How to create and use a national cross-domain ontology and data infrastructure on the Semantic Web

Abstract

This paper presents a model and lessons learned for creating a cross-domain national ontology and Linked (Open) Data (LOD) infrastructure. The idea is to extend the global, domain agnostic “layer cake model” underlying the Semantic Web with domain specific and local features needed in applications. To test and demonstrate the infrastructure, a series of LOD services and portals in use have been created in 2002–2023 that cover a wide range of application domains. They have attracted millions of users in total suggesting feasibility of the proposed model. This line of research and development is unique due to its systematic national level nature and long time span of over twenty years.

1. Extending the layer cake model

The Semantic Web (SW) sees the Web as an interlinked collection of data (Web of Data) instead of only a space of interlinked hypertext documents, Web of Pages. The idea was proposed in the 90’s by Tim Berners-Lee [4], and first recommendations for the SW,1

¹

https://www.w3.org/standards/semanticweb/

such as the Resource Description Framework2

²

https://www.w3.org/RDF/

(RDF), were developed before the millenium. The recommendations constitute the W3C “layer cake model” [9,13] on top of XML, the lingua franca of the WWW, and lay out a new basis of shared semantics for interoperability of data. Founded on using first order predicate logic, the semantics of the SW [19] are independent of application domains and natural languages. This makes the model suitable for dealing with the versatile multi-domain and multilingual data underlying the Web.

To develop applications, the layer cake model is not enough: domain and application specific infrastructures based on shared W3C standards and best practices are needed, too. These can focus on specific domains, such as medicine, biology, cultural heritage, or geography on an international level. However, in practice one also has to deal with national level issues and data available that are represented using national languages, data models, vocabularies, and are created using conventions of local legacy systems. It therefore makes sense to speak about a “national Semantic Web infrastructure” that is compatible with the global SW standards. For example, Cultural Heritage (CH) data in different countries is often nationally specific calling for adapted local solutions for representing and using the data.

Most of the international infrastructure work [5] is focused on collaborations on particular application domains. In contrast, this paper concerns the question: How to Create a National Cross-domain Ontology and Linked Data Infrastructure and Use It on the Semantic Web. This problem is addressed by presenting, discussing, and evaluating approaches and living laboratory experiments developed in Finland during 2002–2023. Presenting lessons learned in this particular endeavour is hopefully useful in a more general setting, as similar challenges are likely to be faced in other countries, too.

The paper is organized as follows: In Section 2, elements needed for a national SW infrastructure are first introduced. The idea and lessons learned in developing a national ontology and a LOD infrastructure are then presented in Sections 3 and 4, respectively. After this, applications of the infrastructures are discussed: as a proof-of-concept, a model is presented that has been used for creating a series of in-use data services and semantic portals that have had up to millions of users (Section 5). Finally, contributions of the work are summarized and related works discussed (Section 6). This paper presents the first consolidated account of this line of research and development, summarizing works reported before in over 500 papers and other publications available on the Web.3

³

Online publications of the Semantic Computing Research Group (SeCo): http://seco.cs.aalto.fi/publications/.

Fig. 1.

Elements needed for a national Semantic Web infrastructure.

2. Elements of a Semantic Web infrastructure

Figure 1 depicts components that are arguable needed in the developing a national SW infrastructure – according to the experiences to be reported in this paper. The system is based on the domain agnostic W3C Web Standards and Best Practices (on the left below in the figure) of publishing Linked Data4

⁴

Best Practices for Publishing Linked Data: https://www.w3.org/TR/ld-bp/.

[16]. Meta(data) Models for metadata [94] are needed for representing knowledge of different application domains, populated by resources taken from shared Domain Ontologies and Ontology Services for interoperability. In this paper, the term domain ontology refers to typically hierarchical, thesaurus-like knowledge organization systems whose concepts are used to populate property values of (meta)data models. The ontologies should be made openly available and easy to access for interoperability and re-use, based on shared ontology services/libraries; cf. [6,66] for a survey of such systems. In the same vein, data services for publishing LD datasets, preferably using, e.g., open Creative Commons licenses, are needed for making re-use of data possible and easy. Also Applications of Linked Data are part of the infrastructure connecting the system to its end users. For making all this possible, Software Tools are needed for aggregating the distributed heterogeneous data from legacy and other data silos involved, and for extracting and linking (disambiguating) entities and relations from data records and textual descriptions [64]. Also tools for data publishing and analysis are needed, as well as tooling for developing new applications for the end users. Here using the FAIR principles5

⁵

The FAIR principles: https://www.go-fair.org/fair-principles/.

for creating Findable, Accessible, Interoperable, and Re-usable data are recommended. The FAIR principles are compatible with the linked data principles6

⁶

Linked data design issues: https://www.w3.org/DesignIssues/LinkedData.html.

and best practices7

⁷

Data on the Web Best Practices: https://www.w3.org/TR/dwbp/.

of the W3C.

