LCSH, SKOS and Linked Data

A technique for converting Library of Congress Subject Headings MARCXML to Simple Knowledge Organization System (SKOS) RDF is described. Strengths of the SKOS vocabulary are highlighted, as well as possible points for extension, and the integration of other semantic web vocabularies such as Dublin Core. An application for making the vocabulary available as linked-data on the Web is also described.


Introduction
Simple Knowledge Organization System (SKOS) -a vocabulary and data model for expressing Knowledge Organization Systems (KOSs) such as thesauri and classification schemes for referencing and re-use in Semantic Web applications -was developed by successive projects and working groups from the late 1990s through its publication in August 2009 as a World Wide Web Consortium (W3C) Recommendation. 1 This paper describes the work of the W3C Semantic Web Deployment Working Group, which was chartered in 2006 to carry SKOS Core, a W3C Working Draft, through the rigorous review required by the W3C Recommendation Track process. 2 The final results of that process are recorded in the formal specification for SKOS [14].This paper, in contrast, focuses on the process itself.By reconstructing the discussion around alternative features and design options and presenting the rationale for key decisions, the paper aims at providing insight into how SKOS turned out as it did, and why.Assuming that SKOS, like any other successful technology, may eventually be subject to revision and improvement, the critical account offered here may help future editors approach such a task with deeper understanding.
After presenting a brief history of SKOS from 1997 through 2009, the paper outlines the rationale for a language, other than existing formal ontology languages, for expressing Knowledge Organization Systems.Drawing a contrast between logically precise conceptual structures and more intuitive, pragmatic knowledge representations, the section describes the principle of "minimal ontological commitment" that guided the design of SKOS.
The middle section of the paper walks through the components of the SKOS model -SKOS Concepts (and how they differ from formal-ontological Classes), Concept Schemes, Semantic Relations between concepts, Lexical Labels, Documentation Properties, and Collections of concepts.The section considers several proposed features of SKOS deemed by the working group to be out of scope.
While SKOS was developed for expressing KOSs as "concept schemes" -sets of interrelated concepts -without modeling those concepts as formal "classes," the data model for SKOS itself is defined as an ontology, i.e., as a set of formal properties and classes expressed using the W3C Web Ontology Language (OWL).The final section of the paper reviews the semantics of SKOS properties and classes as defined by axioms supporting inference and "integrity conditions" for when data can be considered "not consistent" with the SKOS data model.The section also considers the compatibility of SKOS with different variants of OWL and with previous versions of SKOS itself.The paper concludes by highlighting issues that could provide starting points for a future revision of the specification.

History of SKOS from the late 1990s through 2009
Today's SKOS can be traced back to work on improving search interfaces in the European project Desire (1997-2000).The original W3C Resource Description Framework (RDF) Model and Syntax Working Group (1997-1999), aware of Desire, raised the question of expressing thesauri in RDF as an issue. 3Phil Cross, Dan Brickley, and Traugott Koch turned the Desire results into a proposal, published jointly by the Institute for Learning and Research Technology (ILRT) in the UK and the Lund University Library Netlab in Sweden, "for encoding a core set of thesaurus relationships using an RDF schema." 4 This draft schema was picked up by the European project LIMBER (Language Independent Metadata Browsing of European Resources, 1999-2001), which defined a vocabulary based more explicitly on "concepts" labeled by terms in multiple languages. 5he results of the LIMBER Project fed into the SWAD Europe project (Semantic Web Advanced Development, 2001-2004). 6In SWAD Europe, Alistair Miles of Rutherford Labs solicited input from experts on thesaurus and classification standards, creating a community of interested users, for whom W3C set up a community mailing list, public-esw-thes, 7 and the revised vocabulary was published under the name "Simple Knowledge Organization System."This draft was picked up in 2004 by the W3C Semantic Web Best Practice and Deployment Working Group (2004-2006), whose Porting Thesauri Task Force 8 created a home page for what was now called "SKOS Core."In 2005, the working group published "SKOS Core Vocabulary Specification" as a W3C Working Draft. 9SKOS Core was taken as a starting point for the review process described in this paper.
The Semantic Web Deployment Working Group began by distilling requirements for SKOS out of use cases solicited from early adopters about present and future applications [7].Successive revisions of the 2005 SKOS Core specification were posted for public comment as Working Drafts, then as Candidate and Proposed Recommendations, prior to finalization as a W3C Recommendation.The editors of the specification were supported by two working-group chairs with active input from a dozen or two working-group members and a wider circle of external reviewers and mailing-list followers.Discussion took place on the working group's mailing list 10 and on public-esw-thes for the wider community. 11The group met over a period of 35 months in three face-to-face meetings and 110 near-weekly teleconferences.Teleconferences used W3C's bot-supported telephone bridge, which assigned URIs to actions "scribed" into a shared chat channel and automatically generated draft minutes, complete with pointers to the agenda, previous minutes, actions past and current, mailing-list postings, and document drafts.As technical or design issues were formally raised they were assigned URIs and added to an Issue Tracker 12 that automatically collected links to any minutes or postings in which the issues were mentioned.Each such URL cited in this paper leads the interested reader into a web of richly interlinked working-group resources.
The following discussion will make reference to online resources produced during the WG process.Rather than peppering the narrative text with URIs, references to issues, requirements and axioms will be handled as follows: • Issues.Details of all issues are documented in the Working Group's issue tracker at http://www.w3.org/2006/07/SWD/track/issues.Issues will be cited in the text by number, e.g., Issue 27.
• Requirements.Requirements are documented in the SKOS Use Cases and Requirement document at http://www.w3.org/TR/skos-ucr.Requirements will be referred to in the text by their handles, e.g., R-GroupingInConceptHierarchies.
• Axioms.SKOS axioms are listed in Tables 2 and 3. Details of these axioms are given in the SKOS Reference document at http://www.w3.org/TR/ skos-reference.Axioms will be referred to in the text by their "S" handle, e.g., S1.
For all of the above, full URI references will be available in digital versions of the paper.

