FAIR + FIT: Guiding Principles and Functional Metrics for Linked Open Data (LOD) KOS Products

Table 1 presents the number of datasets of the original KOS types, including thesauri, classifications, taxonomies, terminologies, and lists, plus the number of SPARQL endpoints provided. Table 2 shows the formats as well as the number of datasets which make data available in that format. The major findings reveal that: (1) Many serialization formats have been used for KOS’ deliverables, such as JSON, HTML, Turtle, N-triples, RDF/XML, CSV, and more. The default/auto format is usually a html table. (2) The number of KOS products offering operational SPARQL endpoints is still very limited, despite a gradual increase. We should understand that, enabling a SPARQL endpoint allows for the targeting of specific bits of data, and the content types of query results can be selected based on the intended usage in applications. (3) The majority of KOS datasets made available via SPARQL endpoints have been implemented with tools including Virtuoso, PoolParty, Fuseki, and ARC SPARQL+. Others have used their own custom-built implementation or do not indicate what tool they are using. (4) Supported SPARQL features differ based on the service implementation, yet we found that all endpoints offered a number of serialization formats for query results.

Our recommendation is that a KOS vocabulary should be delivered in consumable formats: available in various data serialization formats and accessible through a SPARQL endpoint.

Our findings have shown that nearly 80% of endpoints reviewed are operational. Though encouraging, it is still evidence that more than 20% are no longer working. Being able to rely on an endpoint or to anticipate when it might be unavailable will be critical for some users. It is our recommendation that institutions should commit to ensuring the sustainability of access to their KOS dataset deliverables by providing a persistently available SPARQL endpoint.

An assessment of the properties used inside a LOD KOS dataset is necessary because even though the datasets are made available, it is not immediately obvious what classes and properties are involved and what links exist between datasets. Understanding the properties used can help us evaluate how suited a dataset may be for reuse in other contexts; it also allows users to better understand and integrate it in various applications. One of the metrics of dataset performance quality when evaluating a data structure is assessing whether it has complied with W3C standards and if independent specialized properties are used. Our assumptions were that LOD KOS datasets would rely heavily on W3C recommended standard properties from SKOS, OWL, and RDFS. Yet, the initial findings in 2015 alerted us to the fact that a noticeable number of independent specialized properties are used. These are represented by the orange-colored rows in the following partial example from the 2015–16 sample. A similar property check was applied to the 2019 data collected to assess changes over time. In Figure 4, the independent properties used in the datasets are colored in grey.

The findings reveal that properties from standard vocabularies are from SKOS, Dublin Core and Dublin Core Terms, OWL, RDFS, DBPedia ontology, FOAF, and Schema.org. SKOS was highly used especially among datasets tagged as thesauri. RDFS and OWL are increasingly being used. The majority of datasets have included independent specialized properties (as additional or as primary) to represent their structures, which could directly impact their interoperability and reusability. The situation is especially troublesome when there is no hint of how to query and use these properties through an endpoint. Even for users who know the SPARQL query language, they still must know the internal properties and data structures, in order to use the products.

We recommend dataset properties and structure information be more effectively and readily available. A SPARQL service should at least contain refined query examples to reveal the internal structures of the datasets. This will serve to make vocabulary contents reachable, as the usability and reusability of LOD KOS products is a major hurdle yet to be overcome.

A common and increasingly challenging issue for the full usage of LOD products is that, end-users may have difficulty accessing and using the LOD datasets since they might not have been trained to access data dumps or SPARQL endpoints. Therefore, although SPARQL gives users the opportunity to design unique queries, a non-technical user may find themselves at a loss as they encounter endpoint implementations loaded with a default query and no other information. Among the available endpoints checked, less than a quarter of SPARQL endpoints would load with a default query in the query window. From Table 3, for example, it discloses that, in 2019, about 41% of them loaded with a default query; less than 20% provided query examples for users to explore the data; and only a few endpoints made more than three example queries available.

Figure 5 shows an ideal example for endpoint providers to emulate. In the UNESCO vocabularies SPARQL service, multiple query examples are provided such as those for obtaining lists of all concepts, microthesauri, or data values in certain languages.

