Open Science in the Age of Artificial Intelligence

European Experiences in the Digital Humanities

Christof Schöch
Trier Center for Digital Humanities
Trier University, Germany

International Digital Humanities Symposium:
Humanities Data Infrastructure in the Age of Generative AI
Tokyo, Japan

18 Jul 2026

Introduction

Slides: tokyo.dh-trier.de

Dear Prof. Shimoda,

Dear Prof. Nagasaki, dear Kiyonori,

Esteemed members of the audience,

Many thanks for the kind introduction, and for the opportunity to be with you today. It is a great honor for me to be here.

I would like to speak about “Open Science in the Age of Artificial Intelligence” from a European perspective. In particular, I would like to offer some thoughts on how Open Science, Artifical Intelligence and Research Infrastructures interact, in the context of the Digital Humanities.

I hope this will be a relevant impulse for your debates about these issues.

Before I start, please note that if you have a laptop or tablet with you, you can follow along with the slides at the link shown here.

Abstract

Over the last two decades, the Digital Humanities have flourished and spread across Europe. In the process, Digital Humanities researchers have embraced Open Science principles as a driving force and profoundly transformed their scholarly practices. National and European research infrastructures are being developed to support accessibility, reproducibility and sustainability for research data, methods and training.

Now, Artificial Intelligence is reshaping the landscape, offering both profound challenges and transformative opportunities. Closed datasets and AI models, inequalities in access to resources, and linguistic biases threaten to undermine Open Science’s core principles. Copyright issues, privacy risks, and environmental costs demand urgent attention. At the same time, Artificial Intelligence enhances research in the Digital Humanities in multiple ways, whether by automating rich annotations on large datasets, enabling insights into multilingual resources or democratizing reseachers’ ability to design algorithmic pipelines.

This keynote will explore how Europe’s Digital Humanities are navigating this complex intersection, leveraging the opportunities of Artificial Intelligence to advance research while safeguarding its ethical and inclusive foundations. The future of open knowledge production in the Humanities depends on our ability to bridge these worlds by fostering collaborative infrastructures, developing creative solutions, and maintaining scholarly best practices.

Overview

1. Introduction
2. Open Science
3. Artificial Intelligence
4. Research Infrastructures
5. Conclusion

This is what I would like to speak about today:

I will start with a quick reminder about some fundamental aspects of Open Science, including how Open Science is relevant to the Digital Humanities.

In a second step, I would like to discuss the role that Artificial Intelligence has recently taken on in Digital Humanities. In the process, we will see that it is quite challenging to bring Open Science and Artificial Intelligence together.

In the final step, I would like to argue that, at least in some ways, research infrastructures can provide essential help in this situation. They have been supporting researchers in Digital Humanities for a long time and are important advocates and supporters of Open Science. Now they can (and must!) support scholars in reconciling Open Science and Artificial Intelligence.

But let’s see all of this step by step, starting with Open Science.

Open Science
(and the Digital Humanities)

Aspects of Open Science

What this visualization shows is that Open Science is a lot more than Open Access, or even Open Data.

I think there are good reasons that often, we focus in particular on these two aspects. However, they are closely linked to other aspects as well.

• For example, Open Educational Resources are very important to disseminate competencies about best practices in research, including those relating to Open Science.
• Similarly, Open Infrastructures provide essential tools to support Open Access in scholarly publishing.

I think it is important to keep these relationships in mind, and not reduce Open Science to one particular aspect only.

But, I have said “open” up to now as if it was obvious what that meant. However, I think that it is not that obvious. In fact, “open” in the context of “Open Science” means many things:

• It is true, of course, that making resources “accessible” to others without financial barriers is fundamental. However, the “openness” extends much further.
• There should also. be as few technical or legal barriers as possible. This means that no proprietary tools should be necessary to access and use the resources.
• The data formats should be based on standards maintained by the community (for example XML-TEI for textual resources; or RDF for descriptive metadata).
• Whenever possible, widely-used identifiers, authority data and vocabularies should be used, for example from Wikidata or the relevant national library, to enable connections between datasets.
  
• Finally, openness is not just something technical, but also a value within a community, one that implies respect for others, recognition of difference, and a desire to create equity for all involved.

Source: Open Science Potsdam, uni-potsdam.de/en/openscience/.

Why practice Open Science?

Building on the essential principles of academic freedom, research integrity and scientific excellence, open science sets a new paradigm that integrates into the scientific enterprise practices for reproducibility, transparency, sharing and collaboration resulting from the increased opening of scientific contents, tools and processes.

– UNESCO, Recommendations on Open Science, 2021.