Rationale for SKOS
Many institutions develop and maintain Knowledge Organization Systems (KOSs) -thesauri, classification systems, subject heading lists, folksonomies, and the like, holding concepts and terminologies for a wide range of domains -as backbone structures for their information systems.The potential of such KOSs to serve as components in knowledge-rich applications has been recognized since the rise of the Web in the 1990s.
Porting an existing KOS for use in Semantic Web applications, however, is not a trivial problem.The Semantic Web languages for expressing domain knowledge are mathematically formal in nature.The vocabulary description language of the Resource Description Framework (RDFS) and the Web Ontology Language (OWL), in particular, provide ways to define classes and properties and to associate those classes and properties with formal reasoning rules that enforce constraints or produce new knowledge by inference.KOSs, on the other hand, have typically been designed not as formally precise representations of domain knowledge, but as informal structures reflecting the intuitive knowledge of human users in a form useful for resource discovery (e.g., through supporting query expansion).KOSs have variously been classified as "term-based" or "concept-based" depending on how explicitly they are intended to represent conceptual structures. 13Traditional KOS standards have never included the sort of formal axioms expressed by Semantic Web ontology languages.
Informally defined KOSs cannot typically be translated into the language of RDFS and OWL properties and classes, with their formal-logical implications, without introducing potentially false or misleading logical precision.Informal KOSs may be converted into formal ontologies (see [6]), but the process of assigning appropriate formal semantics to the elements of a KOS may require a long, hard modeling effort.Hierarchical relationships, for example, must be disambiguated into relationships of class instantiation, class subsumption, part-whole, or other types -a process that cannot usually be automated.An analysis of the thesaurus of the National Cancer Institute [4] (as reported in [3]), for example, found conceptual structures that are incompatible with formalized frameworks that assume stricter modeling principles.The AGROVOC thesaurus of multilingual agricultural terminology, the product of many people over many years working from multiple perspectives, was straightforwardly converted into a hierarchy of OWL classes many years before the finalization of SKOS.While the maintainers of AGROVOC-in-OWL intended to increase its ontological precision over time, through editorial correction and refinement, it eventually proved to be more practical simply to convert AGROVOC back into the formally less "committed" form of a SKOS concept scheme, leaving it to designers of specific implementations to upgrade parts of the thesauri into class-based ontologies when required to support reasoning [2].
The traditional use cases for which KOSs were typically designed are still relevant in the Web context.One key role of a controlled vocabulary, for example, is to improve precision when retrieving objects from an indexed collection.The hierarchical and associative relationships of thesauri enable users to browse for search terms, and information retrieval applications can use this structure to automatically expand queries, which improves recall.Applications such as simple search or browsing of documents or "conceptual spaces" can all benefit from a shared basis for data exchange and linking.For such purposes, Semantic Web technology is indeed a game changer, as it allows users and developers to seamlessly re-use data from different contexts, or to link together multiple KOSs, in order to achieve broader or deeper search, even across languages.Expressing KOSs as Linked Data allows the library community to create pools of trusted URIs citable by catalogers in resource descriptions in support of such applications [1, 19].
SKOS aims at providing a path for migrating KOSs to a Semantic Web context at low cost by expressing features common to a wide range of KOS types.
The SKOS properties for "broader," "narrower," and "related," for example, are intended to capture the native, sometimes ambiguous semantics of existing thesauri and similar structured vocabularies.Using SKOS, no additional intellectual work is required to represent these relationships in RDF, allowing the maintainers of controlled structured vocabularies to leverage their existing investments.
The design of SKOS followed the principle of making a minimal ontological commitment to the nature of concepts and of relationships between concepts.As explained by Thomas Gruber [5]: An ontology should require the minimal ontological commitment sufficient to support the intended knowledge sharing activities.An ontology should make as few claims as possible about the world being modeled, allowing the parties committed to the ontology freedom to specialize and instantiate the ontology as needed.
The principle of avoiding over-commitment guided many of the discussions about possible extensions to SKOS.Where the use cases collected by the working group demonstrated no clear requirement for a candidate feature, or in the absence of clear usage experience, the group tended to opt for a "safe" course of action.As a result, SKOS captures the basic, informal semantics most commonly required by the use cases.Where there was doubt that a particular feature would be easy to understand or use, the working group generally chose to omit the feature from the specification.
The working group was particularly focused on keeping SKOS compatible with the thesaurus standards ISO 2788 and ISO 5964, 14 with the result that the SKOS data model reflects standard thesaurus construction principles.SKOS does not, however, express all of the best practices described in the ISO standards, nor does it include the elements needed to capture all of the features of any given, existing KOS standard, such as specializations of broader and narrower hierarchical relations (see Section 3.3).Experience indeed shows that best practices are not always followed-a problem revealed, for example, when generic "See also" references in the Library of Congress Subject Headings were converted into standardized thesaurus relations [18]-and that some KOSs use idiosyncratic constructs for meeting very specific requirements.The working group felt that fully committing SKOS to supporting the creation and validation of any particular type of concept scheme, such as a standard thesaurus, would create an obstacle to the wide-spread adoption of SKOS by users of other types.
Lightly specified by design, SKOS is intended to prevent data publishers from introducing false precision into their data and to prevent inference engines from drawing unwarranted conclusions.In some cases, however, the specification recommends usage conventions, such as best practices for KOS design.The SKOS model thus presents two layers of specification: formal, enforceable axioms, along with weaker "guidelines."Guidelines are not represented formally, nor they are considered to be inviolable integrity constraints; rather, they are considered to be advisory.
Opting for such a minimal approach is made dramatically easier by the vocabulary extension mechanisms offered natively by Semantic Web technology.Applications that require more constrained behaviour may define compatible extensions to SKOS [8].For example, modelers may coin sub-classes and sub-properties of SKOS properties or associate those properties with specific formal axioms.The RDF data model allows properties from such extension vocabularies to be used alongside properties from SKOS in expressing data.Where properties seen as required were already provided elsewhere, such as the Dublin Core property dc:subject, the working group deferred to existing vocabularies.

Components of SKOS
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
The SKOS data model enables features listed above-identifying, labeling, documenting, linking, and mapping concepts, and aggregating concepts into concept schemes or collections-by defining the elements depicted in Figure 1.This section looks at the design choices made in modeling those components.