For developing, maintaining, and using the infrastructure in a sustainable way a Human Infrastructure is needed (on the left in Fig. 1), too. This involves, e.g., educating people about the technology, introducing SW courses in university curricula, and production of documentations and learning materials for the community using national languages. In the Finnish case, for example, online materials have been created,8

⁸

See, e.g., the open self-study video lecture course “Linked Data Technologies for Cultural Heritage and Digital Humanities: Introducing the Semantic Web in Video Lectures” at https://seco.cs.aalto.fi/teaching/sw-introduction/.

a Finnish text book about the SW was produced [22], over 40 public seminars and other events have been organized in Finland and beyond,9

⁹

Homepages of the events organized by the Semantic Computing Research Group: https://seco.cs.aalto.fi/events/.

and hackathons10

¹⁰

E.g., as part of the Helsinki DH hackathon series: https://www2.helsinki.fi/en/helsinki-centre-for-digital-humanities/helsinki-digital-humanities-hackathon.

were organized on using the data and tools. The work reported in this paper has also been a core part of the activities of the Helsinki Centre for Digital Humanities HELDIG,11

¹¹

https://heldig.fi

established in 2016 with a particular focus on Digital Humanities research [23,26].

3. Domain ontology infrastructure

In the early 2000’s, the focus in SW research was on ontologies [76], arguably the “silver bullet” of the SW [10]. Accordingly, a series of projects called “FinnONTO” in 2003–2012 were conducted12

¹²

FinnONTO project homepages: https://seco.cs.aalto.fi/projects/finnonto/.

in Finland.

A National Effort The goal in the FinnONTO initiative [45] was to develop a national ontology and content infrastructure, based on W3C standards, that would be cross-domain,13

¹³

In this paper, cross-domain ontology refers to a knowledge organization system encompassing several interlinked application domains.

multilingual (Finnish, Swedish, and English), and openly available. The consortium behind the initiative included finally some 50 companies and public organizations that represented a wide spectrum of functions of the society, including libraries, health organizations, cultural institutions, government, industry, media, and education. By integrating the interests of several functions of the society it was possible to acquire substantial and long-standing funding for the infrastructure and application development work.

The initiative produced 1) metadata models contributing to national standards,14

¹⁴

Such as the Public Recommendation for Geographic Metadata, Ministry of Internal Affairs, http://www.jhs-suositukset.fi/suomi/jhs158.

2) domain ontologies [45,85] to be used for populating the metadata models, 3) a living laboratory called ONKI of public ontology services [85,91], and 4) tools for metadata creation and application development, such as Skosify [77] for SKOS vocabulary quality assessment and the SAHA editor [52,59] for editing RDF repositories. The infrastructure was tested by using it in case studies in different application domains, including e-culture [36,61], e-health [44,78], e-government [74], e-learning [48], and e-commerce [54].

A central goal of FinnONTO was to create an interlinked cloud of national ontologies [11] based on existing thesauri that were already used in different areas of the society. The rationale for this was that metadata available in national databases had already been catalogued using these thesauri, which would make it much easier to develop applications. According to the FinnONTO vision, the ontologies should be served not only through human readable browser interfaces,15

¹⁵

Two human interfaces were created, ONKI.fi (https://onki.fi) for RDFS and ONKI Light for SKOS domain ontologies (https://light.onki.fi).

but also as centrally managed national ontology services using REST APIs. In this way, common functionalities of the services, such as (semantic) autocompletion [35], URI fetching, and query expansion [86], could be shared on a national level, and everybody would get access to up-to-date versions of the ontologies. This kind of collaboration would be cost-efficient on a national level and gradually lead to better interoperability of the data catalogued in different organizations. Availability of the centralized services is needed especially for smaller organizations that do not have much expertise and resources for developing their own web services.

Table 1

Linked domain ontologies of the original KOKO ontology cloud. Contemporary versions of these RDFS ontologies are available transformed automatically into SKOS format at the national ontology service https://finto.fi

Ontology	Application domain	# of concepts
YSO	General upper ontology	27 200
AFO	Agriculture and forestry	7000
JUHO	Government	6300
KAUNO	Literature	5000
KITO	Literary research	850
KTO	Linguistics	900
KULO	Cultural research	1500
LIITO	Economics	3000
MAO	Museum artifacts	6800
MERO	Seafaring	1300
MUSO	Music	1000
PUHO	Military	2000
TAO	Design	3000
TERO	Health	6500
TSR	Working and employment	5100
VALO	Photography	2000

From Thesauri to Ontologies The FinnONTO project transformed 16 key national thesauri used in Finland into light-weight ontologies listed in Table 1. The transformation process was more ambitious than just transforming the traditional standard thesaurus format [2] into an RDF-based model, such as SKOS.16

¹⁶

SKOS Reference: https://www.w3.org/TR/skos-reference/.