But why practice Open Science? The description provided by UNESCO in 2021 explains this:

“Building on the essential principles of academic freedom, research integrity and scientific excellence, open science sets a new paradigm that integrates into the scientific enterprise practices for reproducibility, transparency, sharing and collaboration resulting from the increased opening of scientific contents, tools and processes.”

I think this aptly describes why Open Science is worth aspiring to:

• Transparency means enabling others to understand exactly how we have come to the conclusions we present, based on what data and what methods;
• Reproducibility means enabling others to repeat our experiments and investigate the robustness or generalizability of our results.
• And I’d like to add sustainability to the list; meaning that we can avoid duplication of effort, and make sure our work remains useful in the long run, through Open Science practices.

Essentially, my point is that Open Science makes research better.

Sources: UNESCO (2021) and 10.5281/zenodo.1285574.

The Open Science Staircase

How does that work, you might ask. I try to answer this question with the following diagram.

The basic idea of this diagram is to show that, the more aspects of Open Science we consider, the more good practices of research we also enable.

• The baseline is publishing a PDF in a closed-access journal
• Going open access is the first step up, and enables everyone to participate in the scholary conversation, especially when the avenues for publishing are also open, that is, when journals for example do not charge publication fees.
• Also making the underlying research data available adds the ability for others to re-use the dataset for similar or different research. Data can be enhanced and re-shared, building on each others’ efforts in a sustainable way.
• Finally, also making the code or software used available adds the ability for others to actually replicate the findings from the publication using the real data and code. Often, this is a great first step towards enhancing the research as well.

Shamelessly adapted from 5stardata.info/en/ (CC0).

Open Humanities, or how does Open Science apply to the (Digital) Humanities?

• Open Access publications (85% of journals)
• Open Data: TextGrid Digital Library, ELTeC, DraCor, PoeTree
• Openness as scholar-led standards (XML-TEI)
• Open software: stylo, Gephi, Voyant, …
• Openness as a community value

But what are examples of these aspects in the Digital Humanities, and how widespread is Open Science here? I would argue that Open Science is really fundamental to DH.

• There are many Open Access journals; in fact, more than 85% of DH journals are Open Access. 48 out of 66 are Diamond Open Access (meaning: no publication fee), and another 9 are Gold Open Access (that is, with such a fee).
• Publishing datasets openly is also the standard for many projects; just think of corpora like TextGrid Digital Library, ELTeC, DraCor, or PoeTree.
• Many important pieces of software are, of course, released as open source software: for example stylo, Gephi or Voyant.
• Shared and community-driven standards are also essential for DH. The foremost example is of course the Text Encoding Initiative (TEI).
• The are several platforms for Open Educational Resources, like “The Programming Historian” or “DARIAH-Teach”.
• Finally, openness is also a fundamental value of the community, in the sense that accessibility and equity are important considerations when organizing activities or events.

Sources: Vaucher and Thomas (2026) and mapping-dh.wikisource.cloud.
Open corpora: ELTeC, DraCor, PoeTree, TextGrid Repository.

Case study: “Mining and Modeling Text”

• Open Access, e.g. Hinzmann et al. (2026)
• Open Dataset: Corpus of French Novels (in XML-TEI)
• Open Standards: Linked Open Data (RDF, Dublin Core, etc.)
• Open Infrastructure: Wikibase, Wikidata, OpenRefine
• Open Licences: Creative Commons
• Open Educational Resources: Tutorial

To make things a little bit more concrete, I’d like to briefly speak about a project I have led until recently. This project is called “Mining and Modeling Text” and was concerned with a evidence-based history of the French novel in the 18th century. We tried to be as open as possible in everything that we did:

• All of our publications appeared in Open Access.
• All of our data, or almost, is available as open data, using community standards; for example a corpus of 200 French novels in XML-TEI
• Beyond TEI, we also used standards relevant to the project database, which follows Linked Open Data princples.
• Of course we used open source software and infrastructure, among them Wikibase and Open Refine, and linked our data to Wikidata and other authority files.
  
• All of our data is also licensed openly, that is using Creative Commons licenses.
  
• And we even created a SPARQL tutorial that is available as an Open Educational Resources. In this project, we learned to respect the care and effort that is required to make the ideal of Open Science a reality. In particular, there are several challenges we recognize:
• Practicing Open Science often means additional effort, and may slow us down, individually. At the community level, however, it is more efficient.
• Also, there are often issues like copyright law and personal data protection to be considered;
• And of course there is still limited prestige associated with, for example, dataset publication, so the proper reward structures are not in place.

Of course I think these are challenges that we can and must address.

But how is this situation affected by the advent of Artificial Intelligence? That’s what we’ll see next.