SKOS Concepts (and how they differ from OWL Classes)
A wide diversity of concepts.. SKOS is designed to express, in an interoperable way, different types of Knowledge Organization System-sets of terms or concepts, whether listed with definitions (glossaries), in hierarchical structures (basic classifications or taxonomies), or characterized by more complex semantic relations (thesauri, subject heading lists, or other advanced structures).Each type of KOS has its own specific characteristics.Yet they all organize knowledge by gathering a coherent set of lexical entities (terms, words, headings, captions. . . ) around more abstract notions that the SKOS model represents as Concepts.In a thesaurus, for example, a concept is the construct that clusters a preferred term (the one used for describing resources in a document retrieval system) with near-synonymous alternative terms (or variants).A KOS may link such concepts among themselves with various types of semantic relations, such as class-subclass, part-whole, or looser associative links.
SKOS leaves ample room for interpreting the notion of concept, and many artifacts from information science and other fields fall in scope.As the SKOS Reference puts it [14], "a SKOS concept can be viewed as an idea or notion; a unit of thought.However, what constitutes a unit of thought is subjective, and this definition is meant to be suggestive, rather than restrictive." If the objective is information retrieval via the use of a knowledge organisation system as a subject indexing language, then one can take an operational view and define concepts as units of indexing and retrieval [21].The subject indexing process can then be viewed as the action of linking documents (such as a textbook about butterflies) to concepts (such as a concept labeled "butterflies"), and the retrieval process involves selecting one or more concepts to use as a subject query and retrieving the sets of documents linked to those concepts.One possible formulation of this view is to describe the set of documents linked to the same concept as a kind of "document extension" of that concept.Some approaches to mapping are based on this notion, as when conceptual equivalences between concepts are derived by measuring the overlap between the document extensions with which they are associated [25].
Information retrieval use cases are an important motivation for SKOS, and the SKOS data model is perfectly compatible with this view.However, subject indexing and retrieval are not the only uses for knowledge organisation systems, so SKOS does not attempt to normatively define or formalize any relationship between documents and concepts.This flexibility enables SKOS to represent knowledge organisation systems used in a variety of applications, as well as enabling implementers of information retrieval systems to explore alternative retrieval strategies and algorithms such as query expansion.
Moreover, conceptual vocabularies need not be intended primarily for describing documents for information retrieval.At the most basic level, applications merely require that a concept have identity and that it have features which distinguish it from other concepts within a KOS, such as natural-language labels, definitions, and semantic relations to other concepts.The group felt that this simple, flexible model would cover most of the available use cases and requirements while enabling a broad range of applications, whereas formally reconciling traditional KOS models at a higher degree of granularity would have been both more difficult and of less obvious utility.SKOS concepts vs. OWL classes..The concepts from Knowledge Organization Systems (and hence, SKOS concepts) are often wrongly interpreted as classes from formal ontology languages like OWL.Some SKOS concepts indeed reflect universal categories that also appear in in OWL ontologies, such as "animals" and "cats" in a vocabulary animals.Yet, as seen previously, SKOS concepts are not by default provided with precise extensional semantics, while an OWL class explicitly describes a collection of individuals.Following the example above, the interpretation of an OWL Class "Butterfly" would be the set of butterflies.
From a technical perspective, SKOS Concepts are simply individuals in an OWL ontology (see Section 4).These individuals are interpreted as arbitrary elements in the domain that might (or might not) correspond to collections of indexed documents.
Hierarchical relationships between SKOS Concepts, moreover, might (or might not) correspond to sub-class relationships between OWL classes (owl:Class).Where a KOS might assert broader links between "Dog," "Collie" (a type of dog), and "Lassie" (a particular dog), an OWL ontology might specify that "Collie" is a subclass of "Dog" (rdfs:subClassOf), then "Lassie" could be described as an instance of "Collie" (rdf:type).Some KOS standards distinguish a class-instance variant of "broader," such as the "broader instantive" relationship of ISO 2788 [9].(Indeed, a preliminary draft SKOS Extensions Vocabulary Specification with "broader instantive" and other such variants of the semantic relation properties was posted for discussion in 2004. 15) Many KOSs, however, use the same general relation for all hierarchical links.
As the SKOS Reference says, concept data could be "facts about the thesaurus or classification scheme itself, such as 'concept X has preferred label "Y" '; not facts about the way the world is arranged within a particular subject domain, as might be expressed in a formal ontology."As hinted in Section 2, converting a traditional KOS into an OWL ontology may thus require some hard "cleaning."Conversion into a straightforward, lightweight representation is often the most appropriate course of action [2, 24].One may see formal (OWL) ontologies as KOSs, as they organize knowledge in a domain; but clearly not all KOSs can be directly interpreted as formal ontologies.
Disjointness of the SKOS Concept Class..The previous paragraph highlights that KOS concepts function in a quite specific way: one can view them as proxies that establish an "indirection layer" between lexical entities and "real-world" ones, either represented as individuals (say, a person) or classes (say, all persons with a specific role) in the OWL sense.A test for identifying such resources could be for example the "date of creation" associated with them.The resource that stands for a person in a name authority file (thus, represented as an instance of SKOS concept) will probably have a different date associated to it than the one associated to the resource that stands for that person as a "real person" (represented using the foaf:Person class). 16his observation leads to a first kind of representation pattern, which distinguishes separate KOS concepts from entities in the real world (or classes thereof) and represent the referential link between them using properties such as foaf:focus (defined as "The underlying or 'focal' entity associated with some SKOS-described concept." 17).This is for example what was retained for the Virtual International Authority File (VIAF), 18 which creates for each cluster of authority records an instance of foaf:Person and (at least) one instance of skos:Concept, linking the latter (e.g., http://viaf.org/viaf/sourceID/SELIBR%7C317488#skos:Concept)to the former (http://viaf.org/viaf/85312226)using foaf:focus.
Such approaches would fit well a modeling choice making skos:Concept disjoint with other classes of entities, such as foaf:Person or meta-modeling classes like owl:Class.This would rule out that different "modelling streams," each coming with different kind of possibly incompatible data, are "crossed" within one same graph.(The cultural reference in working-group discussions on this topic was that of the dire warning, from the 1984 film Ghostbusters, 19 never to "cross the streams" of proton beams from multiple particle throwers because, vaguely but ominously, "it would be bad.")However, the working group opted for not asserting explicit disjointness between SKOS concepts and non-SKOS classes.The first reason is quite pragmatic: the world of ontologies beyond SKOS is wide, and choosing specific classes for disjointness statements would have been an incomplete, biased effort.skos:Concept is only formally disjoint with other classes in the SKOS namespace: skos:Collection, skos:ConceptScheme, skosxl:Label, which will be introduced below.Advocates of the disjointness option may argue that SKOS could have featured a new class of "Non-Concepts" to handle the case, next to a property to relate the two disjoint categories. 20At that time, however, the group felt that the proposed patterns were still not mature enough and quite out of the scope defined in our charter, because such information is usually not present in the data defining concepts in existing KOSs.
The second reason for not declaring SKOS concepts to be disjoing with "nonconcepts" stems, again, from the requirement for minimal commitment, as well as from a concern not to rule out valid patterns.As pointed out above, a skos:Concept is intended to provide a neutral target for migrating a wide diversity of KOS concepts to the Web of Data.This includes cases where concepts are also elements of formalized ontologies.Some OWL properties and classes can be seen as members of a concept scheme as, for example, in applications that would not handle the full complexity of OWL reasoning but would require lexical annotations richer than those supported by OWL.The Library of Congress, for example, represents MARC relators21 both as SKOS concepts and OWL properties.The SKOS Primer discusses cases in which it might make sense to treat an instance of SKOS Concept also as a class. 22

Concept Schemes
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
Sets of concepts were referred to as "schemes" as early as the Limber Project (1999-2001). 23It is worth noting that the Dublin Core community began talking in 1997 about "schemes," 24 one type of which came to be called a "vocabulary encoding scheme"25 -a notion much less specific than, but not incompatible with, the SWAD Europe project's notion of a SKOS Concept Scheme.
The "ability to explicitly represent the containment of any SKOS individual or statement within a concept scheme" was accepted by the working group as a candidate requirement for SKOS (R-ConceptSchemeContainment).The ability to express the containment not only of particular concepts, but also of particular statements using SKOS predicates, such as skos:broader, was seen as necessary for tracking the provenance of a concept scheme's informational content, for example to establish trust.Two properties from the 2005 SKOS Core specificationskos:inScheme and skos:hasTopConcept-already provided a way to relate SKOS concepts (along with instances of other classes, such as skos:Collection) to a given concept scheme.
Expressing the containment of statements, on the other hand, implied a mechanism for denoting an entire set of statements as a named entity-a challenge faced by any RDF-based application and thus not specific to SKOS.The use cases for containing relations between concepts also seemed more marginal than for containing concepts.The text used to close this issue (Issue 36)and Section 5.3 of the SKOS Primer point to ongoing work on named graphs and RDF Datasets in SPARQL 26 -work which at the time of writing in 2012 remains on the agenda of the W3C working group developing RDF 1.1. 27

