Large-Scale Research with Historical Newspapers: A Turning Point through Generative AI

by Sarah Oberbichler

Digitized historical newspapers have been a crucial source for my research over the past decade. Exactly ten years ago, in 2014, I began my research journey with historical newspapers at the University of Innsbruck in Austria. Using a semi-automated multilingual approach to analyze a corpus of over 20,000 articles from German and Italian newspapers, I also started learning how to apply digital methods to historical sources: analyzing large-scale historical newspaper archives simply required such skills.

My research journey with newspapers has not always been easy. The effort required to create representative corpora, the limitations caused by missing or problematic article separation (see Figure 1), and the high ORC error rates in transcripts have been significant challenges. I have written multiple papers addressing these issues (Oberbichler et al. 2020; Oberbichler and Pfanzelter 2022; Pfanzelter et al. 2021; Oberbichler and Pfanzelter 2021).

Working with newspapers often meant spending more time creating a research corpus than doing historical research, it meant being frustrated with digital methods that – even though of groundbreaking nature – didn’t deliver the necessary quality of results for historical research.

When I began my new position at the Leibniz Institute of European History, I once again chose to work with multilingual historical newspaper corpora. However, this time, the process of building the corpus and preparing the data needed to be just the tip of the iceberg, with the historical research forming the substantial base. Generative AI and especially large language models (LLMs) turned out to be a turning point for my research with historical newspapers, making this change possible. Before I explain this further, let me start by giving a glance at my new research project.

The project: Tracking Transnational Multilingual News Circulation using the Topics of Migration and Environmental/Natural Disasters

International news networks already began to take shape between 1450 and 1650 (Raymond and Moxham, 2016; Raymond 2013). However, it wasn’t until the expansion and solidification of the telegraph industry in Europe and North America during the mid-1860s that a new era in global connectivity truly began. The establishment of global electronic communication networks quickly brought events from far-flung corners to communities around the world (Winseck and Pike 2007; Boyd-Barrett and Rantanen 1998). Those news were of course influenced by regional and national factors such as politics, culture, and economics, but also feature common narratives that transcend language boundaries (Ponomarenko 2020).

The case studies on (return) migration and environmental/natural disasters provide an opportunity to explore transnational flows of news. Both topics represent significant historical phenomena that triggered extensive global communication, even before modern newspapers existed (see Caracciolo 2016). Examining the multilingual discourse on these topics can yield valuable insights into the formation or manipulation of narratives and the construction of identity across languages and over time.

The project utilizes a cyclical workflow with six core steps. While a sequential progression exists, these steps are interconnected, allowing for revisiting previous stages as required (Figure 3).

By researching how events on migration and environmental disasters were discussed and represented in European historical newspapers, this project aims to shed light on the interplay between language, culture, politics, and economics as well as overcome challenges in multilingual discourse analysis.

But let’s break this down on a concrete example:

In 1908, a devastating earthquake, followed by a tsunami in Italy (specifically Messina–Reggio Calabria), resulted in a catastrophic loss of approximately 80,000 to 100,000 lives, thousands of refugees, and widespread destruction of cities and landscapes. Initially, news of this disaster propagated slowly due to interruptions in telegraph communications. However, once these communication issues were rectified, the calamity dominated both Italian and international news headlines for an extended period of over 10 months (Dickie 2006; Schemper 2019).

By leveraging newspapers, the Italian government sought to portray an image of efficient disaster response, aid to refugees, and national unity. However, this manipulation was laid bare by a number of correspondents and journalists, especially in the Italian newspaper “Avanti!”, the recognized mouthpiece of the Italian Socialist Party.

For instance, the Italian socialist and journalist Oddino Morgaro reported in “Avanti!” on January 5, 1909:

Here I am in Villa, under the pouring rain. The spectacle is similar to that of Catania. The abandonment seems total. The ferry that was caught in the storm that morning has not returned. During the three days since the disaster, no official, engineer, doctor, or any authority has arrived, nor has there been any delivery of bread. A squadron was seen near Messina, but no ship had docked here. The boats are all destroyed by the earthquake, and for four days, the strait has not been crossed [own translation].

The Italian government reacted to the disclosure of their disorganization by labeling it as “fake news”. For example, in official notes from the government printed in newspapers (see Figure 5: Newspaper clipping from “La Patria del Friuli” on January 8, 1909, titled “A Note from the Government on False News”).

They also responded by:

• Releasing letters from victims, refugees, or correspondents that shared their version of the truth.
• Organizing official visits of the king to the destroyed areas and to meet with refugees .
• Having pro-government newspapers print their press releases that described the great work of the government.
• Blaming the refugees and victims themselves as a problem.

But which “media truth” did the world outside of Italy get?

Answering such questions on different case studies between 1850 and 1950 is the main goal of my project. The local disaster in South Italy quickly became a national matter and resonated strongly on an international level. Newspapers worldwide reported on the tragedy, called for help and donations, and contributed to shaping the narratives surrounding the catastrophe.