See: mimotext.uni-trier.de. Background: Schöch et al. (2022), Hinzmann et al. (2026).

The Advent of Artificial Intelligence in Digital Humanities

A timeline of recent developments in AI

• 2003: Topic Modeling
• ~2006: Deep Learning / Neural Networks become competitite
• 2012: Static word embeddings (word2vec)
• 2017: Contextual word embeddings (BERT)
• 2018: Encoder-decoder models (GPTs)
• ~2023: Multimodal models become popular
• 2024: Agentic models (MCP)

There is no time to talk in depth about this timeline. I just want to make a few observations about how these developments have been taken up by Digital Humanities.

• The DH community has been a leader and a pioneer in methods and standards for text encoding, with early work around the “Text Encoding Initiative”, still in the SGML context, strongly influencing the subsequent development of XML.
• The same is not true for Machine Learning and Artificial Intelligence. In this domain, the story is different: fundamentally, Digital Humanities have always adopted and adapted methods from Computer Science for their purposes, with more or less pronounced time lags. It took about eight years before LDA Topic Modeling, first introduced by David Blei in 2003, became integrated into Digital Humanities.
• Recently, the rhythm of adoption has accelerated: Generative LLMs or agentic AI have seen pionieering adoptions in various areas of Digital Humanities within one year or less.

We will see in a minute that this rhythm of increased adoption speed has its limits. In historical comparison, we can afford to slow down a little. From my perspective, the most notable game-change, in recent times, was BERT models, first coming out in 2017. And we have not yet exhausted the potential of BERT models for Digital Humanities.

Another point I’d like to make about this timeline is that, remarkably, it is all covered by the term “Artificial Intelligence”. This term has become an umbrella term, and, to be honest, a marketing term.

This shows an urgent need to differentiate and to be very clear what we are actually talking about, when talking about AI in our research. At the very least, we should be distinguishing between the following kinds of AI:

• “Traditional”, that is, statistical machine learning.
• Word Embeddings, including transformer-based, BERT-style embeddings.
• Generative Large Language Models for text
• Multimodal LLMs
• And agentic LLMs

Note that none of these categories include “real” AI in the strong sense of autonomous, embodied agents experiencing and interacting with the real world and learning from this interaction!

Applications of AI in Digital Humanities

• OCR and HTR for full-text digitization, e.g. of historical documents (Transkribus)
• Text encoding for Digital Scholarly Editions (Scholger, Strutz, and Pollin 2024)
• Stylometric Authorship Attribution (Hicke and Mimno 2023)
• Text annotation in Computational Literary Studies (Brunner et al. 2020)
• Image annotation, e.g. in Digital Art History (Arnold and Tilton 2024)
• Text classification in Computational Literary Studies (Keith et al. 2026)
• Coding support for a variety of applications

Whatever the case may be, it is notable that various kinds of LLMs have been quickly adopted in the Digital Humanities in recent years. Many if not all areas of DH have been involved:

• OCR and HTR for full-text digitization, e.g. of historical documents (think of Transkribus)
• Text encoding for Digital Scholarly Editions
• Authorship attribution (for short texts)
• Text annotation in Computational Literary Studies
• Image annotation, e.g. in Digital Art History
• Text classification in Computational Literary Studies
• Coding support for all kinds of applications, from website building to SPARQL queries.

Case study: Historical Wine Labels

To quickly illustrate how AI can radically scale up annotations of complex materials, I’d like to briefly speak about a project we are currently running at the Trier Center for Digital Humanities, on Historical Wine Labels.

Here you see the wine labels themselves, from our collection of about 5000 such labels. Early ones, from the nineteenth century…

Case study: Historical Wine Labels

… and more recent ones, from the early to later twentieth century.

Multimodal models to generate image descriptions

• Small, local, open-weights model (Gemma4-12B)
• Low hardware requirements (12GB GPU, runtime 60s)
• Description sufficient for search / retrieval

For the moment, we have followed two approaches to enhancing the discoverability of these labels in our database, using a multimodal LLM:

• A description in prose, using a prompt centered around the text, image objects, style, and context. You see the result here. It is not a full interpretation of the label, but it can support search.