The thesauri were developed semantically a bit forward, using the OntoClean methodology [14] and RDFS,17

¹⁷

RDF Schema: https://www.w3.org/TR/rdf-schema/.

in the following ways [45]: 1) Multiple meanings of thesauri terms were disambiguated and relocated in rdfs:subClassOf hierarchies. For example, the concept of child, a unique concept in the underlying General Finnish Thesaurus YSA, can refer to the class of young people, to a family relation type, or a social class (superconcept of street child). These concepts should obviously be in located different places in an RDFS domain ontology, i.e., a hierarchical ontology constructed using the RDFS semantics. The concept was therefore split into several concepts in the corresponding General Finnish Ontology YSO [73]. 2) The thesauri that were transformed did not differentiate whether the standard Broader Term (BT) relation [2] means the part-of or hypernymy relation. This distinction was crafted manually in the ontologies. 3) The rdfs:subClassOf hierarchies were completed: all concepts were given at least one superclass except the root. 3) Inheritance of being an instance over subclass hierarchies was checked as specified by the RDFS semantics, so that the hierarchies could be used for reasoning, e.g., in query expansion and when using faceted search in applications, such as the Sampo portals [27].

Linked Ontology Cloud KOKO The domain ontologies in Table 1 contain similar and related concepts [11]. Especially, the largest ontology YSO ( $27200$ concepts) – transformed from the most used thesaurus YSA of the National Library in Finland – shared lots of concepts with all other ontologies, in some cases more than 50%. This suggested that the ontologies should be linked together using YSO as the top ontology. This idea resulted in creating the Finnish linked ontology cloud called KOKO18

¹⁸

The current version of KOKO is available at the Finto.fi service: https://finto.fi/koko/fi/.

where the top ontology concepts of YSO are refined by subconcepts of interlinked domain specific ontologies [11,45].

While transforming the top ontology YSO from a thesaurus into an ontology its concept labels were also translated into English and the existing partial Swedish translation was completed. This makes it possible to align the ontology with related international domain ontologies. In some other KOKO ontologies translations were already available.

Lessons Learned A key problem to be solved in FinnONTO was that large cross-domain thesauri, especially the General Finnish Thesaurus YSA, could not anymore be maintained easily by its management team. Even if the team included people from different fields, the terminology related to specific areas needed deeper domain specific expertise than was available. Developing the interlinked KOKO ontology cloud mitigates the problem by distributing work on specific concepts to collaborative, domain specific ontology developer teams. However, in this model new problems arise pertaining to maintaining the linked ontology cloud and to coordinating the collaboration network [11]. These new challenges are now being tackled by the Finto collaboration network19

¹⁹

https://www.kansalliskirjasto.fi/fi/content/finto-5

coordinated by the National Library. The FinnONTO initiative pointed out that lots of redundant work had been done in developing the thesauri in Finland as they shared lots similar concepts with each other. In the new, more coordinated KOKO model, redundant work can be better eliminated. To support the ontology alignment work, tools such as MUTU [69] were developed.

A challenge encountered in the ontologization process was that organizing the concepts into class hierarchies cannot in many cases represent correctly the meaning of the original terms that can be complex and fuzzy. The world cannot be represented fully using ontologies and there can be several ways in which this can be done. In spite of such challenges, the idea of adding more semantics needed for application development seems to be a better option than continuing using the original thesauri, whose semantics were too vague from a sofware development point of view in applications, such the semantic portals to be discussed in this paper. A strategic choice made in FinnONTO was to follow the wisdom articulated by Jim Hendler already in the late 90’s in the SHOE project:20

²⁰

https://www.cs.rpi.edu/~hendler/LittleSemanticsWeb.html

A little semantics goes a long way. In our case, the traditional thesauri semantics [2] were refined only a little using RDFS (as explained above, using and completing, e.g., class hierarchies) for interoperability and to help development of web applications. However, already this was a handful of work, as thousands of terms in the thesauri had to be manually checked and refined [73].