By conducting a comprehensive examination of transatlantic news flows across various languages and nations, my objective is to perform a systematic exploration of these flows. This will allow me to reveal hidden patterns and pinpoint transnational discourse threads.

• Which of the competing narratives were reprinted by foreign newspapers and to what extent? Which reprint networks were created? What information was passed on, and what was omitted?
• What information spread “virally” throughout Europe and beyond?
• How did transnational news flows change over time?
• How did the international press react to accusations of manipulation or fake news? Were these accusations acknowledged or reflected?
• What role did refugees play in international press coverage of the disasters?
• How did the disasters contribute to nation-building processes abroad?

So, how are Large Language Models a turning point for answering those questions?

Coming back to the beginning of this blog post, using newspapers for large-scale analysis has many pitfalls. I am dealing with millions of newspaper pages in four different languages, which need to be accessed, organised, tailored to specific use cases, and analysed. This are time-consuming tasks that can be handled a lot faster and more efficient with LLMs.

It is important to outline, that LLMs are a lot more than chatbots! They can be used for many traditional NLP tasks at once with little or no training. So far, I am experimenting with Llama, GPT, and Gemini models. I am accessing the models via APIs, which allows me to work with my data in my own local coding environment.

LLMs for text extraction:

An essential step in creating my corpora involves extracting the articles that pertain to my research questions from the newspaper pages. Article segmentation for historical newspapers can be based on layout information and graphical elements (image) as well as on textual context (data). While the former is very challenging due to the changing and complex layouts of historical newspapers, the latter needs a human-like understanding of the texts provided.

This is precisely where LLMs come into play. The following example shows raw data from a page of the German newspaper “Deutsche Allgemeine Zeitung”, which contains an article about an earthquake. The following code was used to extract the yellow highlighted text using the Gemini 1.5 Pro model.

The extracted text: Die Insel Niphon wurde am 23. Dec. von einem starken Erdbeben heimgesucht, welches die volkreiche und blühende Stadt Ohosaco gänzlich zerstörte und in Simoda große Verwüstungen anrichtete. Auch Jeddo hat empfindlich gelitten. Von den Dschonken im Hafen von Simoda wurden viele landeinwärts getrieben und die russische Fregatte Diana erlitt solche Beschädigungen, daß sie sank. Die Mannschaft wurde gerettet und erhielt von der amerikanischen Fregatte Powhattan sowie von den Japanesen jede Unterstützung. Jetzt befindet sie sich in Hida, 30 englische Meilen von Simoda, wo sie bis zum Frühjahre bleiben wollte.

Another advantage of the semantic understanding of LLMs is that irrelevant articles can be filtered out simultaneously during the extraction process. For example, the term “earthquake” can also be used in different contexts, such as a “political earthquake”. The response of the LLM can appear as follows:

Respond: I'm sorry, but the provided text does not contain any specific reports about earthquakes. The text describes military actions and sieges during a historical conflict, without direct mention of earthquake events.

Click on “Open in Colab” to see the full code or go to the IEG Github page to see all notebooks:

LLMs for data cleaning:

Another task where the context-understanding of LLMs helps to process data is the correction of OCR errors. OCR quality has been a long-standing issue in digitization efforts. Historical newspapers are particularly affected due to their complexity, historical fonts, or degradation. Additionally, OCR technology faced limitations when dealing with historical scripts.
The following example from the Italian newspaper “La Stampa” shows how bad OCR (caused by the low quality of the scan) can be impressively improved using the Llama 3 model:

Scan	OCR	OCR corrected (Llama 3)
	La popolazione non volle rientrare nell] caîe per la paura di qualche ‘altra sco: 20. AlPalba futta In gente era ancora rac. colla; per le vie, della città. S’improvutsano ‘processioni per implorare il santo Andrea di ‘Avellino, Molte donne del popolo. în preda a vinissima a	La popolazione non volle rientrare nelle case per la paura di qualche altra scossa. All’alba tutta la gente era ancora raccolta per le vie della città. S’improvvisavano processioni per implorare il santo Andrea di Avellino. Molte donne del popolo, in preda a un’invincibile angoscia…

Click on “Open in Colab” to see the full code or go to the IEG Github page to see all notebooks:

LLMs for entities extraction:

Entities such as individuals, organizations, places, etc, are important data points, while events are interactions between entities that deviate from typical interaction patterns. Detecting those events (locating them), and extracting event arguments (type of event, who, where, when, etc.) helps to discover sets of events matching patterns of events in an event context. The code in Figure 7 shows how the OpenAI model GPT-4o was used to extract events on earthquakes in Italian newspapers. While zero-shot learning (handling tasks it wasn’t explicitly trained for using its vast pre-trained knowledge) was used for OCR post-correction and article separation, for the event detection, in-context learning (quickly adapting to new tasks with minimal examples) was applied. This means not only a prompt but also an example of what kind of response is expected was provided.

