• Georgiana D. Hedesan (Dr) University of Oxford

University of Oxford

Started in 1998, the Newton Project (https://www.newtonproject.ox.ac.uk/) was one of the first initiatives to provide full texts of historical documents online and in open access. Ever since, the Project has amassed around 12 million words that were authored by Isaac Newton or closely related to him. In connection to the tercentenary of Isaac Newton’s death in 2027, an ambitious new project has started under the aegis of the Newton Project: the production of an AI-enabled digital edition of the three versions of Newton’s masterpiece Naturalis Philosophiae Principia Mathematica (1687, 1713, 1726), together with material relating to their conception and revision. One of the chief milestones of the Newton Principia Project is the transcription of manuscripts related to the Principia and housed at the Cambridge University Library, and in particular MS 3965. This manuscript comprises c. 1530 pages, most of it handwritten by Isaac Newton himself. While some of these are written in relatively fair hand, the majority are drafts that contain heavy corrections, tables and calculations, some of which are jotted down almost chaotically on any piece of free paper that Newton could find. 

The proposed presentation focusses on our design of the optimal MS 3965 transcription workflow, involving both automatic transcription and manual correction. The design involved the empirical testing of different transcription AI-based models before settling on a bespoke solution involving two generative AI (Anthropic Claude and Google Gemini Pro 3.0), with incorporation of the LaTex-language outputs of a math-specialised LLM (MathPix). Claude was employed as virtual assistant and computer vision specialist, while Gemini was used for its transcription capacity. By interacting with Claude and Gemini, we were able to create a 5-tier classification system for the pages in the manuscript, together with specific workflow for each of the tiers. The complexity of the transcription task and the possibilities of generative AI further led to the AI-assisted construction of a bespoke transcription platform that optimises the interactions between human and the AIs and models used, while capturing the complexity of the manuscript pages (multiple corrections, changes, stains and blots, deteriorated paper, complex mathematical notation). The platform is designed in view of creating TEI XML code that includes genetic markup and genetic encoding. We consider that the classification system and the transcription platform could in due course provide a methodology for triaging manuscript pages, transcribing and capturing them in complex TEI-XML standard.

 

• Jamie McLaughlin University of Sheffield

University of Sheffield

“Marking up” text, specifically the process of assigning metadata to text spans in order to encode meaning, has been central to digital humanities practice for decades. Thousands of person-hours have been devoted to this task using standards such as TEI, either performing the process manually or sometimes partially assisted using conventional NLP software.

 

Huge advances in neural networks and deep learning have profoundly impacted NLP since the early 2010s. This has greatly expanded the range of language tasks which can be performed automatically or semi-automatically. In particular, deep learning models known as “Transformers”, using a multi-head “attention” mechanism [1] have improved the performance of tasks such as Named Entity Recognition (NER) and Relation Extraction (RE) to the point where markup tasks previously performed by human researchers can potentially be automated or partially automated. This offers huge cost-savings for DH projects. But more importantly it increases the size of the text collections which can be processed and therefore the scale and scope of feasible research questions.

 

The DHI in Sheffield has been involved in many large scale projects over the past decades which involved extensive human markup work. We are now beginning to work on projects which perform markup using deep learning models. We have insights to share about which tasks are better suited to, for example, encoder-only models such as BERT, and which are better suited to the generative Large Language Models (“LLMs”) which have become synonymous in the popular consciousness with “AI”. We have developed strategies for producing the most accurate results from locally run smaller models rather than remotely hosted commercial services such as ChatGPT or Gemini. For example, the absolute minimum number of ground truth human examples required to fine tune a BERT model for good results in a particular domain, or using spans generated from a partially trained encoder to in turn formulate more effective LLM prompts.

 

This paper aims to demystify what is possible with deep learning in our domain, where material is highly specialised, markup requirements atypical and appropriate training data often scant or nonexistent. With respect to completed projects, the DHI has amassed a huge tranche of training and test data over the years against which we can evaluate deep learning approaches. For new projects, we are quickly learning how conventional wisdom about “AI”, often “throw everything at a commercial LLM”, can be a flawed strategy.

 

Deep learning should allow digital humanities scholars to allocate more time to research and less to data processing and markup. This paper aims to help researchers fulfil the promise of “AI”; that we do less of the repetitive work we want to avoid and more of the higher thinking and creating that defines us.

---

1. A. Vaswani et al., "Attention is all you need," in Advances in Neural Information Processing Systems, vol. 30, 2017, pp. 5998–600

• Martina Scholger University of Graz,
• Sabrina Strutz University of Graz

University of Graz

The transformation of historical correspondence from print to TEI-compliant digital editions requires systematic person identification and authority linking—a process traditionally dependent on extensive manual labour and domain expertise. This paper presents a hybrid workflow combining retrieval-augmented generation (RAG) with authority database reconciliation to semi-automate person identification while maintaining scholarly standards. Our case study is the correspondence of Joseph von Hammer-Purgstall (1774–1856), with 3,500 transcribed and annotated letters.

The corpus was recently converted from a print edition (Höflechner et al. 2018) into TEI-XML digital scholarly edition (Höflechner et al. 2025). While the encoding captures the structural components of the letters, editorial interventions remain largely unaddressed. Person references remain encoded as <persName> tags without authority identifiers, while contextual information resides in unstructured editorial notes. This represents a common challenge in correspondence digitization: moving beyond mere structural markup toward what Sahle (2016) defines as “digital editions”, which exploit the capabilities of the digital, rather than replicating the print paradigm.

