Featured Image: Mining (ca. 1916) – Louis Schaettle (1867-1917)
When we look at notions of privacy across a large corpus of literary sources, a rich tapestry of different nuances emerges. In this blogpost we will delve into digital methodologies that can be used to explore such a corpus. Some researchers have embraced digital methodologies as these methods’ usefulness and dynamic nature resonate with their perspectives and pursuits. Others see these digital tools – especially LLMs (large language models) – as testaments of academic degeneration and a threat to academic integrity. In this text we hope to challenge and leverage these concerns against the digital by showing the utility that a digital approach can bring. The intention of this text is to captivate potential users of digital methodologies by showing how digital tools such as text-mining, but also to some degree LLMs, can elevate traditional historical approaches.
You might be thinking that the paragraph you just read seemed a bit off? The keen-eyed would have noticed a few words that today should ring warning bells as indicators of a text having been manipulated, or even wholly written by, an LLM. It is in fact not; but it has purposely been written to resemble one. One of the main tells would be the fact that the paragraph contains, by intention, the top ten words most commonly used by ChatGPT. (1) The same year ChatGPT was released to a broader public, 2023, an enormous spike appeared in the usage of the LLMs preferred words within academic articles. The word “delve” for example was used in close to six times as many publications in 2023 compared to previous years. (2) The critique that there is a threat to academic integrity is therefore not unfounded, nor untrue. As a response, especially in regard to how common generative text has become within student assignments and projects, many teachers across the educational system are currently working towards an intelligent and nuanced integration of AI and other digital tools within the humanities. However, one of the general reactions among scholars in the humanities has been one of fear and adamant resistance, and proponents of zero tolerance policies when it comes to using AI are numerous. It is easy to criticize what we do not understand, but it is also hard to argue how to regulate something without any personal experience. Do these LLMs really only pose a threat to academia? In many cases the answer is yes, but in some we can see a non-intrusive and plagiarism-free utility with these LLMs.
This post is not meant to belittle the very real problems that come with LLMs and AI-plagiarism but rather to show how some parts of LLMs, and other digital tools, can be used to both support and effectivize traditional approaches in the humanities. The digital methodology, and the associated tools, that will be presented here relate to the navigation and identification of relevant passages within a large and varied text corpus that would otherwise be hard, or even impossible, to grasp with traditional close-reading methods. We take our point of departure in the methodology developed for Viktor Wretström’s PhD-thesis Reaching and Breaching Privacy in Garden Spaces which is applicable for any analysis of a larger corpora of digitalized texts. The core concept in this methodology is text analysis through digital tools, or more specifically, text-mining through the software Voyant.
In Reaching and Breaching Privacy in Garden Spaces the central concepts relate to privacy in garden spaces. This is pursued in a longue dureé perspective from antiquity to the medieval period and over a wide geographical region surrounding the greater Mediterranean area. The project is part of the Centre for Privacy Studies and is situated within the field of history. This article is a collaborative effort of Wretström together with Julia Kjellgren Sten and Ellen Saarinen, both of whom were research interns on the project during the fall of 2025. In the PhD-project Voyant is utilized to search for keywords related to garden spaces to effectively locate passages that are connected with either reaching, or the breaching, of privacy within these spaces. During their internship Kjellgren Sten and Saarinen worked with the preparation and categorization of sources, as well as adding new sources from other databases to the PhD-project database.

Compared to traditional close readings of sources, this digital approach contributes an important factor related to scale; digital text analysis allows us to work with texts and corpora that would be too large to study through traditional close reading practices. (3) Digital tools can aid us in finding relevant passages in large bodies of text, thereby allowing us to conduct so-called “distant readings”. (4) A risk that comes with distant-reading is the fact that macroscopic approaches to sources can both disenfranchise the scholar from the corpora that they study and make it hard to contextualise the text they find within their “big data”. (5) At the same time big data approaches can give us a false sense of security as highlighted by literary scholars Ryan Heuser and Long Le-Khac “even with millions of texts, the evidence generated can still be anecdotal.” (6) To avoid this disenfranchisement we have followed a middling approach, sometimes called a mixed method large-scale text analysis. This method uses both so-called distant reading – that is, text-mining – to identify and find passages that are of relevance for the study but also uses more traditional historical approaches – close-reading – once these passages have been identified and categorized. (7)
A general problem we face with a textual digital approach is the conflict between what is called signal and concept. Signal is defined as the data, the words that are being identified by the digital tool, and concept is the phenomenon that signal is meant to represent. Because the tools can only measure the signal – i.e. the words – it is up to the digital scholar’s analysis to bridge this gap between signal and concept. This is primarily done by applying traditional approaches to the digital textual findings; close-readings and contextualisation. For example, the signals used for the thesis this text is based on are not directly related to its concept – privacy. The signals that are used are instead related to the locus, the gardens, in which we can look for the concept – issues and notions of privacy.