Semantic Relations
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
The properties skos:broader, skos:narrower and skos:related are referred to collectively as the SKOS semantic relation properties.They have their origins in thesauri-controlled structured vocabularies used primarily for keyword indexing of collections of documents or other objects [9].Thesaurus standards provide guidance on the use of hierarchical and associative relationships when constructing a thesaurus.However, because these relationships exist primarily as aids to information retrieval, some ambiguity is permitted.There has been no need for them to support precise formal entailments (such as those supported, for example, by a class subsumption hierarchy in an ontology).Given this context, a number of design decisions needed to be made during the standardisation of SKOS concerning the formal definition of the properties skos:broader, skos:narrower and skos:related.Some constraints were deemed uncontroversial.For example, skos:broader and skos:narrower form an inverse property pair (S25), describing the two directions of a hierarchical relationship.If the concept "mammals" is linked to the broader concept "animals," then the concept "animals" is linked to the narrower concept "mammals."The property skos:related is symmetric (S23) because the fact that two concepts are associated with each other is independent of direction (although sub-properties of skos:related may be defined as directional, i.e., non-symmetric).If the concept "birds" is related to the concept "ornithology," then the concept "ornithology" is related to the concept "birds."In SKOS, hierarchical and associative relationships are declared to be disjoint (S27).
Other constraints, less obvious, concerned transitivity, sub-properties of semantic relations, reflexivity and cycles.Hierarchies and transitivity..It was decided that the properties skos:broader and skos:narrower would not be transitive, and that by convention these properties should only be used to assert direct (i.e., immediate) links between concepts.This decision was made to simplify implementation.For example, many applications will render hierarchical relationships as a tree, and so need some convenient way to differentiate immediate links (parent/child) from indirect links (ancestor/descendant).To support the fairly common use case where applications do want to use the transitive closure of skos:broader or skos:narrower (e.g., to expand a search query), transitive super-properties skos:broaderTransitive and skos:narrowerTransitive were defined.Note that exactly how retrieval applications make use of skos:broader, skos:narrower and skos:related to improve recall is not defined by the SKOS specifications and is left to the application.Some applications will take the view that if document D is indexed with concept A, then document D will always be relevant to a query for any concept that is an ancestor of concept A in the hierarchy.This is equivalent to the view that if a document is about growing vegetables, then the document is necessarily also about gardening.If the property dc:subject is taken to represent the "aboutness" relationship between document and concept, then this behaviour may easily be implemented, for example, by computing the transitive closure of skos:broaderTransitive.However, other applications may take the less categorical view that relevance is likely to degrade as a query is expanded away from some focal concept, and that the different properties skos:narrower, skos:broader, and skos:related might correspond to different shapes or rates of degradation [23].
Sub-properties of skos:broader and skos:narrower.. Some thesauri disambiguate the hierarchical relationship into one of class subsumption, instantiation, or partwhole relationship.The working group discussed whether the SKOS standard should define sub-properties of skos:broader, such as the putative properties broaderGeneric (for class subsumption), broaderInstantive (for class instantiation) and broaderPartitive (for part-whole relations), corresponding to distinctions made in thesaurus standards [9, 10] (see also Issue 56, Issue 150, Issue 178).There would be obvious value in having a standard set of properties, rather than leaving it to third parties to define their own extensions to the detriment of interoperability.However, the working group was also conscious that there is overlap here with RDFS and OWL and was reluctant to define new properties that might be redundant with existing standards.For instance, one might have seen broaderGeneric as equivalent to rdfs:subClassOf and broaderInstantive to rdf:type. 28Whether or not it would be appropriate to use such RDFS or OWL properties in these cases would require a deeper understanding of use cases in which SKOS and OWL are used in combination.The working group decided to postpone this decision, leaving it open for a future revision of the standard.
Reflexivity and cycles..For a conventional thesaurus or similar vocabulary, it is an error for a concept to be in a hierarchical relationship with itself, or to be associated with itself (reflexivity).It could be argued that these constraints should become part of the SKOS data model by stating formally that skos:broader and skos:related are irreflexive properties, and that this would promote consistent implementation.Similarly, in a thesaurus it is an error for there to exist any cycles within the concept hierarchy, which could be enforced by declaring skos:broaderTransitive to be irreflexive.However, the working group was also conscious that there are possible advanced usage patterns (or extensions to SKOS) where both SKOS and OWL 28 http://lists.w3.org/Archives/Public/public-swd-wg/2008Mar/0037.html would be used together within the description of the same knowledge organisation system, and that more work was required to understand these patterns.Although this may be an uncommon edge case, under some of these patterns, it is conceivable that inferences such as "<A>skos:broader <A>" could arise (for example, if someone were to assert that rdfs:subClassOf is a sub-property of skos:broader).It was therefore decided that no formal statements on the reflexivity of the SKOS semantic relation properties would be made, although some informal guidance would be given to application developers on how to detect structural features that are likely to represent errors in the majority of use cases.

Mapping Relations
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
The use cases for SKOS confirmed a strong requirement for mapping between related concepts in different concept schemes.Indeed, the prospect of enabling machine-readable mappings between concept schemes developed in a diversity of contexts, and possibly on the basis of different modeling principles, was expected to be a key advantage of expressing those systems in the common language of SKOS.Taking as its starting point an unfinished SKOS Mapping Vocabulary Specification from 2004,29 the working group settled on five mapping properties: skos:broadMatch, skos:narrowMatch, skos:relatedMatch, skos:closeMatch, and skos:exactMatch, all of which were declared, either directly or by inference, to be sub-properties of skos:mappingRelation, itself a sub-property of skos:semanticRelation.
Much of the discussion about mapping properties revolved around clarifying how they differed from analogous semantic relation properties.The mapping properties skos:broadMatch, skos:narrowMatch, and skos:relatedMatch were declared to be sub-properties, respectively, of skos:broader, skos:narrower, and skos:related.However, these "parallel" properties were not otherwise distinguished in a formal sense.The question indeed arose whether, given this lack of formal-semantic distinction, separate properties for broader, narrower, and related matches were needed at all.
The dilemma, as the working group saw it, was that large parts of the KOS community saw inter-KOS mapping relations and intra-KOS semantic relations as fundamentally different things, perhaps even disjoint from each other.From the standpoint of the working group, the intended distinction between mapping relations and semantic relations depended, conceptually, on the ability to "contain" a concept scheme, along with its intra-KOS relations, as an entity distinct from other concept schemes -an issue, as discussed in Section 3.2 above, which the working group considered to be out of scope for SKOS per se.Even if a distinction between mapping relations and semantic relations might, in principle, be anchored in a formally solid notion of concept scheme containment, the group recognized that the evolution of concept schemes over time could mean that related concepts in two separate schemes could become aggregated into the same scheme, or vice versa -situations in which the use of formally disjoint mapping and semantic properties would prove to be most inconvenient.
The solution adopted by the group was to make the formal-semantic distinction between mapping and semantic properties very weak while emphasizing the "conventional" difference between the two types.As explained in the SKOS Primer, "By convention, mapping properties are used to represent links that have the same intended meaning as the 'standard' semantic properties, but with a different application scope.One might say that mapping relationships are less inherent to the meaning of the concepts they involve. . . .By convention, mapping relationships are expected to be asserted between concepts that belong to different concept schemes."The authors of this paper are not aware that the lack of a strong formal distinction between mapping and semantic properties has been flagged as a problem in the three years since the publication of SKOS as a W3C Recommendation.
The two other mapping properties, skos:closeMatch and skos:exactMatch, were positioned in part as alternatives to owl:sameAs, at the time much overused as a mapping predicate for Linked Data.skos:closeMatch was intended for use with concepts sufficiently similar to be used interchangeably in a given context.The property was not defined as transitive in order to avoid the uncontrolled propagation of the similarity relation to further contexts.skos:exactMatch, defined as a transitive sub-property of skos:closeMatch, was intended to express a degree of similarity close enough to justify such propagation.