Controlled vocabulary + structured results

• Multilingual taxonomy with authority data
• Output in pre-defined JSON structure

{
  "identifier": "SMW-D-E124",
  "designation": "1918er Maringer Rosenberg",
  "objects": [
    {"wid":"Q941", "eng":"man", "deu":"Mann"},
    {"wid":"Q939", "eng":"woman", "deu":"Frau"},
    {"wid":"Q938", "eng":"wine glass", "deu":"Weinglas"},
    {"wid":"Q881", "eng":"grape", "deu":"Weintraube"},
    {"wid":"Q879", "eng":"wine leaf", "deu":"Weinblätter"},
    {"wid":"Q875", "eng":"flower", "deu":"Blume"},
    {"wid":"Q966", "eng":"abstract form", "deu":"Abstrakte Form"},
    [...]
  ]
}

And a more formalized annotation of the objects depicted in visual form. Here, we use a multilingual, controlled vocabulary as the basis for annotation, as well as a structured output format in JSON for easy ingest into the database.

Result: Map and database of historical wine labels

The resulting data is modeled as Linked Open Data and integrated into a Wikibase instance.

This enables us to connect the labels to contextual information, such as wine makers, wine making villages, and vineyards, and enable searching and visualizing the data.

See: mosel.wikibase.cloud and Weis and Schöch (2024).

Challenges and Lessons Learned (especially for Open Science)

• BERT-models and generative LLMs are a game changer for research in DH
• They should only be used for tasks that can be evaluated

• However: LLMs challenge Open Science
  • High requirements for access (financial / hardware)
  • Models are not truly open: limited reproducibility
  • Models are driven by commercial interests, not scholarly needs

• Additional concerns
  • Environmental impacts
  • Respect of copyright law
  • Protection of personal data
  • Bias towards English language and culture

Consequence: Use small, local, open-weight models whenever possible

The wine label project is just one among several projects in which we have recently gained experience with using LLMs for research in Digital Humanities. Based on these experiences, I’d like to formualte a few lessons that we have learned so far, especially with respect to Open Science.

• LLMs can enhance and enrich our access to the research data, and they are a real game-changer in that sense. But they can not, and should not, be expected to answer research questions on their own.
• I am aware that agentic systems are challenging this view. However, my second point still puts a limit on this approach: I think LLMs should only be used for tasks that can be clearly evaluated, so that we know exactly how well, and how reliably, a certain task is performed by the LLM.

With respect to Open Science, and its focus on accessibility, transparency and replicability, the new very large language models are a huge challenge:

• The financial hurdles to access them are significant, whether that is API costs or investments in hardware. This is already reinforcing a “digital gap”, both between smaller and larger Higher Education Institutions within a given country, and between the educational systems of countries of difference economic power.
• Transparency has always been an issue with statistical machine learning. This is significantly increased with deep learning and generative LLMs that have highly complex architectures where feature weights are difficult to extract.
• Similarly, with respect to replicability, the fact that API-based models may change at short notice, and older models may not be available any more, means that replicability is strongly limited. In addition, the very high sensitivity of such models to even rather small changes in the prompts mean they are not as robust as we would like them to be for replicability.
  
• There are additional concerns as well: The ecological cost of deploying ever larger models on ever more powerful hardware running in ever larger datacenters, is something that we need to see critically. Just a few days ago, the Guardian reported that the annual energy use of Microsoft, Meta and Google in 2025, according to their own sustainability reports, corresponds to half of the annual energy use in a country of the size of France, with more than 60 million inhabitants!

The good news is that we don’t always need to rely on these kinds of models. Remember, the wine label annotations were done with Gemini4, a small model that runs on a desktop computer!

So it doesn’t always have to be the latest and largest model. Progress is impressively quick, and trickles down within months from the frontier models to smaller, open-weight models that can be deployed locally on moderately-powerful hardware. Admittedly, for some particularly complex tasks, it still seems true that “bigger is better”. But certainly not always!

In fact, the most fundamental game changer, at least for my own research in Computational Literary Studies, have been BERT models, that is, non-generative models that came on the scene almost 10 years ago!

• They enable access to texts that goes beyond counting word forms, to handling meaning in context.
• They can be fine-tuned for all kinds of annotation tasks.
  
• They exist as multilingual models or as models for specific languages (e.g. ModernGBERT).
• And they are not very resource-hungry, run very quickly, can be deployed locally, and are often more predictable than the generative models.

In any case, my point is: It is important not to just go to the latest, biggest model by default, but balance performance and resource requirements every time, depending on the task at hand. And smaller, open-weight models that can be used on local hardware have many advantages in terms of Open Science.

Research infrastructures for Digital Humanities (in the context of AI and Open Science)

The question now is, what roles research infrastructures have to play in the context of Open Science and Artificial Intelligence. And in particular, how they might help researchers overcome the conflict between OS and AI.

National Research Infrastructures: Germany

I start with an outline of some if the key research infrastructures in Germany, before moving on to the European level.