This relates to the earlier mentioned issue between signal and concept. Many concepts that we look for within the humanities are abstract or complex and therefore hard to “pin down” with a single or even several different signals. In general we can approach this from two angles: one direct and one indirect. The first is attempting to find signals that match the concept as closely as possible – for example the terminological approach to finding privacy through priv*-words. The second, applied in the thesis project, is using a signal that locates the concept in the proximity of that signal. In this approach we can look for the concept by using signals that help us find locations, individuals, spaces or other objects which we expect to be related to different notions and issues of privacy. This is more related to the approach of finding privacy through the heuristic zones – through for example different privacy related loci – or through interlinked words that are related to notions of privacy as done in the approach of semantic mapping for privacy. (8)
A requirement for any digital approach relating to source texts is that the digital tools themselves can read and use the text that is to be analysed. This readability is directly related to something known as OCR (optical character recognition) which measures how well a program can recognize individual characters and written language from, for example, a scanned document. The programme Voyant, which is used for this project, is a text analysis tool that reads texts and returns data; for example, word frequencies, co-occurrence mapping, phrase pairings, and topic modelling. (9) To utilize Voyant at its fullest requires texts that are already transcribed or documented with a high quality and clear OCR-readability. Where can we find such sources we might ask. Well, the internet!
Online databases are exceptional treasure troves of textual sources, particularly for texts with expired or non-existent copyrights. There are a multitude of such databases and their quality and focus can differ quite a bit. While some are specialized on certain topics, others are more general, and whereas some are small and organized others are enormous and chaotic. Some databases are part of research projects or hosted by universities while others lean more heavily on hobbyist and crowd-sourcing efforts. Due to this non-conformity of online collections, it can at times be challenging to find and verify what a database contains. For this reason, a more conventional source critical approach will always be required when working with the texts in these databases. It is important to be aware of when, where, and – if possible – by whom the text was written, as well as by whom it was transcribed and uploaded. This information is not always provided by the website and must thereby be found through other sources. Specifically for this project, when a text in an online database has been flagged as relevant, it is logged and saved in a separate database maintained for the project. Here metadata regarding the spatial, cultural, and temporal origin of the sources is noted to make later processing, analysis, and trackability easier.
Since the project will remain active over several years, the continued availability of these sources must be guaranteed. To ensure that access to the sources will not be compromised if the original website becomes unavailable, a backup is essential. In some cases, this can be achieved by the simple act of copying and pasting text from the website into a document. Paradoxically, the more developed and user-friendly an archive is, generally, the more difficult it is to extract the data. A typical example of such an archive is Wikisource, where the text is often divided into multiple subchapters linked to the main index page, and the time-consuming, tedious task of transferring the text by opening each and every link, can be intimidating, if not a complete waste of human resources.
Through programming languages, such as Python, a so-called scraper script can be created to request a data extraction from a chosen website. As historians, we are seldom well-versed in coding, which is why LLMs like ChatGPT can prove useful in producing functioning scripts. The LLM will simply write, and produce, a functioning script for you in seconds if prompted correctly. Here the challenge is communicating the specifics of the requested script to the AI. Different websites and databases vary in structure, and factors such as HTML and other metadata must be accounted for in each script. Depending on previous experience, or lack thereof, it can be difficult to know what information the AI needs in order to produce successful results. The process may therefore initially involve some troubleshooting and trial-and-error approaches with the script. Once a working script exists, it can be run and the “scraped” text gets copied, ready for use. A scraper can extract and copy text from thousands of links resulting in several thousands of pages of text in the matter of minutes.
Once the sources have been prepared and backed-up, we enter them into a digital text-analysis tool, in this case Voyant, for text-mining. Keywords relating to garden spaces – signal – are then used to find relevant passages – concept – within the prepared and backed-up sources. The text-mining for this project is therefore mainly reliant on only one of Voyant’s many functions: keyword searchers. Voyant both simplifies and widens the scope of the regular word-search available in most word processing programmes by allowing one to search for several keywords simultaneously across an entire corpus. (10) The found passages that relate to garden spaces are then collected in a separate “garden space”-database where it can be further analysed, through close-readings, to identify and categorize different notions and issues of privacy that occur within these garden spaces.
In summary, the digital approach of large-scale text analysis through text-mining tools such as Voyant, can help us expedite the processes of finding passages in our sources while at the same time allowing us to look at sources that would otherwise have been ignored because of the size, or pre-perceived idea of non-relevance. At the same time it effectivizes traditional methodologies – such as close-readings – by allowing us to quickly locate passages that are of interest and value to our research. As scholars within the humanities, we need to start changing our perspective on both digital tools and on LLMs; both are here to stay and will change the academic world with or without us. The approach should not be one of complete acceptance and surrender to the pervasive and plagiaristic nature of generative text; but rather one of understanding and interest in how we can utilize these tools in an academically correct way to effectivize and broaden academic research.
