Linked Data is attracting increasing interest from the Swiss public administration. The Swiss Federal Archives are playing a leading role in this respect by investing significantly in the deployment of an infrastructure for publishing data in LD. This approach has enabled the institution to acquire in-depth knowledge on the subject and to consider integrating LD into its core applications and into the services it offers to the public.
aLOD related activities The archives participating in aLOD activities have set themselves the following goals: 1. To examine the opportunities for LD to achieve the mission of archival institutions. 2. Based on real descriptive metadata sets from their respective AIS, to transform and unify these metadata into LD datasets. 3. In doing so, to formulate "best practices" for transforming existing inventories (metadata) into LD. 4. To communicate and disseminate the achievements of the project within the archival community and the internal users, but also beyond, for example within the community of researchers in digital humanities, and also to exchange with the actors who contribute to the implementation of LD technologies (GLAM and beyond). 5. Demonstrate the potential for third party reuse of descriptive metadata in LD when made freely available (OGD), as for example in the context of hackathons on cultural data.
The different archives exported data from their AIS in the form of csv files that were
then converted into LD using an ad hoc data model, since the RiC-O data model [
The contents of the inventories had very different levels of detail from one institution to another. The data model therefore had to be enriched as new datasets were integrated.
The language of these contents was different: French or German in this case. Fortunately, this aspect is well handled with language tags in RDF2.
The dates had variable numerical formats or even were in textual form. This is one of the aspects that took the longest time to be dealt with, without resulting in a clean and reusable solution.
For each institution, a data export procedure had to be put in place. Even for institutions that used the same AIS, it was not possible to have a generic solution, as the content structures were quite different.
The data conversion produces triples like those associated with the AFS record with the signature "B0#1000/1483#3792*" (see Fig. 1). Thanks to this uniform representation of the data from the different archives and to the fact that these Linked Data are directly accessible on the web via LINDAS and its SPARQL interface, it has been possible to have an experimental prototype for the visualisation of all these data (see Fig. 2). This representation includes a histogram of the number of records per date, which is unusual in archival web portals but could be useful to identify the density of information over time on a specific subject. 2 Resource Description Framework: a formal model to define graph structures. Fig. 1. Example of an entity from the AIS of the AFS, of type "File", converted to LD with the ad hoc data model used in 2015.
LINDAS as a linked data hosting infrastructure has been enhanced since its first release to become a productive infrastructure. This enhancement was carried out between 2017 and 2020. Its general structure is described schematically below (see Fig. 3). In the centre, there are several triplestores to allow testing, integration and finally production of new datasets. Data conversion can be a recurring or a one-off process. In any case, an ETL pipeline is implemented, the execution of which can be scheduled according to the updating of the source data.
To ease the definition of these conversion processes, the Data Cube Creator tool,
specialised in the conversion of OLAP cubes [
To complete the infrastructure, the graphical visualisation of data hosted in LINDAS
can be parameterised using the Visualize tool [
We believe that LD is the optimal solution for publishing data and making it accessible on the web. The question we asked ourselves in relation to our core activities as an archive is whether LD and more specifically the RDF model could be used as the central database of an AIS. In 2018 and 2019, two studies3 were conducted by research institutes to verify certain aspects of this technology in relation to our own issues.
It appeared from this work that there are suppliers of triplestores able to deliver solutions that are perfectly suited to our needs. Thus, Stardog version 5.2 allowed us to build a graph of 10B triples by reading files of 100M triples with an average and stable execution time of 20 min. per file. This amount of data is much more than what we estimate we will eventually have to manage in our AIS: 100-500M triples.
Updates are crucial operations that are implemented by Delete and Insert functions. In our test, 1M updates were performed in 12 sec. on average. And finally SPARQL queries of different complexity combined with Insert, all at different frequencies, have had sub-second response times.
Therefore, we are confident that, if the triplestore is installed on suitable servers, this technology will be performing well as the core database of an AIS.
Another issue that has been studied is whether the RDF model is as powerful as
Property Graphs (e.g. Neo4j [
In our opinion, there is no reason why an AIS should not be developed with a triplestore at its core as a central database. 5
If LD can be implemented at the core of an AIS, it can also have other roles. Here are two areas we are considering working on in the coming years. 5.1
By publishing datasets in LD, public administrations take a first step towards publishing
entire databases for public reuse. However, archives also hold databases in their archive
holdings, ideally in SIARD format [
RiC-O in its current version 0.2 is a very promising proposal that still needs to be tested in the very different contexts of Swiss archives. To this end, LINDAS and its data conversion environment will allow us to test the conversion of the descriptive metadata of 3 These reports have not been published but the author can provide you with a copy if desired. our inventories according to the RiC-O standard. Only then will it be possible to identify possible gaps in the model and to establish best practices in the way of proceeding with this conversion task.