The most important player in Germany here is certainly the “German National Research Data Infrastructure”, NFDI. It consists of 30 domain-specific consortia. The government has just recently, a few days ago, decided to continue funding it until 2038 with almost 100 million EUR per year. So it is a massive effort.

There are four consortia that support the humanities: NFDI4Culture, 4Objects, 4Memory and Textplus.

In addition, and especially in the context of AI, another national infrastructure seems important: The “NHR”, meaning “National High-Performance Computing”. It is a network of data centers that researchers in Germany can use when their local computing resources are not sufficient for their needs.

But let’s move on to the European level, where things are quite a bit more complex.

German National Research Data Infrastructure (NFDI): nfdi.de and High-Performance Computing Network (NHR): nhr-verein.de.

European-Level Resarch Infrastructures for SSH

• Research Infrastructures
  • DARIAH-EU (Digital Research Infrastructure for the Arts and Humanities)
  • CLARIN-EU (Common Language Resources and Technology Infrastructure)
  • OPERAS (Open Scholarly Communication in the European Research Are for HSS)
  • OpenAIRE (Open Scholarly Communication Infrastructure)
• Specific Services
  • ORE (Open Research Europe): mega-journal platform)
  • Zenodo: long-term data deposit
  • EOSC (European Open Science Cloud): collaboration platform

I’ve listed some of the major European-level research infrastructures here that are relevant to the humanities. There are many more, and often they also have national nodes or centers.

• DARIAH-EU, the Digital Research Infrastructure for the Arts and Humanities
• CLARIN-EU, with a focus on linguistic resources
• OPERAS, which supports open access publishing in the Humanities and Social Sciences
• OpenAIRE, which runs open scholarly communication infrastructure

And I would also like to highlight some specific services offered at the EU level:

• Open Research Europe, which is a mega-journal platform for open access publications.
• Zenodo, a brilliant service for long-term data deposits, run by CERN in Geneva.
• And EOSC, the European Open Science Cloud, which is a collaboration platform that also offers computing resources

Case Study: European Open Science Cloud (EOSC)

Just to give you a quick impression of how EOSC looks like, this is the dashboard screen of the platform, where you can access various services. For example file-sharing, iPython notebooks, or CPU and GPU computing resources.

How RI support Open Science and Artificial Intelligence

• Promote FAIR principles (NFDI, DARIAH)
• Enable and support Open Access Publishing (OER, OPERAS)
• Offer digital collaboration tools: filesharing, notebooks, CPU/GPU resources (EOSC)
• Offer dedicated high-performance computing resources (NHR)
• Enrich, improve, and make available data resources (DARIAH, CLARIN)
• Offer training materials and opportunities (DARIAH, NFDIs)
• Advocate for the needs of the SSH
• Serve as community hubs and DH incubators
• Perform basic research for infrastructure development

Of course, this is just one way in which research infrastructures support Open Science in the Humanities. Generally speaking, one can say that they are active in the following domains:

• They promote FAIR principles
• They enable and support Open Access Publishing
• They offer digital collaboration tools: filesharing, notebooks, CPU/GPU resources
• They offer dedicated high-performance computing resources (NHR)
• They enrich, improve, and make available data resources (DARIAH, CLARIN)
• They offer training materials and opportunities (DARIAH, NFDIs)
• They advocate for the needs of the SSH
• They serve as community hubs and DH incubators
• And they Perform basic research for infrastructure development

For this last point, fundamental research, I’d like to offer one more case study, this time regarding solutions for copyright issues in research.

Case Study on copyright: The ‘window of opportunity’ effect

• Positive impulses
  • Open Science
  • Academic Freedom
• Constraining factors
  • Copyright Law
  • OCR Technology

Basically, copyright law brings Open Science and Academic Freedom into a conflict that seems hard to solve:

• We want to practice Open Science by making our corpora publicly available
• We want to freely chose our object of analysis, e.g. contemporary French novels (in my case)

But then, we run into the reality of the “window of opportunity” effect: - The state of OCR technology means that there are a lot less materials, still, for the period ca. pre-1800 - And while copyright law allows us to build and use in-copyright corpora, it does not allow us to publish them

The good news is: OCR is constantly improving, and there is at least a partial solution for the copyright problem.

Solution with ‘Derived Text Formats’

Derived Text Formats (DTFs) are the result of a strategic transformation of textual materials that are protected by copyright in their original form, such that the resulting data is useful for computational analyses and can be openly shared following best practices of Open Science without infringing copyright law.

• Four steps to create a DTF
  • Legal access to originals
  • Information enrichment
  • Information reduction
  • Open sharing possible

That solution can be called “Derived Text Formats”. I’ve recently described them in the following terms:

“Derived Text Formats (DTFs) are the result of a strategic transformation of textual materials that are protected by copyright in their original form, such that the resulting data is useful for computational analyses and can be openly shared following best practices of Open Science without infringing copyright law.”

