases. While this system also meets the data citation requirements, it is defined for working with relational databases and there is no extension to RDF graphs.
[Groth, et al., 2010] proposed the nano-publication model where a single statement (expressed as an RDF triple) is made citable in its own right; the idea is to enrich a statement via annotations adding context information such as time, authority and provenance. The statement becomes a publication itself carrying all the information to be understood, validated and re-used. The model proposed by Groth et alii is close to the RDF reification process [Klyne, et al., 2014] where we can make claims about an RDF statement; in the nano-publication model a URI is assigned to the statement in order to make it a dereferenceable entity to be used in the RDF graph enriching it. Then, a name is associated to the RDF graph making it citable.
This model is not specifically defined for citing RDF sub-graphs with variable granularity, but it is centered around a single statement and the possibility of enriching it. Nevertheless, in the following we extend and improve this very idea in order to cite RDF graphs satisfying the four requirements outlined above.
In this context, it is interesting to mention the "Research Objects" initiative which has the aim of bringing together several international activities with the common goal of defining a new approach to publications in order to improve reuse and reproducibility of research. LOD plays a central role in this context and several activities comprised by the Research Objects initiative are based on LOD-related methodologies and technologies e.g. the Open Archive Initiative for Object Re-Use and Exchange (OAI-ORE), which exploits the RDF Framework for sharing compound objects on the Web. This initiative does not propose a methodology for citing LOD subsets, but it could exploit it within the research objects it defines; from this perspective the methodology proposed here is a companion of research objects rather than an alternative to them.
3 Use case: Search Engines Experimental Evaluation
We present a use case based on experimental evaluation of search engines, which produces scientific data that are highly valuable from both a research and financial point of view [Rowe, et al., 2010]. Experimental evaluation of search engines is a demanding activity that benefits from shared infrastructures and datasets that favor the adoption of common resources, allow for replication of the experiments, and foster comparison among state-of-the-art approaches. Therefore, experimental evaluation is carried out in large-scale evaluation campaigns at an international level, such as the Conference and Labs of the Evaluation Forum (CLEF) in Europe and the Text Retrieval Evaluation Conference (TREC) in the USA. The evaluation activities produce huge amounts of scientific and experimental data, which are the foundation for all the subsequent scientific production and development of new systems. For this reason, these data need to be discoverable, understandable and citable [Harman, 2011].
As a consequence, the Distributed Information Retrieval Evaluation Campaign Tool (DIRECT) system [Agosti, et al., 2012] has been defined with the aim of modeling the experimental data and developing a software infrastructure able to manage and curate them. The data made available by means of DIRECT have been mapped in RDF with the purpose of exposing them as LOD on the Web in the near future. This will increase the discoverability and the re-use of the experimental data; furthermore, it will enable a seamless integration of datasets produced by different international campaigns as well as the standardization of terms and concepts used to label data across research groups [Ferro and Silvello, 2014].
In Figure 1 we report a portion of the RDF graph and its triple representation showing a sample of experimental data3 shareable by means of DIRECT. In this case we show two sample systems (system A and system B) which produce two experiments (exp A and exp B) submitted to an evaluation campaign (CLEF 2009); by considering a certain evaluation measure (precision) each experiment achieved a certain value (0.70 for system A and 0.46 for system B). Furthermore, the evaluation campaign is associated with a descriptive statistic indicating that the average precision of all the considered systems is 0.53.
|