A mundane challenge of developing large domain ontologies, at least in Finland, is how to convince the funding organizations, year after year, that this never-ending work should be supported on a regular basis, not only as separate short-time projects. This is possible if the benefits of using ontologies in practical applications can be demonstrated. In our case, it took some ten years of project-based work before the KOKO ontology infrastructure and the current Finto.fi services could be funded in a more sustainable way by two Finnish ministries. The strategy taken in FinnONTO was to move forward in baby steps, and after each step show a demonstrator on how the ontologies can be applied in practise for creating something useful.

The idea of creating a “living laboratory” of ONKI ontology services [85,91] on the Web turned out to be important for deploying the infrastructure. The participating FinnONTO organizations were supported by the project in connecting their legacy systems to the APIs of ONKI for testing and evaluating the services. Finally, the “point of no return” was reached where pulling off the plug of the services was not an option anymore as the number of ONKI API users were counted already in hundreds.

The FinnONTO project series 2003–2012 started with a smallish one-year project, but eventually grew into a national effort of substantial size on the Finnish scale with dozens of funding organizations involved. A reason for this was that in addition to public organizations, such as museums, libraries, and archives, also companies got interested in the technology, which convinced the main funding organization Tekes (called today Business Finland) that something useful and of monetary value is happening related to semantic web technologies. It is usually easier to get funding for technology development than for research in humanities.

The KOKO ontologies are based on keyword thesauri whose terms usually correspond to the classes. FinnONTO worked also on various “instance-based” ontologies, such as national geogazetteers, person and organization registries, biological taxonomies of species [87,88], and nomenclatures and terminologies of medicine [78], such as Medical Subject Headings MESH.21

²¹

https://finto.fi/mesh/fi/

Creating a national ontology and data infrastructure is a never-ending job and research goes on today, e.g., in the Linked Open Data Infrastructure for Digital Humanities initiative22

²²

https://seco.cs.aalto.fi/projects/lodi4dh/

[24] that is part of the national FIN-CLARIAH research infrastructure for Digital Humanities.23

²³

https://seco.cs.aalto.fi/projects/fin-clariah/

Here the idea is to combine – on a national level as in the CLARIAH initiative in the Netherlands – the work related to the pan-European infrastructure CLARIN,24

²⁴

https://clarin.eu

the research infrastructure for language as social and cultural data, and DARIAH,25

²⁵

https://dariah.eu

the infrastructure for arts & humanities scholars.

When developing ontology-based applications in FinnONTO, much of the time of the developers was “waisted” in cleaning and aligning the data from different organizations for interoperability. Obviously, it would be more cost-efficient do this work already when cataloging the data using ontology services. This would also enhance the quality of the linked data, which is a critical problem [93] on the SW. The local cataloguers know best their own data and should have the best interest in quality of their data. The motto for the FinnONTO work was therefore taken from a wisdom of Albert Einstein: Intellectuals solve problems – geniuses prevent them; a key goal of FinnONTO was to prevent interoperability problems rather than to solve them afterwards when the damage has already been done in cataloguing [21].

A major outcome of FinnONTO was the ONKI ontology server with its ontologies [91] that were published first in 2008. As a next step, the ONKI Light service26

²⁶

https://seco.cs.aalto.fi/services/onkilight/

[79] was developed and deployed in 2014 [80] by the National Library of Finland as the national Finto.fi service.27

²⁷

Available at: https://finto.fi

ONKI Light finally evolved into the open source Skosmos tool28

²⁸

https://skosmos.org/

[81] in use in several other organizations in Finland and internationally.29

²⁹

For a list of international services, see https://www.kiwi.fi/display/Finto/Skosmos-ohjelmisto0.

ONKI Light was based on a SPARQL endpoint. The idea was to separate the data service fully from the user interface, and use only SPARQL to access the data. This idea turned later useful when developing the Sampo model [27] and Sampo-UI tool [46,70] for semantic portals to be discussed later in this article.

The Finto.fi service has grown into a popular national open service. In 2019 it was used by $280000$ different users and its APIs were called 32 million times. The users include, e.g., museums, whose cataloging system get their keywords with URI identifiers from Finto. These developments suggest that the fundamental ideas of FinnONTO are feasible; they have actually made a paradigm change in Finland in developing and using linked light-weight ontologies on a national level instead of thesauri.

4. From 5-star to 7-star Linked Data deployment scheme

The SW infrastructure model of Fig. 1 includes a platform for publishing datasets and (re-)using them via web services. A key component in LD publishing is the SPRAQL endpoint, but the platform should also support other functions [16]. The Linked Data Finland service LDF.fi30

³⁰

https://ldf.fi

[43] was therefore developed in the “Linked Data Finland” follow-up projects31

³¹

https://seco.cs.aalto.fi/projects/ldf/

of FinnONTO.

