Digital resources in the Social Sciences and Humanities OpenEdition Our platforms OpenEdition Books OpenEdition Journals Hypotheses Calenda Libraries OpenEdition Freemium Follow us

Rethinking Libraries in the Age of AI: Impressions and Takeaways from TPDL 2025

Rethinking Libraries in the Age of AI: Impressions and Takeaways from TPDL 2025

This year, the 29th International Conference on Theory and Practice of Digital Libraries (TPDL 2025) took place in Tampere, Finland, from September 23–26. I attended the conference together with Sarah Oberbichler from the DH Lab to present our joint paper, “Studying Model Design Biases in LLMs for Multilingual Historical Newspaper Extraction; The Messina Earthquake Case Study”, in which we propose an explainability-driven framework to detect and analyze biases in large language models used for multilingual historical newspaper article extraction.

Focusing on the conference’s many LLM-related contributions, this blog post explores how digital libraries are evolving under the continuous advancement of AI and what digital humanities can contribute to the field.

How I got to TPDL as a Master Student

What led to my journey to the land of a thousand lakes, moomin trolls, and delicious cinnamon rolls was an internship that I began with in October 2024 at the DH Lab of the IEG, supervised by Sarah Oberbichler as part of my master’s studies at the University of Mainz in Digitale Methodik in den Geistes- und Kulturwissenschaften. This 3-month internship involved supporting her research project on natural disasters in historical newspapers by testing and evaluating different LLMs for the task of extracting topic-specific articles from newspaper transcripts. We already shared some preliminary results in a previous blog post that you can read here if you are interested.

After the internship had ended, our collaboration expanded over the course of the following months to include two colleagues from computer science, Carlos-Emiliano González-Gallardo (Université de Tours) and The Trung Tran (University of Science and Technology of Hanoi), culminating in the paper we submitted to and now presented at TPDL. Thanks to the support of the DH Lab, I was able to attend the conference, so of course I took this chance to travel to Tampere – A journey and experience that I would certainly not have expected a year ago when starting the internship and am very grateful for.

Some Notes About TPDL and Digital Libraries

Before looking at how recent generative AI technologies are reshaping libraries, it’s worth stepping back and looking at TPDL itself. One of the largest conferences in the field of digital libraries, TPDL is held annually in different European cities and brings together researchers and practitioners from a heterogeneous field of participants, both academia and industry, perhaps most prominently with backgrounds in computer science or library and information science, who do work on or with digital libraries.

The adjoining question might be, what exactly are digital libraries? Of course, the composites of this term (digital and libraries) already give some idea, but digital libraries are an umbrella covering more than you’d might expect. According to the conference’s proceedings:

TPDL encompasses the many meanings of the term ’digital libraries’, including new forms of information institutions; operational information systems with all manner of digital content; new means of selecting, collecting, organizing, and distributing digital content; and theoretical models of information media, including document genres and electronic publishing.”1

It is important to note in this context that digital libraries are a quickly evolving field: The rise of new technologies affects how they operate and opens a vast variety of research directions. This reflects in TPDL covering Document Analysis and Recognition and Information Retrieval systematically as new research areas this year beyond their previous topic scope.

Relationship between Digital Libraries and Digital Humanities

The line between digital libraries and digital humanities is quite fluent and blurry. In a 2012 article titled “What is Digital Humanities and what’s it doing in the library?” Micah Vandegrift opens with following statement: “tl;dr – Libraries and digital humanities have the same goals.”2) The central common goal here, according to Vandegrift, is accessibility of information. Further, Vandegrift describes libraries as the central place where DH work can, should and is being done. However, according to him it is mainly the digital library that provides a service to the DH scholar by making contents accessible, while digital humanists can work in libraries, but the overarching digital humanities field doesn’t have its place in the library. While some time has passed since Vandegrift’s reflections on the convergence of digital libraries and digital humanities, newer research by Hermann Kroll et al. supports the notion of a rather one-sided support relationship: In their article “Requirements for a digital library system: A case study in digital humanities” they are asking how libraries can support DH researchers.3 Libraries are digitizing content and making it accessible through new technical ways, e.g. by making metadata available and contents searchable through keywords (or in recent times increasingly AI-driven methods) – but what are digital humanities doing for libraries? Is it a one-sided road of profit between the communities? Or how can, in return, digital humanities support digital libraries? Maybe it can be argued that DH, with their understanding of both the computational and interpretational world, are especially well suited to develop new methodologies that enhance access to library contents, which then in return supports digital libraries. This is supported in DH’s presence at a conference like TPDL that seeks to advance digital libraries.