Figure 5

We highly recommend KOS producers adopt and adhere to best practices like this to enhance usability. Datasets with SPARQL endpoints should provide query examples or forms and templates to enable the easy creation of queries allowing users to interact with the data. This could be accomplished through a showcase of example queries which loads in the query window and is adjustable to user needs (e.g. Wikidata query service

https://query.wikidata.org/

Although such user-friendly products are rare, they illustrate how LOD KOS datasets can be potentially useful to researchers and eventually become knowledge bases (i.e. not just published RDF triple stores of value vocabularies). This will be the primary strategy that will enable LOD KOS to be Transformable (to be discussed in section 4.3).

In summary, this whole section on Functional presents four major criteria and addresses the importance of delivering the vocabulary in consumable formats, ensuring a product is accessible through persistently operational SPARQL endpoint. At the minimal level, a KOS SPARQL endpoint should contain refined query examples to inform the dataset properties and internal structures. The usability and user friendliness can be enhanced by providing default or refined query examples that enable users to explore the data structure and contents of the KOS. When a KOS is Functional, it could further its Impacts and potential Transformable usages.

The first recommended approach is to expose the LOD KOS vocabulary through terminology services such as vocabulary registries and repositories. The direct result of this action would be an increase of visibility. In our “KOS in the Semantic Web” study, major vocabulary registries and services are listed and explained, including (a) vocabulary registries and (b) vocabulary repositories/portals (Zeng, 2018).

https://www.isko.org/cyclo/interoperability.htm#app2

The impact of a KOS can be measured through its usage by data providers in two categories: (a) used as a primary value vocabulary and (b) used in semantic enrichment processes.

Figure 6

Figure 7

Institutions that use name authority data to semantically enrich their digital collections can easily embed the identifiers within individual datasets. By using specific properties of established schemas, for example, owl:sameAs, datasets can link to the HTTP URIs from value vocabularies such as ULAN, TGN, GeoNames, LCSH, and portals such as VIAF and Wikidata. Successful cases can be seen in libraries, archives and museums (LAMs) as well as project-based digital collections (see details at Zeng, 2019). LOD KOS that can be used for semantic enrichment of originally structured, semi-structured, and unstructured data have directly impacted the quality and effectiveness of those data’s delivery on the web, greatly enhancing their FAIR compliance.

To achieve the semantic interoperability of existing KOS vocabularies, activities establishing relationships between the contents of one vocabulary and those of another have seen increased engagement. Mapping is a common process of establishing relationships between the concepts of one vocabulary and those of another. A top-down or centralized full vocabulary mapping could be initiated by one source vocabulary (e.g. AGROVOC) and mapped to other target vocabularies. Alignments require interoperability in syntax & structure. The levels of mapping might be clearly defined based on SKOS, such as the skos:exactMatch and skos:closeMatch. (See example of AGROVOC Alignments report

http://aims.fao.org/standards/agrovoc/linked-data

The bottom-up alignment product Mix’n’match

https://tools.wmflabs.org/mix-n-match/#/

Another approach distinct from vocabulary-based mapping is value-based mapping. A similar volunteer-contributed outcome is the “Authority Control” section in Wikipedia pages for agents, places, works (distinct intellectual or artistic creation), and historical events where identifiers of a THING are provided (e.g. for Leonardo da Vinci

https://en.wikipedia.org/wiki/Leonardo_da_Vinci

Cases of KOS showed and discussed at professional conferences and academic publications provides another way to disseminate, discover, and measure the impact of a LOD KOS. Established methods such as content analysis and bibliometrics would be appropriate for studying their impact. (Since those methodologies are pretty mature, we will not explain them here in detail.) Notable professional conferences include the NKOS workshops

https://nkos.slis.kent.edu/#workshop

In summary, the discussions about the Impactful metric in this section reveal ways to measure the new impacts of KOS vocabularies brought by their advanced LOD products in the 21^st century. Any vocabulary can be exposed through terminology services following the FAIR criteria, used as a primary value vocabulary as well as in semantic enrichment processes, mapped with other KOS vocabularies (whole or part) and aligned with entry-level entities of Wikimedia. Research projects and usages showed or discussed at professional conferences and publications provide evidence of the impacts. All these will help to maximize the impact of a LOD KOS vocabulary, which may affect the investment decisions of the vocabulary itself as an open resource.

LOD brings effective new features to KOS vocabularies, enabling derivation of components from the original datasets in a few seconds. For example, from the UNESCO Thesaurus, dozens of micro-thesauri (e.g. “Geography and oceanography”, “Culture”, “Religion”, “Social policy and welfare”, “International relations”, “Finance and trade”) can be obtained quickly. A micro-thesaurus is a designated subset of a thesaurus that is capable of functioning as a complete thesaurus (ISO 25964-2:2013). Their individual concepts can be also obtained in other languages.