LDF.fi has two user-groups: 1) For application developers, LDF.fi provides SPARQL endpoints and a suite of standard Linked Data (LD) services, including content negotiation, APIs for downloading datasets, LD browsing and editing, and additional tools for, e.g., data documentation and visualization. 2) For data publishers, the idea is to support and automate the data publishing process in the following way: The publisher creates a service description of the dataset and its schemas, using an extended version of the W3C Service Description recommendation.32

³²

http://www.w3.org/TR/sparql11-service-description/

Based on such metadata, LDF.fi then 1) automatically sets up the technical services, 2) generates a dataset “homepage” that explains the dataset, schemas, and 3) provides additional related services for querying, documenting, inspecting, and validating the data. LDF.fi is used primarily for reading RDF data by SPAQRQL queries, not for writing, although also this could be done using the SPARQL endpoint. The general Linked Data Platform recommendation33

³³

Linked Data Platform 1.0: https://www.w3.org/TR/ldp/.

that was under development at the same time has not been used in LDF.fi.

Linked data publications on the SW are typically evaluated with the W3C “5-star” deployment scheme,34

³⁴

https://www.w3.org/community/webize/2014/01/17/what-is-5-star-linked-data/

using a quality scale analogous to evaluating hotels. In LDF.fi, the 5-star model is extended to a 7-star model: there are nowadays also a few 7-star hotels around.35

³⁵

Such as the Burj Al Arab in United Arab Emirates.

The 6th star is given to a data publication if it includes not only the 5-star data but also the schemas of the data with documentation. This makes re-use of data easier. The 7th star is given to a data publication, if the publication includes some kind of evaluation that the data actually conforms to the provided schemas using, e.g., SHACL36

³⁶

https://www.w3.org/TR/shacl/

or ShEx37

³⁷

https://shex.io/

[53]. The idea here is to encourage publishers to publish high quality data as data quality of LD is a severe issue on the SW.

Schemas can be documented automatically in LDF.fi for the human reader using a schema documentation generator, in our case using SpecGen38

³⁸

https://bitbucket.org/wikier/specgen/wiki/Home

and LODE.39

³⁹

https://essepuntato.it/lode/

Datasets in the LD world often use schemas (vocabularies) for which definitions or descriptions are not available, but are embedded in the data itself. In order to find out how schemas are actually used in a dataset, including both published and unpublished schemas, a service vocab.at40

⁴⁰

http://vocab.at

was created that analyzes a given dataset from this perspective and creates an HTML document that lists, e.g., statistics of vocabulary usage and raises up issues detected if an IRI is not dereferenceable. The input for vocab.at is either an RDF file, a SPARQL endpoint, or an HTML page with embedded RDFa markup.

LDF.fi is implemented by a combination of the Fuseki SPARQL server41

⁴¹

https://jena.apache.org/documentation/fuseki2/

for storing the primary data and a Varnish Cache web application accelerator42

⁴²

https://varnish-cache.org

for routing URIs, content negotiation, and caching. For deployment of applications with a data service (cf., e.g., the MMM system [31]) a microservice architecture with Docker containers43

⁴³

https://www.docker.com

is used. Each individual component (the application, Varnish, and Fuseki) is run in its own dedicated container, making the deployment of the services easy due to installation of software dependencies in isolated environments. This enhances the portability of the services. The server environment of LDF.fi is provided by the CSC – IT Center for Science, a company of the Ministry of Education and Culture of Finland providing computational infrastructures for the national universities in Finland.

Lessons learned The Linked Data Finland platform has turned out to be useful for data-analytic research purposes and in developing applications (cf. Section 5). LDf.fi has been used for publishing some 100 linked datasets. Many of them are in use in semantic portal applications and via SPARQL querying combined with query editing and scripting tools using the open CC BY 4.0 license. Some datasets are used only internally in related research projects, and for some datasets licensing policy of the data owners prohibits open use. LDF.fi hosts several instance-based ontologies, too, such as an RDF-based version of the ca. $800000$ official Finnish geographical places based on data of the National Survey.

The LDF.fi service is still maintained by the Aalto University and University of Helsinki that developed it on an academic project basis, but with the hope that some day it will be deployed and be maintained in a more sustainable way – this is at least what happened to the related ONKI/Finto ontology services. A step towards this is that in 2020 the idea of providing Linked Open Data services on a national level using LDF.fi was accepted on the new research infrastructure roadmap of the Academy of Finland as part of the larger FIN-CLARIAH infrastructure.