Historical correspondence presents challenges for this transformation. The Hammer-Purgstall collection exemplifies these through variant spellings (Photius vs Photios), abbreviated titles and names (Freih[errn] <persName>v[on] Buol</persName>), split annotations (<persName>v[on]</persName> <persName>Müller</persName>), partial annotations (<persName>Abu‘l-Fida</persName> Imad ad-Din Ismail ibn Ali al-Ayubi), and implicit contextual information in editorial notes attached to <persName> elements. Our goal is to develop an LLM-based approach to resolve the annotation errors by capturing full names (including forenames and roles), and linking them to authority files. 

Amid ongoing debates on the role of generative AI in digital scholarship  (Pollin et al. 2025), the transformation of print editions into “digital editions” constitutes a primary application area for LLMs in the editorial workflow (Pollin et al. 2024; Strutz 2026; Strutz forthcoming). Unlike rule-based systems, LLMs can apply contextual reasoning and semantic interpretation to make informed judgments based on textual evidence.

Our methodology implements a multi-stage RAG pipeline that progressively resolves person references through different matching strategies. <persName> tags in TEI-annotated letters are extracted with surrounding textual context. In a first stage, these entries are queried against an existing persons index represented in a vector database, resolving straightforward matches. The system then enters an iterative cycle: unmatched names are submitted to the GND (Gemeinsame Normdatei) via the lobid API, with LLM-based disambiguation leveraging letter context to select among multiple candidates (Strutz 2026). Successfully disambiguated persons are added to the vector database, and remaining persons are re-queried against this expanded index, iterating until resolution falls below a meaningful threshold. At this point, unmatched persons are queried against a vector database of the full correspondence corpus, exploiting cross-letter contextual patterns to identify individuals through co-occurrence and descriptive mentions. Finally, the remaining cases are flagged for human reconciliation.

This hybrid architecture addresses critical limitations in existing approaches. Simple string matching fails on variant spellings; pure LLM approaches risk hallucinating authority identifiers; population-based heuristics ignore historical context. By separating RAG retrieval from authority linking and employing LLMs only for contextual disambiguation, we minimize hallucination risks while maximizing computational efficiency and scholarly transparency.

We evaluate this workflow through comparative testing across proprietary models (GPT-5.2, Claude Sonnet 4.6) and open-source alternatives (OLMo2-32B-Instruct, Qwen3-14B), addressing sustainability concerns in DH infrastructure. This cross-model comparison contributes to community discussions about balancing performance with long-term reproducibility requirements.

This work contributes to emerging editorial practices by demonstrating how generative AI can complement human editorial work in enriching structurally digitized texts with the semantic depth characteristic of genuine digital editions, recognizing that as transformer-based LLMs integrate into digital scholarship, their capabilities and limitations must be critically assessed.

 

Deutsche Nationalbibliothek (n.d.) Gemeinsame Normdatei (GND). https://www.dnb.de/DE/Professionell/Standardisierung/GND/gnd_node.html. 

Höflechner, Walter, Alexandra Wagner, and Gerit Koitz-Arko (2018): Joseph von Hammer-Purgstall. Briefe, Erinnerungen, Materialien. Graz: ADEVA. 

Höflechner, Walter, Martina Scholger, Alexandra Wagner, Sabrina Strutz, Elisabeth Steiner (eds.) (2025): Joseph von Hammer-Purgstall. Korrespondenz. Graz. http://gams.uni-graz.at/hpe, http://hdl.handle.net/11471/559.20.

 

Pollin, Christopher, Franz Fischer, Patrick Sahle, Martina Scholger, and Georg Vogeler (2025): When it was 2024 – Generative AI in the Field of Digital Scholarly Editions. In: Zeitschrift für digitale Geisteswissenschaften 10 (2025). 10.07.2025. HTML / XML / PDF. DOI: 10.17175/2025_008.

 

Sahle, Patrick (2016): “What is a Scholarly Digital Edition?” In Driscoll, Matthew James, and Elena Pierazzo (eds.), Digital Scholarly Editing. Cambridge: Open Book Publishers. https://books.openedition.org/obp/3397. 

 

Scholger, Martina, Sabrina Strutz, and Christopher Pollin. 2024. “Empowering Text Encoding with Large Language Models: Benefits and Challenges.” https://doi.org/10.5281/zenodo.13969082. 

Steeg, Fabian, Adrian Pohl, and Pascal Christoph (2019). lobid-gnd – Eine Schnittstelle zur Gemeinsamen Normdatei für Mensch und Maschine. Informationspraxis 5(1), 52673. https://doi.org/10.11588/ip.2019.1.52673.

 

Strutz, Sabrina (forthcoming): “A Multi-Dimensional Evaluation Framework for Assessing LLM Performance in TEI Encoding.” Journal of Open Humanities Data.

 

Strutz, Sabrina, Martina Scholger, and Georg Vogeler (2025): Beyond Rule-Based Processing: LLM-Assisted TEI Encoding of Editorial Interventions in Historical Correspondence. TEI Annual Conference and Members' Meeting 2025, Jagielloninan University in Kraków. Zenodo. https://doi.org/10.5281/zenodo.17176474.

Strutz, Sabrina, Martina Scholger, and Georg Vogeler (2026): LLM-Assisted Metadata Extraction and Normalization for Historical Correspondence: A Multi-Stage Pipeline Approach. DHd 2026 Nicht nur Text, nicht nur Daten (DHd2026), Wien, Österreich. https://doi.org/10.5281/zenodo.18696370.