Lexical labels
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
The ability to annotate a concept for purposes of display or search is met by properties for preferred, alternative, and hidden labels (skos:prefLabel, skos:altLabel, and skos:hiddenLabel), all sub-properties of rdfs:label.These properties are typically used to link an instance of skos:Concept to an RDF plain literal, which the working group took to mean a character string (such as the word "love") combined with an optional language tag (e.g., "en-US").Note that because RDF lacked a class for RDF Plain Literal, the property definition axiom S12 could not be expressed as a formal range assertion in the normative RDF/XML or informative OWL 1 DL expressions of SKOS, as shown in Table 2 (see also Section 4.3).The 2005 SKOS Core specification included properties for symbolic labels-skos:symbol and skos:prefSymbol-which allowed for resource types other than RDF plain literals as labels; however, these were dropped for the 2009 SKOS Recommendation due to a lack of clear requirements (see Issue 76 and Issue 180).
The notion of "preferred label" derived from what the thesaurus community calls a "preferred term" or "descriptor"-i.e., a "term specified by a controlled vocabulary for use to represent a concept when indexing." 30Preferred terms are in principle unique within the representation of a concept scheme in a given natural language.In order to give formal expression to this convention, the working group 30 http://www.willpowerinfo.co.uk/glossary.htm#preferred_termoperationalized the notion of language as meaning language tag, noting that language tags can be extended to distinguish arbitrarily specific regional variants of, say, English, Portuguese, French, or Chinese.The integrity condition axiom S14, therefore, specifies that a resource "has no more than one value of skos:prefLabel per language tag" (see Section 4.2 and Table 2).The properties for alternative and hidden labels were intended for non-preferred indexing terms, whether displayed to users or not.The three labeling properties are considered pairwise disjoint (S13), such that assigning the same literal as both a preferred and alternative label is formally considered an error.
In keeping with the principle of minimal ontological commitment, the SKOS labeling properties have no explicit domain contraints.This follows the example of Dublin Core, which does not specify domains for many of its properties.The lack of specific domains allows the SKOS labeling properties to be used in contexts other than concept schemes, providing Semantic Web applications with a generic vocabulary for labels-a usage already seen in various OWL ontologies [13] and supported by non-SKOS-centric tools such as Protégé.
In order to address the need for associating concepts with alphanumeric codes such as "M1495-2199" (meaning "Vocal music" in Library of Congress Classification), the working group introduced a property skos:notation (see Issue 79).A SKOS notation is intended to uniquely identify a concept within a given concept scheme.It differs from a lexical label "in that a notation is not normally recognizable as a word or sequence of words in any natural language."As explained in SKOS Reference, Section 6.5.1.,"By convention, the property skos:notation is only used with a typed literal in the object position of the triple, where the datatype URI denotes a user-defined datatype corresponding to a particular system of notations or classification codes."31Relations between labels..The ability to model binary relations between lexical labels was identified as a candidate requirement for SKOS (R-RelationshipsBetweenLabels).It should be possible, for example, to assert that the label "FAO" is related to the label "Food and Agriculture Organization" via a relation "acronym for."The proposals initially considered for enabling such assertions offered combinations of three basic ideas: creating a class for instantiating a "term" to which a plainliteral label could be associated; dropping range restrictions on the SKOS labeling properties so that they could be associated with either RDF plain literals or with instances of such a class; and viewing relations, such as the "acronym for" relation, as classes.Instances of such classes would be linked from a concept, via (for example) a seeLabelRelation property, and would link, via an n-ary relation pattern [15], both to a full form (ex:fullForm "Food and Agriculture Organization") and to an acronym form (ex:acronymForm "FAO"@en)-a pattern which, it was recognized, would involve replicating the label literals. 32he solution that emerged-split off into an optional appendix, "SKOS eXtension for Labels (SKOS-XL)," with its own SKOS-XL namespace URI, in order to keep the main SKOS specification as simple as possible33 -defined a class, skosxl:Label, instances of which are associated with exactly one literal form (see SKOS-XL axiom S52 in Table 3).The properties skosxl:prefLabel, skosxl:altLabel, and skosxl:hiddenLabel were coined, with the class skosxl:Label as their range.The property skosxl:labelRelation was coined as a common super-property for applications defining their own specific label relations.The working group felt that defining properties for specific types of label relation was out of scope due to insufficient consensus on what would comprise a reasonably complete set.
In order to ensure the interoperability of data created using the SKOS and SKOS-XL labeling properties, three axioms were formulated to declare a property chain composed of a SKOS-XL labeling property with a literal form.For example, the chain "(skosxl:prefLabel, skosxl:literalForm)" is a sub-property of the corresponding SKOS labeling property (in this case, skos:prefLabel) (see axioms S55, S56, and S57 in Table 3).In other words, SKOS-XL labels can be "dumbed down" to corresponding SKOS labels.It is worth noting that the skosxl:literalForm property chain is analogous to a pattern described in the 1999 W3C Recommendation for RDF, whereby one of the properties of a "structured value" is marked, using the property rdf:value, as "the principal value of the main relation" of a subject to a value resource. 34efining labels, optionally, as individuals that could be annotated or related among themselves in arbitrary ways allowed the working group to resolve an issue raised with regard to the assertion of mapping relations between the labels of different concept schemes (Issue 49) and an issue requiring the capability of applying annotations to the lexical items used as labels (Issue 27).Two concerns that arose during discussions of modeling alternatives for label relations were: identity conditions (When are two instances of the class skosxl:Label the same individual?), and the formal relationship between the class skosxl:Label and the set of RDF plain literals (Can instances of the class skosxl:Label have more than one literal form?). The working group decided to assert that instances of skosxl:Label have exactly one literal form in order to avoid ambiguity, but that sharing a common literal form should not be sufficient to infer that two instances of the class skosxl:Label were the same individual.In other words, two distinct instances of skosxl:Label might have the same literal form; there is no one-to-one mapping between the class extension of skosxl:Label and the set of RDF plain literals.

Documentation Properties
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
SKOS provides a number of documentation (or note) properties.These allow for a variety of annotations including general notes, change notes, definitions, editorial notes, examples, historical notes, and scope notes.These seven note types provided are not intended to be exhaustive, and it is expected that specific application domains may extend the documentation properties (potentially via sub-properties of the given properties, thus allowing generic SKOS machinery access to information asserted using bespoke properties).As with labeling properties, no domains are given for these properties, allowing their usage outside of SKOS concept schemes.
In addition, the documentation properties have no ranges asserted (in contrast to labels).As discussed in the SKOS Primer, 35 this allows for a number of different documentation patterns, including the use of literals, the use of blank nodes for structured annotations, and the use of document references.