5. Sampo model: Applying the Semantic Web infrastructure

Table 2

Sampo model principles P1–P6

P1	Support collaborative data creation and publishing
P2	Use a shared open ontology infrastructure
P3	Make clear distinction between the LOD service and the user interface (UI)
P4	Provide multiple perspectives to the same data
P5	Standardize portal usage by a simple filter-analyze two-step cycle
P6	Support data analysis and knowledge discovery in addition to data exploration

When developing the Finnish SW infrastructure, applications that test and demonstrate its usability were constantly developed. This work evolved gradually into a set of principles for developing LOD services and semantic portals on top of them, called the Sampo Model, and the Sampo Series of LOD services and semantic portals [27]. The novelty of the Sampo model44

⁴⁴

The model is called “Sampo” according to the Finnish epic Kalevala, where Sampo is a mythical machine giving riches and fortune to its holder, a kind of ancient metaphor of technology according to the most common interpretation of the concept.

lays in its attempt to formulate a set of re-usable design principles or guidelines for creating LOD services and semantic portals, especially for Cultural Heritage applications and Digital Humanities research [12]. Based on six principles, the model is a kind of consolidated approach for creating LOD services and semantic portals, something that the field of the Semantic Web is arguably still largely missing [18].

The Sampo Model is an informal collection of principles for LOD publishing and designing semantic portals listed in Table 2. Principles P1–P3 can be seen as a foundation for developing data services; P4–P6 are related to creating semantic portals. The model is based on the idea of collaborative content creation (P1). The data is aggregated from local data silos into a global service, based on a shared ontology and publishing infrastructure (P2). The local data are harmonized and enriched with each other by linking and reasoning. In this model everybody can arguably win, including the data publishers by enriched data and shared publishing infra, and the end users by richer global content and services. The model argues for the idea of separating the underlying Linked Data service completely from the user interface via a SPARQL API (P3). This arguable simplifies the portal architecture and the data service can be opened for data analysis research. For example, YASGUI45

⁴⁵

https://yasgui.triply.cc

 [71] interface for SPARQL querying and visualizing the results can be used, or Python scripting in Google Colab46

⁴⁶

https://colab.research.google.com/notebooks/intro.ipynb

and Jupyter notebooks47

⁴⁷

https://jupyter.org

[83].

The general idea of principles P4–P6 is to “standardize” the UI logic so that the portals are easier to use for the end users and for the programmers to develop, as demonstrated in the Sampo-UI framework tool [46,70]. Principle P4 articulates the idea of providing different thematic application perspectives by re-using the data service. The application perspectives can be provided on the landing page of the Sampo portal system or be completely separate applications by third parties. According to P5 the application perspectives can be used by a two-step cycle for research: First the focus of interest, the target group, is filtered out using faceted semantic search [39,84,90]. Second, the target group is visualized or analyzed by using ready-to-use data analytic tools of the application perspectives. Finally, the Sampo model aims not only at data publishing with search and data exploration [63] but also to data analysis and knowledge discovery with seamlessly integrated tooling for finding, analysing, and even solving research problems in interactive ways (P6) [25].

The Sampo model principles are compatible with the FAIR principles for creating Findable, Accessible, Interoperable, and Re-usable data,48

⁴⁸

https://www.go-fair.org/fair-principles/

but were developed in the context of publishing and using LOD. The principles P1–P6 can be used directly for creating semantic portals. However, its is also possible to apply them first to create an application domain specific framework and reuse it for developing different related application instances, which is arguably cost-efficient. This idea of re-using domain-specific “Sampo frameworks” has been demonstrated in the LetterSampo system for publishing epistolary data [34] and in FindSampo for archaeological data [38].

The Sampo model has evolved gradually in 2002–2023 via lessons learned in developing a series of semantic portals and LOD services, starting from MuseumFinland – Finnish Museums on the Semantic Web49

⁴⁹

This application at https://museosuomi.fi got the Semantic Web Challenge Award at the ISWC 2004 conference.

(online since 2004) [37], CultureSampo – Finnish Culture on the Semantic Web 2.050

⁵⁰

https://seco.cs.aalto.fi/applications/kulttuurisampo/

(online since 2009) [36,61], and BookSampo51

⁵¹

https://kirjasampo.fi

(online since 2011 with some 1.6 million annual users today) [58]. They demonstrated how CH content of dozens of different kinds, both tangible and intangible CH content, can enrich each other. WarSampo – Finnish World War II on the Semantic Web52