It is actually a rather straightforward process to create a DTF:

• Legal access to the in-copyright textual data is required, according to European copyright law.
• Then, you can do some enrichment, for example through token-level annotations or syntactic parsing.
• However, you also need to remove some information, in order to make sure the data doesn’t fall under copryright anymore: for example by randomising the word order; or by masking some of the word forms

And you’re done! You can publish your in-copyright data.

Further background: Schöch (2026), Trippel, Barth, Calvo Tello, et al. (2026).

Example from Sherlock Holmes

original

Mr. Sherlock Holmes, who was usually very late in the mornings, save upon those not infrequent occasions when he was up all night, was seated at the breakfast table. […]

(Arthur Conan Doyle, The Hound of the Baskervilles, 1902)

annotated

Mr._Mr._NNP
Sherlock_Sherlock_NNP
Holmes_Holmes_NNP
,_,_PUNC
who_who_WP
was_be_VBD
usually_usually_RB
very_very_RB
late_late_JJ
in_in_IN
the_the_DT
mornings_morning_NNS ,_,_PUNC
…

randomized

save_save_IN
,_,_PUNC
who_who_WP
infrequent_infrequent_JJ
table_table_NN
very_very_RB
occasions_occasion_NNS
he_he_PRP
the_the_DT
up_up_RP
was_be_VBD
usually_usually_RB
Mr._Mr._NNP
…

Here you see the first two sentences of a famous novel by Arthur Conan Doyle.

• In the original, on the left.
• Then with annotation.
• And then with word order randomized (this would be done in 1000-token segments, to preserve some global structure).

Of course, DTFs can take many forms. Another very useful DTF is a simple term-document matrix. In every case, you can still do quite a lot of analyses with DTFs, like Topic Modeling or Stylometric Authorship Analysis. But you can’t read it, so it is not protected by copyright.

Current research on DTFs

• Usefulness of DTFs
  • Stilometric Authorship Attribution (Du 2023)
  • Topic Modeling (Kocula 2021)
  • Sentiment Analysis (Du and Schöch 2024)
  • DTF for fine-tuning LLMs (Du and Schöch 2024)
• Reconstructibility
  • InvBERT (Kugler et al. 2024)
  • Fine-tuned T5 (Du et al. 2025)
    
• Standardization
  • DTFs as formal standard (Trippel, Barth, Tello, et al. 2026)
  • Certification of the pipeline (in progress)

There is actually quite a bit of research into DTFs going on recently in CLS:

For example, how useful are specific DTFs for a particular method of analysis?

• Stilometric Authorship Attribution
• Topic Modeling
  
• Sentiment Analysis
• Fine-tuning LLMs for specific tasks

How reliably can LLMs reconstruct the original full text from a DTF?

• Inverting a BERT-model
• Using a Fine-tuned T5
• How can we standardize DTFs, so that they can be trusted?
• Parameters of DTFs (Trippel, Barth, Tello, et al. 2026)
• Certification of the pipeline (in progress)

50,000 in-copyright, popular novels held by the German National Library will be published in a DTF format very soon.

In the sense that this work is fundamental research aiming to make better research possible, I think it makes a lot of sense that we are doing this work in the framework of the NFDI, the German National Research Data Infrastructure.

Conclusion

But it is time to conclude.

Open Science, Large Language Models, and Research Infrastructures

• Open Science remains a cornerstone for excellent research
• LLMs are a game-changer for research in the Digital Humanities
• Very large, web-based LLMs are largely incompatible with Open Science
• Research infrastructures are already doing essential work for Open Science
• Now they also need to train open models and provide computing resources
  
  
• Research infrastructures can bring Open Science and Artificial Intelligence together

I want to conclude, very briefly, by simply reiterating a few more general points I hope to have illustrated in my talk:

• Open Science remains a cornerstone for excellent research; it ensures transparent, reproducible, sustainable, “convivial” research.
  
• LLMs are highly valuable to the Digital Humanities, because they enable context-sensitive, meaningful annotations of scholarly objects at the scale we need in Digital Humanities
• However, very large, API-only LLMs are the antithesis of open Science: the involve financial and technical hurdles, the pose substantial challenges to reproducibility, they are highly problematic from an ecological perspective); and they are corporate products that are not under scholarly control.