Fully benefiting from the original faceted and hierarchical structures of a KOS vocabulary, a LOD KOS gives users the autonomy to determine what structure and information is desired and can be reproduced. One case worth sharing is the Art and Architecture Thesaurus (AAT). AAT’s Linked Data SPARQL endpoint

http://vocab.getty.edu/queries#Finding_Subjects

Another illustration of this point is the Global Agricultural Concept Scheme (GACS)

http://browser.agrisemantics.org/gacs/en/

These cases demonstrate the benefits of giving users the autonomy to determine what and how they will use the data provided, which acts as incentive to reproduce it in unique applications.

The above cases also apply to T3, as KOS are being extended to fit the diverse needs of language, culture, domain, and structure. This concept is not new for KOS, since several have been internationally adopted and used worldwide in the 20^th century. For KOS to be appropriately adopted, reused, and reproduced in these contexts, the provenance data of the whole KOS or parts are essential for its quality and trustworthiness. Such data can be very well documented and used in the LOD version. Commonly used properties such as dcterms:created, dcterms:modified, skos:changeNote, prov:wasGeneratedBy, and prov:used have been applied to a KOS’s entry level to express changes made to fit language, culture, structure, and domain

Example: http://vocab.getty.edu/aat/300196975

As we consider these cases, especially the last example, we can further ask the question: how can LOD KOS products become something beyond a value vocabulary? Next, we will outline our final T: Supports innovative and transformative uses beyond normal “value vocabularies.”

Through various case studies, we found the newest and most important function of KOS datasets which should be considered as “knowledge bases,” beyond being normal “value vocabularies” (Zeng & Mayr, 2018). With the advancement of the RDF model, a graph data model is considered to be one of the most flexible formal data structures. Among the knowledge bases, “knowledge graphs” have increasingly become a more widely used concept and label.

A Knowledge Graph (KG) is a graph-theoretic knowledge representation that (at its simplest) models entities and attribute values as nodes, and relationships and attributes as labeled, directed edges (Kejriwal, Sequeda, & Lopez, 2019). Prior to coinage of the term “Knowledge Graph”, proponents of the Semantic Web pressed for the use of graph-theoretic models, pattern-matching query languages, graph data management and use of publicly available KGs like DBpedia, GeoNames and Wikipedia for information retrieval as well as knowledge acquisition and alignment (Kejriwal, Sequeda, & Lopez, 2019). Among the many benefits of knowledge graphs, one of the most noticeable is the potential for discovery of hidden knowledge. The contextual information which enables this discovery is provided by the RDF triples which follow the Linked Data principles and are embedded in trustable value vocabularies and property vocabularies.

Consider the SPARQL query examples provided by UniProt (Universal Protein Resource). One of them is “Select all bacterial taxa and their scientific name from the UniProt taxonomy” which is obviously a function that a value vocabulary provides. It also provides over 20 other query examples, such as “Select the preferred gene name and disease annotation of all human UniProt entries that are known to be involved in a disease” and “Select all triples that relate to the taxon that describes Homo sapiens in the named graph for taxonomy” (UniProt Consortium, 2002– 2020)

Examples provided by UniProt https://sparql.uniprot.org/

The important roles of KOS in the creation of knowledge graphs have been emphasized in the past two years, as knowledge graph development has been considered a major strategy for corporations including Google, Apple, Amazon, Alphabet, Microsoft Corp, Facebook, and more. The Microsoft Academic Knowledge Graph (MAKG) set has over eight billion triples with information about scientific publications and related entities as of 2018–11

http://ma-graph.org/

In summary, the discussions about the Transformable in this section imply great potential for KOS vocabularies to extend their functionality and impact through innovative adaptations. Allowing special KOS products to be derived from the original data, giving users autonomy in reproduction, and enabling the extensions to fit diverse needs are major transformable approaches. More importantly, the innovative use of the originally constructed high quality, contextualized data entries enable the LOD KOS to generate large or specialized knowledge graphs, which function as knowledge bases; and to become foundations of semantic analysis and entity extractions. They consequently become the building blocks of a framework for research in humanities and science. LOD KOS products are thus transformed beyond being just “value vocabularies.”