⁵²

https://seco.cs.aalto.fi/projects/sotasampo/en/

(online since 2015 with several new perspectives published in 2016–2019) [30], an example of applying LOD to publishing and studying Military History [28], is a popular Finnish service that has had thus far some 1.2 million users. A key idea in WarSampo is to reassemble the life stories of the World War II soldiers based on data linking from different data sources. This biographical and prosopographical idea was a source of inspiration for several later biographical applications [29], including BiographySampo – Biographies on the Semantic Web53

⁵³

https://seco.cs.aalto.fi/projects/biografiasampo/en/

(online since 2018) [33], Norssit Alumni [32], U.S. Congress Prosopographer [65], and AcademySampo54

⁵⁴

https://seco.cs.aalto.fi/projects/akatemiasampo/en/

(online since 2021) [56]. NameSampo [47] publishes data about over 2 million place names and places in Finland with old maps. The NameSampo project developed, based on the SPARQL Faceter tool [50] used in many earlier Sampos, the first version of the Sampo-UI framework [46] that has been used after this in all Sampos. It supports implementation of principles P4–P6 from an UI point of view. Sampo-UI has been re-used, e.g., in the portal Mapping Manuscript Migrations (MMM)55

⁵⁵

https://seco.cs.aalto.fi/projects/mmm/

(online since 2020) [31,49] based on metadata about some $220000$ pre-modern manuscripts from the University of Oxford (U.K.), Schoenberg Institute (U.S.), and IRHT (France), in FindSampo56

⁵⁶

https://seco.cs.aalto.fi/projects/sualt/

[38] (online since 2021) for supporting archaeology from a citizen science and metal detectorists’ perspectives. LetterSampo57

⁵⁷

https://seco.cs.aalto.fi/projects/rrl/

[89] is based on early modern epistolary metadata aggregated in the Early Modern Letters Online (EMLO) service58

⁵⁸

http://emlo.bodleian.ox.ac.uk

at the Oxford University, the CKCC corpus underlying ePistolarium59

⁵⁹

http://ckcc.huygens.knaw.nl/epistolarium/

of the Huygens Institute in the Netherlands, and correspSearch60

⁶⁰

https://correspsearch.net

service of the Berlin-Brandenburg Academy of Sciences. During the spring of 2023 two new Sampos of particular societal impact were released: LawSampo61

⁶¹

https://seco.cs.aalto.fi/projects/lawlod/

[42] publishes Finnish legislation and case law based on data from the Ministry of Justice in Finland. ParliamentSampo62

⁶²

https://seco.cs.aalto.fi/projects/semparl/en/

[40,41] publishes LOD of the Parliament of Finland (1907–2023). Its LOD service includes knowledge graphs about nearly one million Parliamentary debate speeches [75] as well as data about the organizational structures and events of the Parliament, politicians, and their networks [57] in 1907–2023. In OperaSampo63

⁶³

https://oopperasampo.fi

(online since 2023) [1] data about thousands of historical music theater and opera performances in Finland (1830–1960) are published.

Developing these systems in use in a university research group64

⁶⁴

Semantic Computing Research Group (SeCo): https://seco.cs.aalto.fi

would not have been possible without re-using the elements of the national infrastructure (Fig. 1) and developing them further step-by-step in a systematic way. The Sampo-UI framework has turned out to be very effective tool in developing the portal user interfaces, and it has been used also by some external developers. In some cases, a first test demonstration of a new Sampo has been developed in a few weeks, but this depends on the case and quality of the data available. In cases like ParliamentSampo several years of development was needed for a finished in-use version on the Web. Natural language processing (NLP) techniques have been another important category of tools in later Sampos, such as LawSampo and ParliamentSampo, where lots of data have been available only in unstructured textual form. During our work, external NLP tools were re-used and new ones developed for named entity recognition (NER) and linking (NEL), for automatic annotation of keywords, and for topical classification of texts [55,82]. For LawSampo also a pseudonymization tool called Anoppi was created [67] as personal information in court decisions cannot be disclosed on the Web.

Data about all over 20 Sampo portals, including links, videos, publications, and further information are available on the Sampo portals homepage.65

⁶⁵

Sampo portals’ homepage: https://seco.cs.aalto.fi/applications/sampo/.

6. Discussion

This paper addressed challenges of extending the SW layer cake model for creating ontology and LOD infrastructures on national and domain specific levels. Lessons learned in developing Finnish ontology and LOD services 2002–2023 for practical applications were discussed. This work has utilized methods of design science [17,62,68] and action research [8], where the idea is to design artifacts, evaluate their value and utility, and to provide improvements in solutions. Rather than creating theoretical knowledge, design science applies knowledge. In this paper, infrastructure elements were designed, implemented, and applied to create the Sampo series of data services and portals as proof-of-concepts. They have had up to millions of end users, which suggests feasibility of the national infrastructures presented. The line of R&D presented is unique in its focus on different domains on a national level, longevity, and series of applications in use on the Semantic Web. Our general strategy has been to develop useful proof-of-concept prototypes and to publish them openly on the Web for everyone to use. The data owners and stake holders, such as memory organizations, saw this as an opportunity to develop their own systems, started to use the services and applications, and in many cases the point of no return has been reached.