This is where, in my opinion, research infrastructures come into play:

• To safeguard Open Science in Digital Humanities when using LLMs, we should be using medium-sized, open models that we can deploy locally whenever possible. This also helps with copyright and personal data protection as well as with energy demand.
• However, training models requires huge amounts of data and computing resources, more than any single researcher or institution can provide. Research infrastructures can leverage their network of libraries and computing centers to overcome this hurdle.
• In parallel, we will also need adequate hardware to run these open, local, mid-sized models: It doesn’t always have to be massive high-performance computing cluster, but national research infrastructures should also help set up a distributed infrastructure of mid-sized computing services where researchers can run small to mid-sized models with support from local staff.

This means, and this is my final remark: We need well-equipped research infrastructures that will enable us, in the Digital Humanities, to make the best use of Large Language Models while continuing to follow best practices of Open Science!

Thank you!

Thank you.

References

Arnold, Taylor B., and Lauren Tilton. 2024. “Explainable Search and Discovery of Visual Cultural Heritage Collections with Multimodal Large Language Models.” CHR2024. https://ceur-ws.org/Vol-3834/paper1.pdf.

Brunner, Annelen, Ngoc Duyen Tanja Tu, Lukas Weimer, and Fotis Jannidis. 2020. “To BERT or Not to BERT - Comparing Contextual Embeddings in a Deep Learning Architecture for the Automatic Recognition of Four Types of Speech, Thought and Writing Representation.” In Proceedings of the 5th Swiss Text Analytics Conference (SwissText) & 16th Conference on Natural Language Processing (KONVENS), Zurich, Switzerland, June 23-25, 2020. https://ceur-ws.org/Vol-2624/paper5.pdf.

Da, Nan Z. 2019. “The Computational Case Against Computational Literary Studies.” Critical Inquiry 45 (3): 601–39. https://doi.org/10.1086/702594.

Du, Keli. 2023. “Understanding the Impact of Three Derived Text Formats on Authorship Classification with Delta.” In Jahreskonferenz der Digital Humanities im deutschsprachigen Raum. DHd-Verband. https://doi.org/10.5281/zenodo.7715298.

Du, Keli, Sarah Ackerschewski, Uygar Navruz, Nazan Sinir, Julian Valline, and Christof Schöch. 2025. “Reconstructing Shuffled Text: Bad Results for NLP, but Good News for Using in-Copyright Texts.” Journal of Computational Literary Studies 4. https://doi.org/10.48694/jcls.4163.

Du, Keli, and Christof Schöch. 2024. “Shifting Sentiments. What Happens to BERT-Based Sentiment Classification When Derived Text Formats Are Used for Fine-Tuning.” In Digital Humanities 2024: Book of Abstracts. ADHO.

Fraser, Nicholas, Fakhri Momeni, Philipp Mayr, and Isabella Peters. 2020. “The Relationship Between bioRxiv Preprints, Citations and Altmetrics.” Quantitative Science Studies 1 (2): 618–38. https://doi.org/10.1162/qss_a_00043.

Fu, Darwin Y., and Jacob J. Hughey. 2019. “Releasing a Preprint Is Associated with More Attention and Citations for the Peer-Reviewed Article.” eLife 8: e52646. https://doi.org/10.7554/eLife.52646.

Hicke, Rebecca M. M., and David Mimno. 2023. “T5 Meets Tybalt: Author Attribution in Early Modern English Drama Using Large Language Models.” https://arxiv.org/abs/2310.18454.

Hinzmann, Maria, Matthias Bremm, Tinghui Duan, Anne Klee, Johanna Konstanciak, Julia Röttgermann, Moritz Steffes, Christof Schöch, and Joëlle Weis. 2026. “Patterns in Modeling and Querying a Knowledge Graph for Literary History.” In Patterns in Language and Communication, edited by Sabine Arndt-Lappe, Sören Stumpf, Milena Belosevic, Peter Maurer, Claudine Moulin, and Achim Rettinger, 135–76. De Gruyter. https://doi.org/10.1515/9783111199740-006.

Huang, Chun-Kai, Cameron Neylon, Lucy Montgomery, Richard Hosking, James P. Diprose, Rebecca N. Handcock, and Katie Wilson. 2024. “Open Access Research Outputs Receive More Diverse Citations.” Scientometrics 129 (2): 825–45. https://doi.org/10.1007/s11192-023-04894-0.

Keith, Allison, Antonio Rojas Castro, Kerstin Jung, Hanno Ehrlicher, and Sebastian Padó. 2026. “A Computational Analysis of Character Archetypes in the Works of Calderón de La Barca.” Journal of Computational Literary Studies 5. https://doi.org/10.48694/jcls.4167.