Concept Collections
Using SKOS, concepts can be identified using URIs, labeled with lexical strings in one or more natural languages, assigned notations (lexical codes), documented with various types of note, linked to other concepts and organized into informal hierarchies and association networks, aggregated into concept schemes, grouped into labeled and/or ordered collections, and mapped to concepts in other schemes.
In thesauri and other structured KOSs, concepts can be grouped into semantically meaningful bundles.For example, arrays are used to group specializations of a concept that share a common feature: the concept "cups" might be specialized into a first group of "cups by form" ("stemware", "tumbler". . . ) and a second group of "cups by function" ("coffee cups", "ice cream cups". . . ) [10].This is especially useful for displaying KOSs: these groups are indeed most often meant as a navigation aid in a conceptual network, not to be used for describing resources.SKOS supports the requirement (R-GroupingInConceptHierarchies), discussed in Issue 33 for representing such constructs using the skos:Collection class and its subclass skos:OrderedCollection for groups where the ordering of concepts matters.
Note that SKOS defines skos:Collection as disjoint with skos:ConceptScheme and skos:Concept.This has important consequences.First, it can raise issues when representing "subsets of vocabularies" such as micro-thesauri in the Eurovoc thesaurus36 or subdivision lists in the Library of Congress Subject Headings37 .The disjointness constraint forces data modelers to opt for using (a sub-class of) either skos:Collection or skos:ConceptScheme, a choice that can be hard to make in the absence of clear guidance in the SKOS documentation.Eurovoc thus now represents microthesauri as concept schemes, while LCSH represents subdivision lists as collections.Fortunately, the KOS community has realized this and started to address the problem, as witnessed by recent advocacy on how to relate ISO 25964 thesaurus standard's "concept groups" to skos:Collection and skos:ConceptScheme [11].
Note, too, that collections cannot be used in combination with semantic relations to assign them a position in the semantic structure of a KOS.It is not consistent with the SKOS data model to declare a collection to be a semantic generalization or refinement of a "normal" SKOS concept with statements using skos:broader.In SKOS, concepts are merely grouped into collections using the properties skos:member and skos:memberList.It may be seen as an obstacle to represent simply semantic hierarchies with collections, and a deviation from the minimal commitment approach.But it is in fact the consequence of a conscious choice to keep data on semantic relations between concepts clearly separate from the display-related considerations that usually motivate the creation of collections.SKOS takes the stance that fitting collections into KOS hierarchies must be handled by specific display algorithms that reflect the need of users in a given navigation environment (see Issue 84).

Issues deemed out of scope
Originally chartered for just 20 months,38 the Semantic Web Deployment Working Group needed 35 months to complete its work.In order to focus its efforts and keep the specification as short and simple as possible, the group declared several topics to be out of scope.
• Concept coordination.Many KOSs are intended to be used as building blocks for constructing "coordinated" concepts, for example to aggregate the "simple" concepts "aspirin" and "side effect" into a "compound" concept "aspirin -side-effects."Compound concepts can be created on a one-off basis by catalogers, as they are needed in resource description, or they can be added as concepts to the KOS itself by its maintainers (which is known as "precoordination," as with the Library of Congress subject heading "China -history").The working group recognized this well-known pattern -"the ability to create new concepts from existing ones, e.g. by using special qualifiers that add a shade of meaning to a normal concept"-as a candidate requirement (R-ConceptCoordination).The group also considered a common practice in the thesaurus world [9] whereby two simple concepts (such as "Road transport" and "Safety") are designated to be used in combination instead of minting a new compound concept such as "Road safety" (see Issue 45).
After much discussion, the group decided to postpone these issues (Issue 40, Issue 131).While the requirements for coordination were not questioned, the group considered them to be relevant more for particular thesaurus and subject heading applications than to the interchange of KOSs generally.The group also noted that the patterns proposed to represent concept combinations were rather complex and largely untested.Finally, it was felt that allowing the core SKOS model to handle such constructs could be seen as a potentially confusing move towards supporting some functions of formal ontology languages such as OWL-languages which support the definition of complex classes or properties from more primitive vocabulary elements.
In retrospect, the authors feel that the decision to postpone was sound.It not only kept untested patterns out of SKOS, avoiding delays in finalizing the standard; it also motivated the community to tackle the issue itself.By the end of 2010, for example, the Library of Congress had developed a first version of MADS/RDF [16], an extension to SKOS which, among other things, supports concept coordination within library subject heading lists.
• Subject indexing.As defined by Leonard Will, subject indexing involves "intellectual analysis of the subject matter of a document to identify the concepts represented in it, and allocation of the corresponding preferred terms to allow the information to be retrieved." 39The working group recognized as a candidate requirement the "ability to represent the indexing relationship between a resource and a concept that indexes it," whereby the SKOS model would include "mechanisms to attach a given resource (e.g.corresponding to a document) to a concept the resource is about, e.g. to query for the resources described by a given concept" (see R-IndexingRelationship, Issue 77).Noting the existence of indexing relation properties in other vocabularies, such as Dublin Core's dc:subject, the working group declared such properties to be out of scope and decided not to carry forward the property skos:subject from the 2005 SKOS Core specification. 40(It should be noted that recent Web crawls show skos:subject to be one of the most-used properties from SKOS, due primarily to its use in DBPedia. 41) For lack of a SKOS indexing vocabulary, a candidate requirement for distinguishing between indexing and non-indexing concepts was also declared out of scope (see 39 http://www.willpowerinfo.co.uk/glossary.htm 40http://www.w3.org/TR/2005/WD-swbp-skos-core-spec-20051102/#subject 41 http://sindice.com/search?q=&nq=%28*%20%3Chttp%3A%2F%2Fpurl.org%2Fdc%2Fterms%2Ftitle%3E%20*%29&fq=&interface=advanced R-IndexingAndNonIndexingConcepts, Issue 46).
• Provenance information about mappings.The ability "to record provenance information on mappings between concepts in different concept schemes" was recognized as a candidate requirement for SKOS (R-MappingProvenanceInformation).
The issue was resolved with a decision not to introduce specific SKOS vocabulary about the provenance of mappings (Issue 47).Rather, the group felt that this issue depended on the use of standard containment mechanisms for encompassing mapping assertions within a context that could be denoted with a URI-an issue relevant for RDF in general, specifically for the future development of standards regarding "named graphs" (see also the discussion of containment in Section 3.2).
• Describing concept schemes.Concept schemes have authors, titles, publishers, dates issued, subject coverage, and the like.The working group felt that the question of what properties to use in describing a concept scheme was an issue best left to communities of practice.Shortly after the publication of SKOS in 2009, for example, a joint DCMI-NKOS task group was formed between the Dublin Core Metadata Initiative and Networked Knowledge Organization Systems community to develop an application profile and a KOS Type Vocabulary for describing KOSs.42 • Concept evolution.The working group acknowledged the importance of mechanisms for representing the temporal evolution of concept schemesan issue that raises questions of granularity (whether to version individual statements, concept descriptions, or entire concept schemes) and of how to represent such versioning information in interoperably machine-readable ways.The group considered this topic best left to the community for research and testing (see [22]).

Formal semantics
This section discusses aspects of SKOS relating to its formal semantics, in particular highlighting the use of OWL.The working group43 was tasked to specify SKOS in accordance with OWL, so as to allow for applications to validate SKOS datasets or to infer new facts from the ones explicitly encoded by publishers of SKOS data.The SKOS model is thus specified by defining OWL classes and properties, which can be interpreted using OWL's formal semantics.A particular SKOS concept scheme is an instantiation of the OWL ontology that defines SKOS in which SKOS concepts are instances of the class skos:Concept with characteristics expressed using the SKOS properties.

Axioms supporting inference
As described above, the SKOS data model contains a number of axioms (stated as S1 to S46 in the Recommendation44 ) relating to the classes and properties of the SKOS vocabulary.
All but six of these axioms, as listed in Table 1, describe how the classes and properties of SKOS are defined, primarily by stating subclass or sub-property relationships or domain and range assertions.These axioms allow the use of inference engines ("reasoners") to derive additional information about the nature of, and relationships among, components of a concept scheme.Note that such inference concerns the concept scheme as an information artefact in itself and says nothing about the nature of the resources or "real-world" entities to which the concepts of a concept scheme may refer.For example, the axiom S4 allows the inference that an object of a triple using skos:inScheme is an instance of the class skos:ConceptScheme. 45Axiom S25 allows an application that is OWL-semantics aware to infer the presence of skos:broader relationships in a concept scheme that asserts only skos:narrower relationships.