(1) These ten words are: explore, captivate, tapestry, leverage, embrace, resonate, dynamic, testament, delve, elevate (https://aiphrasefinder.com/common-chatgpt-words/).
(2) Shapira 2024 (https://pshapira.net/2024/03/31/delving-into-delve/).
(3) Sinclair and Rockwell 2012, 242–244; Wiedemann 2013, 335; Meister 2023, 216.
(4) Cf. Varela 2021, 83–84.
(5) Drouin 2014, 110-111.
(6) Heuser and Le-Khac 2011, 80-81.
(7) Cf. Carr Jones 2018; Howard 2018; Gitre, Kramer, Dawson and Korkmaz 2023.
(8) Bruun 2021.
(9) Voyant is only one of many tools with these types of functions. See Martin 2019 and Klein Käfer 2025 for examples of topic modeling with the digital tool MALLET.
(10) For examples of research relying on other functions of Voyant see Posthumus 2016, Rambsy 2016, Luo 2020 and Swisher 2020.
Bibliography
AI Phrase Finder. 2025. “10 Most Common ChatGPT Words.” https://aiphrasefinder.com/common-chatgpt-words/.
Bruun, M. B. 2021. “The Centre for Privacy Studies Work Method.” Centre for Privacy Studies. https://teol.ku.dk/privacy/research/work-method/.
Carr Jones, P. 2018. “Mining the ICC: Macroanalysis of the Indian Claims Commission” Current Research in Digital History 1.
Gitre, E.J.K., B.L. Kramer, C. Dawson, and G. Korkmaz “Across the Color Line: Using Text Networks to Examine Black and White US Soldiers’ Views on Race and Segregation during World War II.” Current Research in Digital History 6.
Heuser, R. and L. Le-Khac. 2011. “Learning to Read Data: Bringing out the Humanistic in the Digital Humanities.” Victorian Studies 54(1): 79–86.
Klein Käfer, N. 2025. “Private Birth, Public Authority: Topic Changes from the Sixteenth to the Eighteenth Century in German Midwifery Books.” Current Research in Digital History 8.
Luo, J. 2020. “Trade and Missionary Work: A Word Frequency Analysis of Dialogues and Detached Sentences in the Chinese Language (1816) and the Pedagogy of Robert Morrison.” Comparative Literature Studies 57(3): 384–397.
Martin, J. 2018. “Talk-Back Boards and Text Mining: New Digital Approaches in Museum Visitor Studies.” Current Research in Digital History 1.
Meister, J.C. 2023. “From TACT to CATMA; or, a Mindful Approach to Text Annotation and Analysis.” In On Making in the Digital Humanities: The scholarship of digital humanities development in honour of John Bradley, edited by J. Nyhan, G. Rockwell, S. Sinclair, and A. Ortolja-Baird. 213–250. London: UCL Press.
Posthumus, S. 2016. “Reader-Player-User: Engaging Bruno Latour’s AIME Project.” Resilience: A Journal of the Environmental Humanities 4(1): 137–143.
Rambsy, K. 2016. “Text-Mining Short Fiction by Zora Neale Hurston and Richard Wright using Voyant Tools.” CLA Journal 59(3): 251–258.
Shapira, P. 2024. “Delving into ‘delve’.” Philip Shapira. https://pshapira.net/2024/03/31/delving-into-delve/.
Shawn, M. 2019. “Topic Modeling and Textual Analysis of American Scientific Journals, 1818–1922.” Current Research in Digital History 2.
Sinclair, S. and G. Rockwell. 2012. “Teaching Computer-Assisted Text Analysis: Approaches to Learning New Methodologies.” In Digital Humanities Pedagogy: Practices, Principles and Politics, edited by B.D. Hirsch. 241–264. Cambridge: Open Book Publishers.
Swisher, J. 2020. “A Tale of Three Valleys Text Mining, Spatial Analysis, and the Contested Geographies of Colorado and New Mexico’s San Luis Valley” Current Research in Digital History 3.
Varela, M. 2021. “Words as Data.” In Theater as Data: Computational Journeys into Theater Research, edited by M. Varela. 75–93. Ann Arbor: University of Michigan Press.
Wiedemann, G. 2013. “Opening up to Big Data: Computer-Assisted Analysis of Textual Data in Social Sciences.” Historical Social Research/Historiche Sozialforschung 38(4): 332–357.



Modern reconstitution of the 