Kocula, Martin. 2021. “Volltext vs. abgeleitetes Textformat: Systematische Evaluation der Performanz von Topic Modeling bei unterschiedlichen Textformaten mit Python.” Trier University. https://doi.org/10.5281/zenodo.5552486.

Kugler, Kai, Simon Münker, Johannes Höhmann, and Achim Rettinger. 2024. “InvBERT: Reconstructing Text from Contextualized Word Embeddings by Inverting the BERT Pipeline.” Journal of Computational Literary Studies 2 (1). https://doi.org/10.48694/jcls.3572.

Langham-Putrow, Allison, Caitlin Bakker, and Amy Riegelman. 2021. “Is the Open Access Citation Advantage Real? A Systematic Review of the Citation of Open Access and Subscription-Based Articles.” Edited by Sergi Lozano. PLOS ONE 16 (6): e0253129. https://doi.org/10.1371/journal.pone.0253129.

Schöch, Christof. 2023. “Repetitive Research: A Conceptual Space and Terminology of Replication, Reproduction, Revision, Reanalysis, Reinvestigation and Reuse in Digital Humanities.” International Journal of Digital Humanities. https://doi.org/10.1007/s42803-023-00073-y.

Schöch, Christof. 2026. “Derived Text Formats as Strategic Transformations of in-Copyright Materials to Support Open Science: A Survey.” In Leveraging Derived Text Formats to Unlock Copyrighted Collections for Open Science: Book of Abstracts, edited by Trippel, Thorsten, Barth, Florian, Tello, Jose Calvo, Genêt, Phillipe, Lendvai, Piroska, and Schöch, Christof. https://api.zotero.org/users/228821/publications/items/3RZJJCIK/file/view.

Schöch, Christof, Maria Hinzmann, Julia Röttgermann, Katharina Dietz, and Anne Klee. 2022. “Smart Modelling for Literary History.” International Journal of Humanities and Arts Computing 16 (1): 78–93. https://doi.org/10.3366/ijhac.2022.0278.

Scholger, Martina, S. Strutz, and Christopher Pollin. 2024. “Empowering Text Encoding with Large Language Models: Benefits and Challenges.” In TEI Members’ Meeting and Conference 2024. https://doi.org/10.5281/zenodo.13969082.

Trippel, Thorsten, Florian Barth, José Calvo Tello, Phillipe Genêt, Lendvai, Piroska, and Christof Schöch, eds. 2026. Leveraging Derived Text Formats to Unlock Copyrighted Collections for Open Science. LREC.

Trippel, Thorsten, Florian Barth, Jose Calvo Tello, Keli Du, Phillipe Genêt, Daniel Kurzawe, Peter Leinen, Piroska Lendvai, Christof Schöch, Andreas Witt, and Arden Zimmermann. 2026. “DIN 19461: A National Standard for Derived Text Formats.” In Leveraging Derived Text Formats to Unlock Copyrighted Collections for Open Science (Workshop at LREC 2026), edited by Trippel, Thorsten, Barth, Florian, Tello, Jose Calvo, Genêt, Phillipe, Lendvai, Piroska, and Schöch, Christof.

UNESCO. 2021. “UNESCO Recommendation on Open Science.” UNESCO. https://doi.org/10.54677/MNMH8546.

Vaucher, Romain, and Camille Thomas. 2026. “Diamond Is the New Green — Why Green Open Access Is Not a Sustainable Long-Term Model for Scientific Publishing.” Sedimentologika 4 (1). https://doi.org/10.57035/journals/sdk.2026.e41.2397.

Weis, Joëlle, and Christof Schöch. 2024. “Vom Perler Hasenberg zur Lehmener Würzlay – Weinetiketten digital Erschließen.” In Digital ist besser? Sammlungsforschung im digitalen Zeitalter, edited by Katharina Günther and Stefan Alschner. Göttingen: Wallstein. https://api.zotero.org/users/228821/publications/items/KL289SR5/file/view.

Bonus slides

The Open Access Citation Advantage

Citations per year / accesses per month, special issue in International Journal of Digital Humanities, 2023/2024. n=14, cutoff: May 2026.
See also: Fu and Hughey (2019), Fraser et al. (2020), Langham-Putrow, Bakker, and Riegelman (2021), Huang et al. (2024).

What is ‘Repetitive Research’?

• A mode of research that repeats and/or is repeatable
• A conceptual space with three dimensions: research question, dataset, and method/code

• Several forms of repetitive research
  • Follow-up research: similar question, but new or similar data and method
  • Re-use & re-analysis (of data): same data
  • Revision (of method): same question and data
  • Replication (of research): same question, data and method

For a deep dive, see: Schöch (2023). See also: Da (2019).