In contrast to current related ontology library systems [7,66] that typically focus on particular application domains, ONKI and Finto aimed at being a cross-domain ontology service on a national level. For example, the BioPortal [72] of Stanford University is focused on publishing biomedical ontologies. There are lots of LOD services and SPARQL endpoints around.66

⁶⁶

https://www.w3.org/wiki/SparqlEndpoints

The novelty of the LDF.fi service lays on its 7-star model and the idea of integrating the data service with various online tools as well as learning materials to support data re-use. Instead of being a focused data service for particular data, such as DBpedia for Wikipedias, the LDF.fi platform aims at being a cross-domain platform of datasets on a national level. The main application area of the presented infrastructure has been Cultural Heritage and Digital Humanities [12], although also systems for, e.g., e-health, e-government, and e-learning were developed.

Fig. 2.

Timeline illustrating the development of the Semantic Web and work reported in this paper.

Figure 2 illustrates the work reported in this paper on a timeline and in relation to some global developments of the Semantic Web. The development of the SW was boosted by the seminal article in Scientific American in May 2011 by Tim Berners-Lee, James Hendler, and Ora Lassila [3], and this inspired us to organize the conference Semantic Web Kick-off in Finland in the autumn of 2001 together with researchers of the W3C, the Finnish Artificial Intelligence Society, and some other organizations [20]. This event initiated SW research in Finland. During the past 20 years, the SW has evolved in phases [18] (cf. top of the figure) with a focus first on ontologies [76], then on Linked (Open) Data [16], and today on Knowledge Graphs (KG) [15]. In the same vein, our work on infrastructure first focused on ontologies and ontology services (ONKI.fi and Finto.fi) and then on Linked Data and data model services. The Sampo series of applications reflects this development by showing a shift of focus in three generations [25] from data publishing, based on shared ontologies and metadata vocabularies67

⁶⁷

https://lov.linkeddata.es/dataset/lov/

(1. generation portals), to supporting the end-users of KGs with seamlessly integrated data-analytic tools and visualizations needed in areas such as Digital Humanities (2. generation systems). However, the series has also taken first steps forward towards 3. generation portals that can solve problems for the end users based on knowledge discovery, Artificial Intelligence, and computational creativity. There are lots of related works pertaining to the different Sampo systems; discussing them is beyond the scope of this paper, but pointers to such works can be found in the referenced research papers. The bottom line of Fig. 2 depicts introduction of some important standards and important events of the SW for a context.

The ontologies and linked open datasets discussed in this paper constitute together a kind of national Linked Open Data Cloud. The idea is that as new ontologies and applications with new datasets are developed, the open RDF data already available in the infrastructure, say ontologies of places and historical people, can be reused and refined gradually better and better. This applies also to the open source tools, such as the Sampo-UI framework, that has been re-used and extended in all Sampo systems after publishing NameSampo in 2008. A goal of the current national FIN-CLARIAH infrastructure initiative is to foster this development.

The experiences reported in this paper indicate that creating and using a national semantic web infrastructure is useful from the data producers’ and data users’ points of view. However, creating and using linked data has its own challenges, too. More collaboration and agreements on data models and ontologies are needed for interoperability between the data producers, which complicates the publication process. Integration of SW technologies with legacy systems may be challenging, and there is lack of IT personnel competent in using SW technologies and tools. Creating linked data manually is costly but automatic methods may not be available and automation lowers data quality. Using structured semantic data and making the knowledge structures explicit to the end user in the UI calls for new kind of digital data literacy and source criticism68

⁶⁸

https://ranke2.uni.lu/define-dsc/#%20,%20Universit%C3%A9%20du%20Luxembourg

from the end user [51,60]. What the underlying data actually means is not always clear and issues of Big Data quality, such as completeness, veracity, skewness, uncertainty, fuzziness, and errors of data arise. However, in spite of the challenges, enriching data carefully with semantics, with one way or another, is in my mind a way ahead towards creating a more and more intelligent Web in a cost-efficient way. In contrast using “black box” language model-based systems and deep machine learning, such as Chat GPT, the SW makes the data on the Web explicit, transparent, and well-defined, and the already structured curated data in databases can be utilized. This facilitates creation of explainable “white box” AI systems [25,92].