This year’s Conference Program

Of course, in recounting the conference program, I am biased by what I listened to – which was unfortunately not everything because of parallel sessions. The full program can be accessed via the TPDL webpage. For a briefer overview, here is a wordcloud highlighting frequently occurring words across contributions:

Word cloud with most frequent title terms from contributions (full papers, findings papers, short papers, keynotes, workshop proposals)
Fig. 1: Word cloud with most frequent title terms from contributions (full papers, findings papers, short papers, keynotes, workshop proposals)

Out of overall six paper sessions, one was dedicated to knowledge graphs, which highlights their increasing relevance, particularly as a way to structure and interlink information for more explainable and context-aware AI applications. Other sessions covered Digital Archives, Citation Management, Scholarly Issues, Digital Humanities, and Large Language Models.

Some of the conference highlights were certainly the four keynotes. Felix Naumann opened with a talk on data quality in the age of AI, Evaggelia Pitoura continued this line of thought with a keynote on explainability and fairness, discussing their interplay in responsible AI and data-driven research, Eero Hyvönen turned the focus toward digital humanities, presenting the Finnish Sampo Model as a large-scale semantic infrastructure enabling AI-based knowledge discovery across cultural heritage data. Finally, Yannis Ioannidis, one of the founders of the digital libraries community, concluded with an inspiring talk on interactive digital storytelling, exploring how narrative, personalization, and agency can enrich engagement with digital and cultural heritage.

The keynotes and paper presentations were complimented by a poster and a demonstration session, which gave the opportunity to directly talk to attendees who had submitted short papers about their research.

A Closer Look at LLM Contributions

Overall, it was noticeable that LLM/generative AI contributions were very prominent at this year’s TPDL – not only in their designated, but across most sessions, as well as the keynotes, workshops and poster presentation. An analysis of the most frequent terms across all conference contribution titles supports this impression, as “LLM” is number one among the most mentioned:

The most occurring keywords across all contribution titles (stopwords excluded; singular/plural aggregated)
Fig. 2: The most occurring keywords across all contribution titles (stopwords excluded; singular/plural aggregated)

This omnipresence makes sense under the light of LLMs being still a relatively new (and constantly evolving) technology, offering many new applications for digital libraries.

Despite people questioning the hype often having good reasons to do so, interest keeps persisting, which showed in the fact that the LLM paper presentation session was by far the most crowded at the conference: Some people even had to stand in the lecture room. In the LLM session a wide variety of model applications were explored. Christin Kreutz and colleagues looked at how researchers actually engage with LLMs in data discovery (including an amusing detail: Despite a focus on scholarly models, model outputs frequently included unprompted emojis, most notably the 🚀); Davide Caffagni et al. investigated generating synthetic data for low-resource tasks; Durgesh Nandini et al. presented work on prompt-based triple extraction and knowledge graph enrichment; Naoki Sawahata’s team showed approaches for dependency tracking in short event descriptions; and Kehinde Ajayi et al. closed the session with introducing a benchmark for question answering over complex scientific tables.

The BIRDS workshop I took part in on the first conference day, organized by Christin Katharina Kreutz (Herder Institute/TU Mittelhessen) and Hermann Kroll (Technical University of Braunschweig) also shed light on how AI is and will keep transforming the field of digital libraries. Through an interactive format designed to connect domain experts and researchers working on digital library systems, the workshop fostered lively discussion around current challenges and collaborative directions, and featured a very interesting keynote by Jian Wu (Old Dominion University) who put forward the vision that in an age of information overflow, now finally someone may be reading all the papers, namely: bots. While concluding that existing core infrastructures such as search engines and knowledge graphs remain indispensable to provide provenance, scalability and interpretability, he described a future where intelligent digital libraries based on LLMs serve as knowledge hubs to researchers, capable of question answering, hypothesis generation, and claim verification.

We presented our own paper “Studying Model Design Biases in LLMs for Multilingual Historical Newspaper Extraction; The Messina Earthquake Case Study” as part of the Digital Archives session.

“Studying Model Design Biases in LLMs for Multilingual Historical Newspaper Extraction; The Messina Earthquake Case Study” Title Slide
Fig. 3: The opening slide of our presentation

While in featuring LLMs, our paper was far from an outlier at TPDL, what set it apart from other contributions was its humanities-driven perspective and focus on ethical dimensions. Our study investigates how LLMs handle multilingual historical newspapers and where their design choices can introduce systematic biases. Instead of relying on layout or image data, our approach uses purely textual, prompt-based extraction from newspaper pages in English, German and French. Using the 1908 Messina earthquake as a case study, we developed an explainability-driven framework to go beyond just assessing how many percent of relevant articles on the earthquake were found or not, but also detect and analyze biases induced through model design in article extraction tasks. Explainability in this context means to make model behaviour understandable to humans, which is of pivotal importance for ensuring research integrity, but tricky when working in a black-box environment as it tends to be the case with LLMs. Part of our strategy was to integrate explainability directly into our prompt design, e.g. by asking models to give a number on how confident they are about the extraction being correct.

