Building an AI-tool to facilitate the integration, accessibility, and usability of heterogeneous cultural heritage data on medieval manuscripts.

Información del proyecto

ManuscriptAI

Identificador del acuerdo de subvención: 101158067

DOI

10.3030/101158067

Proyecto cerrado

Fecha de la firma de la CE 5 Febrero 2024

Fecha de inicio 1 Abril 2024

Fecha de finalización 31 Marzo 2026

Financiado con arreglo a

• European Research Council (ERC)

Coste total

Sin datos

Aportación de la UE

€ 150 000,00

Coordinado por

STICHTING RADBOUD UNIVERSITEIT
Netherlands

Periodic Reporting for period 1 - ManuscriptAI (Building an AI-tool to facilitate the integration, accessibility, and usability of heterogeneous cultural heritage data on medieval manuscripts.)

Período documentado: 2024-04-01 hasta 2026-03-31

Resumen del contexto y de los objetivos generales del proyecto

Data on our cultural heritage hold enormous potential for Europe’s economic growth, for the construction of a more inclusive narrative of its past, and for the future collective identity of its citizens. The biggest obstacle to unlocking that potential is the current lack of integration and interoperability among the countless datasets describing the holdings of European heritage institutions. ManuscriptAI helps remove that obstacle for the data on Europe’s medieval written heritage, manuscripts. Premodern handwritten books are a pivotal category of our heritage, yet they are currently underrepresented in large research infrastructures and their catalog data locked in digital silos. ManuscriptAI employs machine learning algorithms to construct a model capable of facilitating the autonomous integration of distinct data sources describing medieval manuscripts, under a predefined set of machine-understandable vocabulary terms. The model is accessible through a human engagement interface and was tested during a pilot in a real-world setting.

The project explores two important desiderata: (1) a user-friendly AI-tool to allow heritage professionals to convert their metadata on manuscripts to Linked Open Data, and (2) a dedicated ontology for the description of medieval manuscripts to complete CIDOC-CRM extensions for the cultural heritage domain. The project, building on the achievements of the ERC-2018-stg PASSIM, is supported by a consortium of research institutes, heritage professionals, and (inter)national research infrastructures. ManuscriptAI advances the EU’s agenda for digital heritage. The tool will help democratise datafication, making Linked Open Data accessible to small heritage institutions and actively involving them in its development. The project's results and insights mark an important step forward for the digital preservation and usability of Europe’s unique handwritten heritage.

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

The ManuscriptAI project completed and validated three core research outputs: an interoperable ontology for the structured description of medieval manuscripts, an AI-supported tool for converting existing manuscript descriptions into Linked Open Data, and a pilot study evaluating both developments with representative end users.
The project's principal achievement is the Manuscript-AI Ontology, which provides a comprehensive framework for describing medieval manuscripts in a structured and interoperable way. Medieval manuscript catalogues have traditionally been created using different standards and practices, making it difficult to compare or combine information across collections. The ontology addresses this challenge by providing a shared conceptual model that enables consistent description and data exchange.
To maximise interoperability, the ontology builds upon internationally recognised standards for cultural heritage and bibliographic information, notably CIDOC CRM and LRMoo. Existing manuscript initiatives, including those developed by Biblissima and the Hill Museum & Manuscript Library, together with established codicological vocabularies such as Muzerelle's Vocabulaire codicologique, were analysed to identify best practices and remaining gaps. The resulting ontology models manuscripts, the works they contain, the people involved in their production and transmission, and their material and historical characteristics. It has been formally implemented using internationally recognised semantic web standards to ensure long-term sustainability and reuse.
Building on this conceptual framework, the project developed a prototype AI-supported mapping tool and dedicated user interface to assist institutions in transforming heterogeneous manuscript descriptions into Linked Open Data. The system analyses existing catalogue records, structures the available information, maps it to the Manuscript-AI Ontology, and can identify links with external knowledge resources. The interface provides four workflows with different levels of automation, enabling users to select the level of user control most appropriate for their data and expertise. The software has been deployed on Radboud University's infrastructure, while the source code is maintained through the project's GitHub repository.
The project concluded with a pilot study to evaluate the ontology, the AI-supported mapping workflow and the user interface. Participants tested the system using unseen manuscript data, assessing its effectiveness in structuring information, converting catalogue records into Linked Open Data, and supporting cataloguing workflows. Particular attention was paid to determining which tasks could be automated effectively and where expert human input remained essential.
The pilot was conducted in July-August 2025. Following the pilot, qualitative and quantitative feedback was analysed to refine the ontology, the mapping workflow and the interface. The results were presented during an online dissemination event for project partners, pilot participants, stakeholders and other interested parties. The presentation was recorded and will be made publicly available together with the project's final report.
Overall, the project delivered the planned scientific and technical outputs and demonstrated their applicability through practical evaluation. Together, the ontology, AI-supported mapping tool and pilot provide a foundation for improving the interoperability, accessibility and reuse of medieval manuscript data across European cultural heritage collections.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

The ManuscriptAI project has demonstrated that semantic technologies and AI can support the creation of interoperable manuscript data while substantially reducing the manual effort required to convert heterogeneous catalogue records into Linked Open Data. The project's three main results—the Manuscript-AI Ontology, the AI-supported mapping tool, and their validation through a pilot study—provide a foundation for improving the accessibility, interoperability and long-term reuse of medieval manuscript data.
The Manuscript-AI Ontology establishes a common semantic framework aligned with internationally recognised standards, enabling manuscript descriptions created by different institutions to be integrated more easily with each other and with wider cultural heritage infrastructures. The AI-supported mapping workflow demonstrates that recent advances in large language models can assist users in structuring complex manuscript descriptions and mapping them to this shared ontology through flexible workflows suitable for different user groups.
The pilot study confirmed the technical feasibility of this approach while providing valuable evidence on the conditions necessary for successful adoption. Participants found that AI-assisted mapping can significantly accelerate data conversion but emphasised that automated processing must remain transparent and subject to expert review. For complex historical material, fully autonomous processing is not yet sufficiently reliable for routine cataloguing. Instead, the project demonstrated the importance of human-in-the-loop workflows, in which AI supports expert decision-making without replacing it.
The evaluation also showed that usability is as important as technical performance. Participants identified intuitive editing functions, clear navigation and compatibility with existing cataloguing practices as essential prerequisites. These findings indicate that successful implementation depends not only on model performance but equally on effective workflow design.
A further lesson concerns the development of the ontology itself. While a standardised semantic model is essential for interoperability, medieval manuscripts often contain uncertainty and scholarly complexity that cannot always be represented efficiently. The project therefore identified a layered approach as a promising direction for future development, combining a stable interoperable core with optional extensions that accommodate more specialised scholarly description.
The pilot further demonstrated that successful implementation requires appropriate guidance and training. Several challenges reflected uncertainty about the interpretation of ontology concepts and the operation of the interface rather than limitations of the underlying technology. Clear documentation, practical examples and user training will therefore be important for wider deployment.
These findings provide a clear pathway for further uptake. Continued collaboration with GLAM-institutions will support broader validation of the ontology and encourage community-driven refinement. Future research should focus on improving AI-supported mapping as language models evolve, including greater transparency, multilingual capabilities and the evaluation of open-source alternatives.