DATA CREATION CHALLENGES IN HISTORICAL RESEARCH: OUR FIRST-YEAR JOURNEY

---

A paper by Richard Hadden, Suzana Sagadin, Marcella Tambuscio and Georg Vogeler

---

The first year of the ManMax research project has been a journey of discovery and innovation. Our team has faced numerous technical hurdles, particularly in processing historical data. In this blog post, we aim to share some of these challenges and how we addressed them, hoping to provide insights for other researchers facing similar issues.

---

Digitising Maximilian, based at the University of Graz, focuses on extracting and structuring historical data provided and digitised by the other sub-projects. The aim is to build a prosopographical database according to the Factoid model (Bradley & Short 2005), and to use the collected data to conduct analyses using computational methods.

---

Our database uses the Austrian Prosopographical Information System (APIS: Schlögl & Lejtovicz 2018, 2020). For our project, we expanded the structure of the database beyond the  basic tenets of Factoid prosopography. In the original conception of Bradley, and later expanded by Pasin and Bradley (2015), a Factoid represents an assertion by a researcher that a given source provides a piece of information about a person. This model has been used in other projects, such as the People of Mediaeval Scotland (POMS), to represent relatively simple interactions between persons based on Royal Charters.

---

The scope of the ManMax project is, however, much wider: not only do the research interests of the various sub-projects, and hence the types of data to be captured, differ greatly, but the nature of the interactions is more complex. For example, we aim to capture the relevant interactions contained in complex chains, such as one person ordering a second person to commission the creation of an object by a third person. To enable this, we extended the standard Factoid model to allow nested statements: an order (in addition to being given and received by persons) can contain a further statement (the thing ordered) —theoretically, allowing a chain of orders ad infinitum (and, indeed, ad nauseam)!

---

Of equal importance is the sheer number of possible statement types, ranging from the generic —a person is born or dies; has familial relations; occupies a particular role —to the very project specific— the depiction of a person in a piece of art; the involvement of a person in a secretarial activity pertaining to the creation of a particular printed book from a set of “abstract” texts; the assembly of multiple pieces of armour into a single suit of armour… In doing so, we aim to allow the different sub-projects to create the data they require for their particular areas of interest, while simultaneously connecting the persons involved together in a single prosopographical database, suitable for both direct use by researchers and computational analysis.

---

Factoid Editor (the custom APIS interface for building Factoids)

---

In addition to allowing ManMax team members to enter data into APIS directly, we are also working on automated methods to transform less structured data (also provided by the team) into the Factoid structure. Most of this data consists of short pieces of text. The microgoal we want to focus here is the extraction of the relevant information from these heterogeneous data and the creation of APIS entries. With historical data coming from multiple sources and formats, integrating it into a unified system is complex because the texts varied greatly in terms of language, structure, and content.

---

Example 1: The city council of Augsburg pays 1 fl to Hans Armhaus, Maximilian’s trumpeter (1489), Augsburg, Stadtarchiv, Baumeisterbuch 82, fol. 16r.

---

For instance, if we have a list of payments as in example 1, all presenting the format ENTITY A pays AMOUNT to ENTITY B, it is quite easy to define a method that automatically creates the single Payment statements (the payer, payee and the amount are simple to extract). The only challenge lies in identifying these persons/organisations and this could be done quite efficiently through NER tasks or using regular expressions if the very same structure is repeated across the document.

---

Trumpeters in the Triumph of Maximilian, Graz exemplar (1526), p. 20.

---

Example 2: Hans Mader is instructed to continue the promised stipend to Benedict Zuckenranft, following the death of Jörg Gossembrot, to whom the administration of these payments had earlier been entrusted, Vienna, ÖNB, Cod. 14297, ff. 1v-2r. According to Grantley McDonald —who provided this example— from the context of the statements, we may also infer that the stipend was made to cover specifically musical studies —and that the original order for these payments to be made was given by Maximilian himself.

---

Thus, in terms of the Factoid model we use in APIS, this information comprises multiple (nested) statements:

• The original order given (by Maximilian) for
  • the ongoing payment of a stipend to Benedict Zuckenranft by Jörg Gossembrot
    • for Benedict Zuckenranft to study music
  • The death of Jörg Gossembrot;
  • The new order given (by Maximilian for
    • the new (ongoing) payment of money to Benedict Zuckenranft by Hans Mader;
      • for Benedict Zuckenranft to study music

---

However, it takes a large amount of time (and thought) to fully extract and structure the large informational content contained in even three lines of text —much is implicit, including the ordering and causality of the events. This is leaving aside the difficulty of identifying the persons concerned, even though we have aimed to tackle this by ingesting the persons contained in Angelika Schuh’s index of the Regesta Imperii (Schuh 1993, 1996). For example, “Jörg Gossembrot”  is known more commonly as “Georg”, and his surname has five variant spellings: “Casofinbroth”, “Gossemprot”, “Gossemprott”, “Gossemprutt”, “Gossennbrot”.

---

As such, we are exploring a range of automated methods to tackle this problem. We are currently experimenting with Large Language Models (LLM), deep learning models that are pre-trained on a huge amount of data and that can perform text analysis tasks very well.

---

For instance, regarding the last example, we asked GPT-3 to create a JSON representation using the following instructions (summary) :

• create a JSON isolating the sentences,
• list the entities involved,
• highlight the main actions and associate a label to them

---

And we obtained the following result:

---

These  results are quite promising but of course we need to tune the model to adapt it to the ManMax ontology and realise a uniform set of labels. We should also keep in mind the non-deterministic (probabilistic) nature of these models, which means that the results are not reproducible. In recent months, we have been discussing various approaches with team members to manually check the validity of the generated data.

---

To conclude, the main challenge in this first year has been how to create a flexible platform to incorporate and structure these different data according to the perspective and research goals of our colleagues working in the other subprojects: we are trying different approaches to make the process faster and more efficient.

---

References

Bradley, John, and Harold Short. “Texts into Databases: The Evolving Field of New-Style Prosopography”. Literary and Linguistic Computing 20, Nr. Suppl (1. January 2005): 3–24. https://doi.org/10.1093/llc/fqi022.

Pasin, Michele, and John Bradley. „Factoid-Based Prosopography and Computer Ontologies: Towards an Integrated Approach“. Literary and Linguistic Computing 30, Nr. 1 (1. April 2015): 86–97. https://doi.org/10.1093/llc/fqt037.

Schlögl, Matthias, und Katalin Lejtovicz. „A Prosopographical Information System (APIS)“. In BD-2017 Biographical Data in a Digital World 2017  Proceedings of the Second Conference on Biographical Data in a Digital World 2017 Linz, Austria, November 6-7, 2017, 53–58. [Budapest]: CEUR, 2018. http://ceur-ws.org/Vol-2119/paper9.pdf.

Schlögl, Matthias, and Katalin Lejtovicz. 2020. Die APIS-(Web-)Applikation, das Datenmodell und System. In The Austrian Prosopographical Information System (APIS). Vom gedruckten Textkorpus zur Webapplikation für die Forschung, eds. C. Gruber, Kohlbacher, J., and Wandl-Vogt, E., 31-48. Wien: new academic press.

Schuh, Angelika. Böhmer, J. F., Regesta Imperii XIV. Ausgewählte Regesten des Kaiserreiches unter Maximilian I. 1493-1519. vol. 1,3.: Register der Personen- und Ortsnamen 1493-1495, Köln et al.: Böhlau, 1996; Bd. 2,3. Register der Personen- und Ortsnamen 1496 – 1498. Köln et al.: Böhlau, 1993.

---

Header Image Credits: Universitätsbibliothek Graz

---

---