By comparing six state-of-the-art models, we identified recurring bias patterns that stem from model design rather than data quality, such as overconfidence bias – meaning models tend to give too optimistic numbers about their own performance. Knowing where biases emerge can help assess risks in working with LLMs, craft better prompts and make more informed model choices to – in our example – responsibly make use of newspapers as very rich historical resources that get significantly more accessible for larger-scale research through LLMs.

Rethinking Together: Concluding Reflections

Participating in TPDL 2025 offered me a very enriching glimpse into the Digital Libraries research landscape: Aside from being able to listen to insightful presentations, the exchange with other attendees made the experience particularly inspiring, be it over shared methodologies and challenges or just intercultural conversations.

While large language models may still be in their hype phase, as the sheer number of related contributions at the conference showed, AI technologies are very likely to stay and keep transforming the field. With that comes both opportunity and responsibility. Rather than standing on the sidelines, we can help shape how AI is used: Critically, reflectively, and in line with the values of fairness and openness that have long defined the digital humanities.

Encouragingly, TPDL is planning to introduce a qualitative track next year alongside its established quantitative focus. This might help strengthen the dialogue between digital libraries and the humanities and drive forward a symbiosis between the two.

As we rethink libraries in the age of AI, data quality, truth and responsible technology use are more important than ever. Perhaps digital humanities can play a key role in shaping the vision of how the libraries of the future can and should look like.

Literature

  • Balke, W.-T., Golub, K., Manolopoulos, Y., Stefanidis, K., & Zhang, Z. (Eds.). (23.–26. September 2025). Linking theory and practice of digital libraries: 29th International Conference on Theory and Practice of Digital Libraries, TPDL 2025, Tampere, Finland, proceedings (Preface, p. v). Springer. https://doi.org/10.1007/978-3-032-05409-8
  • Kroll, H., Kreutz, C. K., Jehn, M., & Risse, T. (16. December 2024). Requirements for a digital library system: A case study in digital humanities. In Proceedings of the 24th ACM/IEEE Joint Conference on Digital Libraries (JCDL ’24) (pp. 1‑3). ACM. https://doi.org/10.1145/3677389.3702502
  • Mauermann, J., & Oberbichler, S. (31. January 2025). LLM biases: Expected and unexpected model design effects in historical newspaper article extraction on the Messina earthquake. DH Lab. https://doi.org/10.58079/137qr
  • Oberbichler, S., Mauermann, J., Tran, T., & González-Gallardo, C.-E. (Sep 2025). Studying model design biases in LLMs for multilingual historical newspaper extraction: The Messina earthquake case study. In Proceedings of the 29th International Conference on Theory and Practice of Digital Libraries, TPDL 2025, Tampere, Finland. HAL: 05166022. https://univ-tours.hal.science/hal-05166022v1
  • Vandegrift, M. (27. June 2012). What is digital humanities and what’s it doing in the library? In the Library with the Lead Pipe. https://www.inthelibrarywiththeleadpipe.org/2012/dhandthelib/

Featured Image: Johanna Mauermann, Collage, 2025. All rights reserved. Original photos: Antoine Doucet.

  1. Wolf-Tilo Balke, Koraljka Golub, Yannis Manolopoulos, Kostas Stefanidis, & Zheying Zhang (Hrsg.). (2025). Linking theory and practice of digital libraries: 29th International Conference on Theory and Practice of Digital Libraries, TPDL 2025, Tampere, Finland, September 23–26, 2025, proceedings (Preface, p. v). Springer. https://doi.org/10.1007/978-3-032-05409-8 []
  2. Vandegrift, M. (27. June 2012). What is digital humanities and what’s it doing in the library? In the Library with the Lead Pipe. https://www.inthelibrarywiththeleadpipe.org/2012/dhandthelib/ (last accessed: 29.10.2025 []
  3. Kroll, H., Kreutz, C. K., Jehn, M., & Risse, T. (16. December 2024). Requirements for a digital library system: A case study in digital humanities. In Proceedings of the 24th ACM/IEEE Joint Conference on Digital Libraries (JCDL ’24) (pp. 1‑3). ACM. https://doi.org/10.1145/3677389.3702502 ; Full text freely available here https://www.semanticscholar.org/reader/845a715039347ddd6bb54f2dae46ea4c471e804e . []

OpenEdition schlägt Ihnen vor, diesen Beitrag wie folgt zu zitieren:
Johanna Mauermann (31. Oktober 2025). Rethinking Libraries in the Age of AI: Impressions and Takeaways from TPDL 2025. DH Lab. Abgerufen am 17. November 2025 von https://doi.org/10.58079/152m0


Schreibe einen Kommentar

Deine E-Mail-Adresse wird nicht veröffentlicht. Erforderliche Felder sind mit * markiert

This site uses Akismet to reduce spam. Learn how your comment data is processed.