Response when an event was dedected:

['Tipo di evento: Terremoto  ', 'Origine notizia: Napoli e Avellino  ', 'Data evento: 7 dicembre  ', 'Luogo evento: Napoli, Avellino, Calitri, Italia  ', 'Persone coinvolte: Popolazione di Napoli, Avellino e Calitri, Questore, corpi di guardia, guardie, residenti, donne del popolo  '] 

Response when no event has been detected:

['Nessun evento.']

Click on “Open in Colab” to see the full code or go to the IEG Github page to see all notebooks:

Next step: Using LLMs to discover discursive patterns

As a next step, LLMs will be applied to discover discursive patterns within news coverage. Here again, the human-like context understanding of the models helps to grasp the semantic meaning from the provided text. More blog posts will follow on this topic in the future.

In conclusion, a few critical notes:

Some critical aspects also need to be highlighted: the architecture of LLMs challenges the reproducibility of their outputs (Kosch and Feger 2024), and the training process of the models might raise ethical considerations. It is further crucial to be aware of what LLMs can do and what they are not suitable for. Using LLMs via APIs locally has the advantage that model parameter settings are controllable and can be documented, as well as the use of a specific version of a model. For the ethical aspects of LLMs, initiatives such as “Common Corpus” take important steps towards AI ethics. Common Corpus is an international initiative coordinated by Pleias, involving researchers in LLM pretraining, AI ethics, and cultural heritage. Such initiatives are very important for future developments in AI for the humanities.

---

References:

• Boyd-Barrett, Oliver, and Terhi Rantanen. 1998. The Globalization of News. 1 Oliver’s Yard, 55 City Road, London EC1Y 1SP United Kingdom: SAGE Publications Ltd. https://doi.org/10.4135/9781446250266.
• Caracciolo, Carlos H. 2016. ‘Natural Disasters and the European Printed News Network’. In News Networks in Early Modern Europe, edited by Joad Raymond and Noah Moxham, 756–78. Brill. https://www.jstor.org/stable/10.1163/j.ctt1w8h1ng.40.
• Dickie, John. 2006. ‘Timing, Memory and Disaster: Patriotic Narratives in the Aftermath of the Messina–Reggio Calabria Earthquake, 28 December 1908’. Modern Italy 11 (2): 147–66. https://doi.org/10.1080/13532940600709262.
• Kosch, Thomas, and Sebastian Feger. 2024. ‘Risk or Chance? Large Language Models and Reproducibility in Human-Computer Interaction Research’. arXiv. http://arxiv.org/abs/2404.15782.
• News Networks in Early Modern Europe. 2016. Brill. https://www.jstor.org/stable/10.1163/j.ctt1w8h1ng.
• Oberbichler, Sarah. 2020. Autochthone Minderheiten Und Migrant*innen: Mediale Argumentationsstrategien von 1990 Bis 2015 Am Beispiel Südtirols. Innsbrucker Forschungen Zur Zeitgeschichte 29. StudienVerlag.
• Oberbichler, Sarah, and Eva Pfanzelter. 2021. ‘Topic-Specific Corpus Building: A Step towards a Representative Newspaper Corpus on the Topic of Return Migration Using Text Mining Methods’. Journal of Digital History 1 (1): 74–98. https://doi.org/10.1515/jdh-2021-1003.
• ———. 2022. ‘Tracing Discourses in Digital Newspaper Collections’. In Digitised Newspapers – A New Eldorado for Historians?, edited by Estelle Bunout, Maud Ehrmann, and Frédéric Clavert, 125–52. De Gruyter. https://doi.org/10.1515/9783110729214-007.
• Oberbichler, Sarah, Eva Pfanzelter, Stefan Hechl, and J. Marjanen. 2020. ‘Doing Historical Research with Digital Newspapers – Perspectives of DH Scholars’. EuropeanaTech Insight 16: Newspapers. https://pro.europeana.eu/page/issue-11-generous-interfaces.
• Pfanzelter, Eva, Sarah Oberbichler, Jani Marjanen, Pierre-Carl Langlais, and Stefan Hechl. 2021. ‘Digital Interfaces of Historical Newspapers: Opportunities, Restrictions and Recommendations’. Journal of Data Mining & Digital Humanities HistoInformatics (January). https://jdmdh.episciences.org/7069/pdf.
• Ponomarenko, Lesia. 2020. ‘Migrant Crisis in European Multilingual Media: Identity Construction across Languages’. Journal of Intercultural Communication 20 (1): 17–31. https://doi.org/10.36923/jicc.v20i1.299.
• Raymond, Joad, ed. 2013. News Networks in Seventeenth Century Britain and Europe. 2nd ed. Hoboken: Taylor and Francis. http://public.eblib.com/choice/publicfullrecord.aspx?p=1397022.
• Schemper, Lukas. 2019. Writing the History of ‘Natural’ Disasters. The Case of Messina.
• Winseck, Dwayne Roy, and Robert M. Pike. 2007. Communication and Empire: Media, Markets, and Globalization, 1860 – 1930. American Encounters – Global Interactions. Durham: Duke University Press.

---

Featured image: Image generated by GPT-4. Prompt: “A digital interface emerging from a traditional newspaper, blending old-fashioned print with modern digital elements”.

---

---