Modelling medieval manuscripts in Wikidata

The undertaking described in this blog post was supported by an NFDI4Memory FAIR Data Fellowship at the Herzog August Bibliothek in March 2026. I was supported by Daniela Schulz and Torsten Schaßan on the side of the Herzog August Bibliothek and by Camillo Carlo Pellizzari di San Girolamo and Jonathan Groß on the side of the Wikidata Manuscripts project. Camillo, in particular, provided vital know-how and the guidance needed to work with OpenRefine and QuickStatements.

The stimulus that led me to apply for the NFDI4Memory FAIR Data Fellowship in 2025 came from two directions. In 2021, I completed a three-year Innovating Knowledge postdoctoral project at the Huygens Institute in Amsterdam, during which I surveyed all known surviving early medieval manuscripts of the most important Latin medieval encyclopaedia, the Etymologies (Etymologiae) of the Spanish bishop Isidore of Seville (d. 636). This encyclopaedia was truly a medieval bestseller and the essential go-to for generations of Latin-reading people until the early modern times. It, thus, should not come as a surprise that it was copied vigorously in the period of the handwritten book, so that more than 1,400 medieval copies survive. A printed handlist published in 1966, based on research carried out in the first half of the twentieth century (Fig. 1), suggested that approximately 300 manuscripts of this quintessential medieval reference work survived from before the year 1000, a remarkable testament to the importance of this encyclopaedia. Through re-evaluation of previously collected material, meticulous review of manuscript resources published since the 1940s, and novel discoveries, my project brought this number to more than 500 early medieval copies, showing how influential the Etymologies was in the first three and a half centuries of its circulation and readership.

Fig. 1: The title page and a sample page of a 1966 printed handlist of the medieval manuscripts of the Etymologies identified by August Eduard Anspach in the 1940s. (Scan by the author)

In contrast to the earlier twentieth-century researchers, I wanted the insights I collected in the twenty-first century to be published not as a catalogue, but rather as data that can be shared, reused, and are suitable for further machine-based processing and analysis (Fig. 2). I therefore opted for a double publication pipeline: an online database, in which users could access information about individual manuscripts and discover their mutual connections via a graphic user interface, and a data spreadsheet, which could be reused for further research and was ideally suited for analysis. Nevertheless, since the end of the project in 2021, I have been wondering how to make the data I collected during my project, and which continued to grow in the following years as manuscripts became available online, new manuscripts surfaced in public and private collections, previously unknown connections between manuscripts came to light, and further research involving them was conducted, available in a more findable, accessible, interoperable, and reusable (or FAIR) fashion. The quest became more urgent as it became clear that the Huygens Institute cannot maintain the project database, the main channel through which the researchers and the interested public accessed my research data, in the long term.

Fig. 2: A screenshot of the original Innovating Knowledge dataset as an Excel spreadsheet.

The second important impulse came in 2024, when I was contacted by Jonathan Groß from the Wikidata Manuscripts project (Fig. 3). It was exciting to learn that there was an initiative that now laboured to ingest structured and formalised data about medieval manuscripts into Wikidata, and which could therefore perfectly accommodate the manuscript data I collected during my Innovating Knowledge project. Injecting information about the manuscripts of the Etymologies into Wikidata, moreover, meant making them part of the Linked Open Data (LOD) ecosystem, a significant step forward in their FAIRification.

Fig. 3: The screenshot of the Wikidata Manuscripts project main page, https://www.wikidata.org/wiki/Wikidata:WikiProject_Manuscripts.

The point of departure for my undertaking was an Excel datasheet I published in early 2026, which reflected the status of my manuscript survey in 2025. However, it became clear almost immediately that modelling medieval manuscripts as LOD entities is no straightforward task. Rather, the representation of such entities with complex histories and lives in the LOD ecosystem requires answering important questions about what a medieval manuscript is and how it relates to the modern material objects that populate the special collections of heritage institutions. The manuscripts produced by the hands of medieval scribes too often became fragmented, damaged, or even destroyed. Some of them could end up scattered across several modern heritage institutions, so that it requires a certain ingenuity to piece them together into a single medieval whole. In other cases, several medieval manuscripts might have been combined together. Medieval fragments are sometimes bound together into folders. Moreover, once their original purpose was exhausted, medieval manuscripts could have been given a second life as the maculature of other manuscripts and printed books, to which they may continue to be bound today. In short, the manuscripts as they were designed and circulated in the Middle Ages (i.e., the medieval production units) and the items currently preserved, described, and classified in modern heritage collections (i.e., modern shelfmarks) do not perfectly match.

As a first step in my goal to import my 500+ early medieval manuscripts of the Etymologies into Wikidata, I, therefore, had to create two separate but interconnected datasets: the first consisting of 493 items (i.e., rows in a spreadsheet) corresponding to medieval production units, and the second consisting of 561 items corresponding to modern shelfmarks, in which these 493 production units occur. The splitting of my originally single dataset into two spreadsheets allowed me to more clearly distinguish and model data that properly belong to one object type or the other. For example, the production date, location of production and other locations through which the manuscript passed in the Middle Ages, script, and dimensions of manuscripts are properly aligned with medieval production units, while the current geographical location of preservation, URLs of digital surrogates, IIIF manifests, and references to extant descriptions and online resources, which tend to reference modern shelfmarks, should be tied to the latter.