Integrity conditions
In addition to the axioms described above, a number of integrity conditions (labeled as S9, S13, S14, S27, S37, and S46) are also given.The integrity conditions serve a different purpose to the other axioms stated, in that they are intended to facilitate and promote interoperability by defining circumstances under which data are not consistent with respect to the SKOS data model.Details of integrity conditions are given in Table 2.
The working group was chartered to create a machine-readable specification of the SKOS axioms using the OWL language, which forms the base for exchanging and exploiting formal specifications of ontologies on the Web of Data, as envisioned in the W3C Semantic Web technology stack.The SKOS Recommendation makes no assumptions, however, as to how implementation of the checking of integrity conditions for a particular concept scheme are performed.They could be checked through inference, but other mechanisms could be used, for example querying for particular graph patterns or the use of rule driven approaches such as SPIN 46 or Pellet's Integrity Constraints. 47

SKOS as an OWL Ontology
Historical context..The Web Ontology Language (OWL) was first published as a collection of W3C Recommendations in 2004 48 developed by the Web Ontology working group, 49 first convened in 2001.One key aspect of OWL was the definition of three sublanguages known as OWL Lite, OWL DL and OWL Full.OWL DL supported those users who wanted maximum expressiveness while still retaining computational completeness.OWL Full provided greater expressiveness and syntactic freedom, but with a lack of computational guarantees.OWL Lite was a subset of OWL DL intended to support users needing a classification hiearchy and simple constraints.The working group was tasked to specify SKOS in accordance with OWL Full.
In 2007, the OWL Working Group 50 was convened, with a charter to produce an update to OWL, resulting (in 2009) in a collection of recommendations defining OWL 2 51 (also earlier known as OWL 1.1 during the process).The work of the OWL Working Group overlapped with the work of the Semantic Web Deployment Group, with the consequence that the SKOS recommendation did not have the opportunity of using OWL 2 features in the SKOS recommendation (this point and the related issue of defining SKOS within the limits of OWL DL in is covered in more detail in Section 4.4).
To avoid confusion, this section refers explicitly to the original (2004) recommendation as OWL 1 and the revision (2009) as OWL 2.
SKOS as an OWL ontology..The SKOS data model is represented as an OWL 1 ontology, i.e. a collection of classes and properties with associated axioms.
The SKOS Namespace Document RDF/XML Variant 52 provides definitions of the classes and properties of this model using OWL 1, along with axioms that represent integrity conditions on the data represented using SKOS.As there are limits to the expressivity of OWL 1 (and of its subspecies or fragments), not all of the desired constraints can be fully expressed using OWL 1.This is further discussed below.Where this is the case, the constraint is expressed as a comment in the schema.

Compatibility with OWL 1 DL and OWL 2
SKOS is defined as an OWL 1 ontology, and a requirement R-CompatibilityWithOWL-DL made to the working group was that SKOS should provide a legal OWL 1 DL ontology, primarily to ensure compatibility with editing tools and to facilitate the use of reasoners, many of which operate in the OWL 1 DL space.
This was problematic as OWL 1 DL lacked the expressivity needed to capture some of the assertions.For example, OWL 1 DL has no facility to express hierarchies of annotation properties.Nor does OWL 1 provide a mechanism for stating axioms concerning property chains as used in axiom S55.A further complication was that the work of the Semantic Web Deployment Group overlapped with that of the OWL Working Group, which was defining the OWL 2 Recommendation (also referred to as OWL 1.1 during the process).OWL 2 was likely to introduce features that would support some of these assertions, but as the OWL Working Group was scheduled to finish after SKOS delivery, the normative SKOS reference could not make reference to OWL 2. For example, the particular feature supporting hierarchies of annotation properties was ultimately introduced into OWL 2.
In order to provide some support for reasoning engines and those applications working in the OWL 1 DL space, a "pruned" RDF schema was produced, providing a non-normative resource.This is made available (in a non-normative fashion) as the SKOS RDF Schema -OWL 1 DL Sub-set. 53In particular, the pruning removed axioms stating that SKOS labeling properties are sub-properties of rdfs:label as sub-property axioms are not applicable to annotation properties in OWL 1 DL.
This particular pruning of the schema in order to provide a valid OWL 1 DL ontology is only one of a number of possible ways in which the OWL 1 Full RDF Schema for SKOS can be adjusted in order to sit in the OWL 1 DL space-each of which would have differing semantic consequences.As a result the OWL 1 DL prune was considered non-normative. 50http://www.w3.org/2007/OWL 51 http://www.w3.org/TR/owl2-overview/ 52 http://www.w3.org/TR/skos-reference/skos.rdf 53http://www.w3.org/TR/skos-reference/skos-owl1-dl.rdfOther constraints were also problematic in terms of OWL 1 representation.S14 states that "A resource has no more than one value of skos:prefLabel per language tag."This was not expressible in OWL 1. Nor were property disjointness constraints as expressed in S13, S27 and S46.Issues relating to compatibility with OWL 1 DL-Issue 38, Issue 137, Issue 138-were thus formally postponed by the working group, indicating that, should work resume on an updated recommendation, this should be the focus of attention.
Comments from members of the OWL Working Group (raised as Issue 155 and Issue 157) highlighted areas where an adjustment to the model would potentially provide better alignment with the emerging OWL 2 recommendation.Issue 157 was formally postponed.Following the resolution of Issue 135, labeling properties were defined as owl:AnnotationProperty.

Machine-readable Formalizations, Formal Semantics and Data Quality
As discussed earlier, the SKOS data model is represented as a collection of axioms, some providing definitions of classes and properties, which then support inference, others asserting integrity conditions.When representing these axioms in a machine-readable way (which was the main mission of the SWD group), the implementation creates two "layers" orthogonal to this question of definition versus integrity: 1. Axioms formally represented in the ontology, for example, sub-property relations to skos:semanticRelation; 2. Axioms that are not explicitly represented in the ontology, primarily due to a lack of expressivity in the representations, for example assertions about disjoint properties.
The SKOS RDF/OWL representation thus proposes a core layer for inference and validation of SKOS data.However, as mentioned in Section 4.2, the working group did not assume a specific technique for checking the integrity conditions of the SKOS data model.This flexibility can be explained by the difficulty of representing all integrity conditions in the OWL language (as discussed above).But it is also in line with a more fundamental stance of the SWD group, which allows for a flexible approach to data quality in SKOS, generally.In addition to the two layers described (formal versus informal axioms), the SKOS reference includes what one might call guidelines which are weaker recommendations, for example that skos:closeMatch should be used to relate concepts from different schemes.There is no attempt at formal representation of the latter, nor is it considered an integrity constraint that should not be violated.These assertions are more "advisory," but are still somehow part of the SKOS model.It is left to SKOS implementations to adapt these guidelines-or others from specific domains, such as thesaurus design guidelines [9], which can provide useful "checks" for SKOS data.
Example approaches to validation of SKOS data include the Poolparty Thesaurus Consistency Checker, 54 which runs custom validation rules derived from the SKOS axioms.The qSKOS tool by Mader et al. [12] is used to identify a number of quality issues in SKOS vocabularies, in particular fifteen "guideline" violations.The Skosify tool [20] identifies an overlapping (but slightly different) set of criteria, some of which correspond to SKOS integrity conditions (e.g., S13 concerning disjointness of alternate and preferred labels).
54 http://poolparty.punkt.at/Tables 1, 2 and 3 provide a summary of the axioms in the SKOS and SKOS-XL data models.It also highlights those axioms that lack a formal machine representation in either the normative RDF Schema or the non-normative OWL 1 DL prune (note there is no corresponding OWL 1 DL prune of SKOS-XL).

