Capitularia Tool Stack and Workflow<\/h5>\n
In the capitularies project, we currently rely deliberately on manual transcription and post-collations in the oXygen XML Editor<\/a>, rule-based collation (based on CollateX<\/a>), and traditional editorial standards. The reasons lie partly in the “pre-dating” of the project, whose conception and application goes back to a period when the tools available today \u2014 some of which have since become well-established \u2014 did not yet exist or were still in their infancy. On the other hand, and these are arguably the weightier considerations, the reasons are rooted in the materials being edited. The transmission situation of the capitularies explains why tools deployed profitably in other projects reach their limits here:<\/p>\n Handwritten Text Recognition<\/em> (HTR) with tools such as Transkribus<\/a>, however, works best with large quantities of consistent training data and longer, coherent texts by a small number of scribes. Although most capitularies are written in Carolingian minuscule \u2014 a script for which a Transkribus model exists \u2014 the transmission also includes decidedly earlier as well as considerably later textual witnesses. Overall, it can be assumed that applying existing models to the short, diversely transmitted texts would yield comparatively high error rates. Post-processing would therefore likely take longer than transcription itself. It should also be borne in mind that high-resolution digitizations are not available for all manuscripts, and digitization quality equally determines the proportion of correctly recognized letters and abbreviations. For these reasons, the current workflow \u2014 providing for manual transcription followed by two complete post-collations (six-eyes principle) by trained staff \u2014 appears to be without viable alternative. The philological precision and quality thus achieved could not be delivered by HTR without substantial checking and revision.<\/p>\n Workflow for the Digital Edition (own illustration; created with the help of ChatGPT)<\/p><\/div>\n A further method now quite established in editorial contexts is Named Entity Recognition<\/em> (NER), as it facilitates index creation and consequently the linking of authority data. As noted, capitularies \u2014 with the exception of the relevant ruler(s) \u2014 rarely name specific persons or localities. Compounding this are the difficulties associated with historical place names and their geographic identification. Which person, place, or region is intended can only be determined through deep historical expertise. The identification and disambiguation of named entities<\/em> therefore remains an intellectual, editorial activity that cannot be automated in our context.<\/p>\n Even for applications promising added value or time savings, their adoption warrants critical scrutiny. After twelve years we have proven and established workflows. Retrospective integration of new tools would entail considerable effort \u2014 with rather uncertain benefits. There are also broader scholarly and ecological concerns: AI tools do not guarantee transparency or reproducibility, and it frequently remains unclear how a given output came about. The underlying data of commercial models is moreover potentially questionable regarding provenance and quality. Relying on such tools creates dependencies that are naturally undesirable from a project perspective. Universities often officially prohibit the use of commercial models and providers outright. The open alternatives that must consequently be considered are comparatively limited and lag behind the major players in output quality. Access to the requisite computing capacity \u2014 for instance for training proprietary models \u2014 is also frequently restricted.<\/p>\n The following section explores where we see opportunities in employing AI. It is important to note that, on principle, only applications operating according to the human-in-the-loop<\/em> principle are considered \u2014 ones in which people remain embedded in the work process, verify outputs for errors, and retain decision-making authority. Applications can be assigned to the following levels:<\/p>\n The basis for creating the critical edition in the Classical Text Editor (CTE)<\/a> is the manuscript files encoded in TEI\/XML. The collation tool<\/a> based on CollateX<\/em> supports text constitution through the dynamic display of deviations and variants between manuscripts. CollateX<\/em> operates in a rule-based manner. The editors of the individual texts produce the critical text, variant apparatus, factual annotations, and translation in the CTE, thereby preparing as thoroughly as possible the printed edition appearing in the MGH’s leges<\/em> series.<\/p>\n Workflow print (own illustration; created with the help of ChatGPT)<\/p><\/div>\n After an embargo period, the MGH processes the printed edition for use within their online platform dmgh.de<\/a>. It would naturally be desirable for the critical text and translation to also be available on the Capitularia website and accessible via search \u2014 this would link the digital and printed editions far more closely. In principle the CTE offers export formats; however, these have proven occasionally lossy, necessitating laborious post-checking. Working directly with the CTE file itself would eliminate the risk of loss during transformation to export formats. A subsequent approach would be AI-assisted re-encoding, for example via a transformer model, which would considerably accelerate the digital publication of the critical text without compromising editorial standards. Since the files follow a uniform template, recognizing apparatus entries and transferring them into the corresponding TEI structure ( A potential workflow could be structured as follows:<\/p>\n CTE export (critical text) Beyond re-encoding, several other domains merit consideration:<\/p>\n Potential areas of application for the controlled use of AI (own illustration; created with the help of ChatGPT)<\/p><\/div>\n We currently forgo AI applications not on principle, but for pragmatic reasons: the manuscript tradition, our source material, and the existing workflows make common tools such as Transkribus uneconomical for us. At the same time, we recognize genuine opportunities. Above all, the re-encoding of critical texts by means of AI models appears as a promising scenario; extended search capabilities, visualizations, and AI-ready data preparation are further areas of interest.<\/p>\n It does not seem productive to choose between “yes to AI” or “no to AI,” but rather to assess anew for each task whether there is a genuine benefit and whether deployment can be justified scientifically, resource-wise, and ecologically. Where AI is employed, this happens selectively and transparently. The design of infographics and visualizations for presentations has, for instance, become considerably more straightforward through the new possibilities.<\/p>\n \n <\/a>[1] This contribution is fundamentally based on a talk delivered in December 2025 at the workshop “Digital Editions of Contemporary History between AI and Linked Open Data” (Berlin, 04\/05.12.2025) of the Commission for the History of Parliamentarism and Political Parties (KGParl). A PDF version of the presentation is available on Zenodo<\/a>.\u2191<\/a><\/p>\n <\/a>[2] Cf. Daniela Schulz (8 May 2025). Edit recommends #3: Burchard’s Decree Digital (Resource Roundup Special). Text+ Blog<\/em>. Retrieved 25 June 2026 from https:\/\/doi.org\/10.58079\/13w6w\u2191<\/a><\/p>\n <\/p>\n \n
![\"\"](\"https:\/\/capitularia.uni-koeln.de\/wp-content\/uploads\/2026\/06\/WorkflowDigEd-1024x216.png\")
<\/a>Potential Usage Scenarios for Applications or Methods<\/h5>\n
\n
CTE2TEI<\/h6>\n
![\"\"](\"https:\/\/capitularia.uni-koeln.de\/wp-content\/uploads\/2026\/06\/workflowPrintEd-1024x114.png\")
<\/a><app><\/code>, <lem><\/code>, <rdg><\/code>) should be straightforward. The same applies to emendations. This is not “creative” work, but rather straightforward re-encoding following clear rules, so editorial standards would not be endangered. Initial experiments proceeding from the image of the printed edition are quite promising.<\/p>\n
\n\u2193
\nSpecialized transformer model (trained on 20\u201330 already encoded examples)
\n\u2193
\nTEI-XML (automatically re-encoded)
\n\u2193
\nManual review by editors
\n\u2193
\nDigital publication + web edition<\/p>\nFurther Scenarios<\/h6>\n
\n
![\"\"](\"https:\/\/capitularia.uni-koeln.de\/wp-content\/uploads\/2026\/06\/KIPotenziale-1024x145.png\")
<\/a>Summary<\/h5>\n
Further reading on the topic:<\/h5>\n
\n
How to cite<\/h5>\n