Accordingly, the import into Wikidata had to be done in two stages. First, the dataset corresponding to modern shelfmarks was imported into Wikidata or, in the few instances where items already existed for certain shelfmarks, these were reconciled (meaning that specific rows of my datasheet were matched with entities extant in Wikidata using a special service, on which more below). This initial import provided the foundation on which the medieval production units, with their more complex data, could be laid. In this manner, it was possible to model even complex material situations. To provide but one example, I can present a remarkable set of seven fragments, which are currently divided between four public and two private collections (as one public institution holds two fragments with different shelfmarks), and represent the only remaining remnants of a ninth-century Breton manuscript of the Etymologies containing Latin and Old Breton glosses. What remains of this once splendid codex, therefore, amounts to seven different modern shelfmarks, which had to be first imported into Wikidata separately using the ontological category of manuscript and only then could their relationship as parts of a single medieval production unit be properly expressed in Wikidata using the ontological category of codicological unit (Fig. 4).

Fig. 4: A complex case of a medieval production unit (codicological unit according to the Wikidata ontology) constituted by seven different shelfmarks (here the first two are visible).

While I originally intended to import most of my dataset into Wikidata, during my FAIR Data Fellowship, it became evident that the input of some data categories would require more labour than the fellowship’s limited duration allowed. In particular, the original Innovating Knowledge dataset contains detailed information about texts other than the Etymologies transmitted by the 493 medieval manuscripts included. However, a test reconciliation round showed that only about 30% of these texts already appeared in Wikidata. This meant that, if the content of each manuscript were to be described, I would first have to create a large number of new items in Wikidata, whether manually or semi-automatically, using the same strategies as for the shelfmarks and medieval production units. Of the data pertaining to the contents of the manuscripts, I, therefore, decided to restrict my import to the Etymologies alone, including the information about the pages or folia on which this text occurred in each manuscript.

Fig. 5: A screenshot of my project data in OpenRefine in the process of reconciliation.

The Wikidata import was carried out using two tools: OpenRefine, an open-source desktop program for data wrangling, and QuickStatements, an online tool for pushing large-scale edits into Wikidata.

OpenRefine was used for initial data cleaning and manipulation, such as splitting of columns to generate new data fields relevant for the Wikidata manuscript data model designed by the Wikidata Manuscripts project. OpenRefine also supports data reconciliation with Wikidata, a prerequisite for any import. Using the OpenRefine reconciliation service allowed me, thus, to establish which of my modern shelfmarks already existed as items in Wikidata (and therefore to which my data needs to be merely added), and which should be newly created in Wikidata before my main import of production units could be undertaken (Fig. 5). In addition, in order to perform the import of modern shelfmarks, it was also important to prepare and reconcile data about their holding institutions. Most larger manuscript-owning institutions, such as the Vatican Library and the National Library in Paris, already had corresponding entities in Wikidata, and I could simply use the OpenRefine reconciliation service to express the relationship between my shelfmarks and these institutions. However, some of the small manuscript-holding institutions had not been present in Wikidata and entities corresponding to them were created manually first so that the shelfmarks could be imported correctly.

Fig. 6: A screenshot of the Wikibase schema used to import medieval production units from the Innovating Knowledge dataset into Wikidata.

Once all the relevant columns in my data sheet were reconciled, I used OpenRefine to create import schemas for both modern shelfmarks and medieval production units (Fig. 6). Such a schema clarifies how certain columns in my datasheets have to be parsed into the Wikidata to conform to the data model for manuscript description suggested by the Wikidata Manuscripts project. Passing my dataset through the schema, I exported the output to a QuickStatements .tsv file and performed the import in the QuickStatements environment. While OpenRefine also allows for direct import through its interface, using QuickStatements had particular advantages, such as a more transparent error log, which helped me detect and clear problematic values in my dataset (Fig. 7). For example, it revealed incorrectly formatted URLs of some shelfmarks and a medieval production unit label which had to be shortened to be accepted by Wikidata (which allows maximum 250 characters in each of its fields).

Fig. 7: Error log generated by QuickStatements 2.0 that indicates a problem with one of the medieval production units during the last round of imports. In this case, the issue was caused by using the ‘somevalue’ descriptor in one of the columns and resolved by using QuickStatements 3.0.

The main challenge I faced during my fellowship was the technical issues affecting the Wikidata reconciliation service since the beginning of 2026. This issue, which appears to have its roots in changes to API security on the side of Wikidata, meant that at various times the Wikidata reconciliation service native to OpenRefine would fail, and I had to wait several hours or days before I could resume my work. In spite of this minor technical hiccup, Wikidata is a crucial Open Data platform, which has great potential to benefit researchers of medieval manuscripts. Due to the robust support provided by its user community, the existence of an ecosystem of aligned tools, user-friendly architecture, commitment to Open Science, and international character, it is a vital supplement to more traditional institution-based research platforms, which are often tailor-made and whose long-term sustainability is not guaranteed. Manuscript researchers should therefore seriously consider integrating Wikidata into their project publication strategies.

The final versions of the two datasets imported into Wikidata can be found on Zenodo, at https://zenodo.org/records/20249521.

Dieser Beitrag wurde redaktionell betreut durch Hannah Busch und Karoline Döring.

Evina Steinova

I am currently a VENI postdoc at Huygens ING, an institute of the Dutch Royal Academy of Arts and Sciences… Weiterlesen

Alle Beiträge anzeigen

Nur der Text ist unter der Lizenz Creative Commons Namensnennung-Share Alike 4.0 International nutzbar. Alle anderen Elemente (Abbildungen, importierte Anhänge) sind „Alle Rechte vorbehalten“, sofern nicht anders angegeben.