SKOS Namespace URI
A question that was the focus of much attention during the Recommendation process was that of the URI to be used for SKOS, formally raised as Issue 153 and Issue 175.Earlier work from the Semantic Web Best Practices and Deployment Group resulted in a SKOS Core Working Draft. 55This Working Draft was a key input to the work of the working group, and much of the content of the original Core was preserved in the final Recommendation.The original core defined vocabulary using the namespace URI http://www.w3.org/2004/02/skos/core.
Various possibilities were open to the working group: 1. Provide a new namespace URI for the SKOS Recommendation; 2. Use the existing SKOS Core namespace URI for the SKOS Recommendation, potentially redefining or changing the semantics of URIs defined in that namespace.3. Use the existing SKOS Core namespace URI for the SKOS Recommendation, minting new URIs for those vocabulary elements where semantics had been changed.
As an example of a situation where the semantics of a vocabulary element had changed, consider the hierarchical semantic relations skos:broader and skos:narrower.
In the original core, these properties were declared to be transitive, while in the final SKOS recommendation, as discussed in Section 3.3, they were not (instead a transitive reduction [17] design pattern was used, introducing transitive superproperties skos:broaderTransitive and skos:narrowerTransitive).
Each option had pros and cons.The introduction of a new namespace URI would reduce the problems of inconsistent interpretations of existing vocabularies that may have been producing using the original semantics.However, a new namespace URI would then potentially require changes to existing tools, infrastructure and concept schemes.
The final decision made was for option 2. It was felt that disruption to the existing body of data that had been published using SKOS Core would have been significant if the namespace URI or property names of key elements have been changed.Although this resulted in a change of semantics to some properties, applications should, in principle, be able to make use of the machine-readable published schema to access those semantics.
On a similar note, elements were removed from the SKOS Core vocabulary (see discussion in the SKOS Reference56 ) although historical versions of the schemas remain available. 57lthough SKOS Core had at that point been deployed by early adopters for several years, changing the semantics associated with the URIs was unproblematic strictly from the standpoint of process because the 2005 specification had only attained the status of Working Draft-a type of specification by definition subject to change.
It is worth noting that following publication of SKOS as a Recommendation, it was observed that the the SKOS OWL 1 DL prune ontology had no version IRI, thus breaking a rule specified in the 2009 OWL 2 recommendation, that "If an ontology has an ontology IRI but no version IRI, then a different ontology with the same ontology IRI but no version IRI SHOULD NOT exist."In order to address this, an additional owl:versionIRI triple was added to the ontology 58 .

Conclusion
The intellectual roots of Knowledge Organization Systems go back decades, even centuries.The goal of expressing Knowledge Organization Systems in a generically interoperable way was raised already as a goal when W3C working groups began developing the Semantic Web language, Resource Description Framework (RDF), in the late 1990s.In the twelve years from the beginnings of RDF in 1997 through the finalization of SKOS as a W3C Recommendation in 2009, the torch for this work was passed among a succession of UK and European research projects and of W3C working groups, each of which added features, dropped others, and progressively clarified its underlying concepts.This process illustrated the challenge of developing specifications that depend on related specifications which, in today's continually evolving environment, are inevitably subject to change.As discussed above, SKOS would have looked slightly different if OWL 2, published as a W3C Recommendation just two months after SKOS, 59 had been finalized just half a year earlier.
This paper highlights a number of issues that were "postponed" -a status which marks them as being of potential interest to future working groups: • SKOS and OWL.In the three years since the publication of the W3C Recommendation for SKOS, one of the most Frequently Asked Questions has been that of the relationship between information KOSs, expressed using SKOS, and OWL ontologies.As discussed in Section 3.1, almost anything can be considered a SKOS Concept (as long as it is not a SKOS Concept Scheme, Collection, or Label), and foaf:focus provides a way to link SKOS concepts to things in the world to which those concepts refer.The working group defined SKOS this way so as not to preclude experimentation with usage patterns as yet unforeseen.
• The formal expression of SKOS.ISO 25964 standard [10].The issue of extending SKOS with symbolic labels remains unaddressed, though the authors see the potential for experimentation related, for example, to Web accessibility.The potential refinement of SKOS mapping properties perhaps awaits a stronger push on formulating best practice for mapping in the Semantic Web context generally (see Issue 176).
• Concept coordination.Patterns defined by MADS and ISO 25964 for concept coordination could be evaluated in light of implementation experience, especially as pre-and post-coordination patterns are tested in the context of different types of KOS and in information retrieval applications (see Issue 40, Issue 45, Issue 131, and Section 3.8).
• Concept scheme containment and provenance.As of late 2012, the RDF Working Group is working towards standardizing an approach to naming graphs and datasets.As pointed out in Section 3.2, the identification of graphs is relevant to all issues which require that concept schemes be delimited, or "contained," for the purpose of tracking provenance or expressing precise alignments (see R-ConceptSchemeContainment and R-MappingProvenanceInformation).The Linked Data community is developing relevant practices and vocabularies, such as VoID 60 , which addresses the provenance both of generic datasets and of more specific "linksets."The SKOS community has made some progress in the past years on modeling concept evolution 61 -an issue of particular interest to builders of SKOS "registries" and APIs.
• Best practices for modeling SKOS Concept Schemes.As discussed in Section 3.7 there are some concept groupings, such as micro-thesauri, to which the architecture of SKOS Collections versus SKOS Concept Schemes does not neatly fit, suggesting a need to clarify best practices.
The W3C Semantic Web Deployment Working Group, which carried SKOS forward during the final three years of this process, began its work with a draft specification, at the time called SKOS Core, which had already been widely deployed and tested by early adopters.In the three years since its publication in 2009, SKOS has become one of the most widely used vocabularies in the Linked Data cloud 62a context to which its flexible, generic design, based on the principle of minimal ontological commitment, is uniquely well-suited.As its designers intended, SKOS continues to be adapted and extended to meet more specialized requirements.Some of this work is discussed on the public-esw-thes mailing list 63 and tracked on the W3C SKOS community wiki 64 .The development of SKOS has been the collective result of several dozen contributors working, typically, in the context of working groups or projects of two or three years' duration.The accumulated impact of such incremental contributions becomes clear only in retrospect, looking back with the perspective of a decade or two.

Figure 1 :
Figure 1: Main elements of the SKOS data model 26 http://http://lists.w3.org/Archives/Public/public-swd-wg/2008Jan/0093.html 27http://www.w3.org/2011/rdf-wg/ The formal expression of SKOS axioms could be enhanced in light of OWL 2 (see Issue 38, Issue 136, Issue 137, Issue 138, Issue 155, and Section 4.4).Such enhancements could help consolidate a more consistent approach to validating SKOS concept schemes.The choice of axioms pruned to create the non-normative OWL DL 1 Prune should also be revisited in light of implementation experience (see Section 4.4).A future working group might also want to formulate recommendations on the use of non-OWL semantics based, perhaps, on constraints with a closed-world interpretation.•Inference, validation, and quality control.A future working group might want to incorporate work being done in the implementation community on extending the definition of "conformance," for example to enhance support for specific types of KOS (see Issue 35).