CHExGraph: A Knowledge Graph for Jagiellonian University Cultural Heritage

CHExGraph is envisioned as a unified KG containing interlinked resources with information shared between institutional units. The current version focuses on the early period of the university's history, from its foundation in 1364 to the reforms of 1780 [1], providing a structured data resource for the study of pre-modern and early modern academic and intellectual life. The dataset covers a broad temporal span from the late Middle Ages to the early modern period, encompassing several key historical transformations, including the rise of vernacular writing, the impact of the Reformation, and the evolution of scholarly practices.

The backbone of CHExGraph is provided by data from the AUJ's Corpus Academicum Cracoviense (CAC), a database containing approximately 67,000 records on students and graduates of the University of Kraków¹ for the period 1364–1780 [1]. CAC includes detailed information on academic trajectories as well as occupational and personal events in the lives of individuals. The data are maintained in a PostgreSQL database using a person-centric custom data model. Data from the JL originate from the ALMA system, an integrated library management platform containing more than 3 million bibliographic records. These data are structured according to the MARC21 format and include information on authorship, publication details (place, publisher, and date), language, and subject or genre classifications. Both CAC and ALMA data were provided as database dumps for processing. Finally, data from MUJ are accessed via the MuzUJ system, which exposes records through a private REST API. MuzUJ contains metadata for approximately 75,000 objects described using a custom object-centred data model.

The current scope of CHExGraph is restricted to a subset of records centered on persons represented in CAC that can be matched to at least one additional source (JL or MUJ). Records from JL were included when a matched person appeared as a creator, contributor, or as part of the thematic subject of a bibliographic record. Records from MUJ were incorporated when a matched person was associated with an object in roles such as creator, contributor, depicted individual, owner, or other documented forms of involvement. This selection criterion ensures the presence of explicit cross-institutional links and supports the primary objective of the KG, namely the integration and exploration of relationships between entities across previously disconnected collections.

Prior to the CHExRISH project, no mapping was available for metadata shared across the units. Thus, first a entity reconciliation pipeline had to be developed and applied. Several alternatives were tested, as described in [7], and a pipeline based on a fine-tuned multilingual Transformer (LaBSE) [8] was used for match candidates identification. Domain experts from AUJ verified the matches, so all reported matches have been manually verified. Table 1 summarizes the number of records and matched entities across the three institutional sources after the reconciliation process. A similar supervised strategy was used to matching place names across the institutions.

Another major barrier to CH metadata integration across institutional units is the use of different data models, cataloguing traditions, and levels of semantic granularity. Although the underlying records of the JU collections describe related entities, such as persons, places, objects, and events, they do so in institution-specific ways that make direct interoperability difficult. For this reason, CHExGraph adopts the CIDOC-CRM as its core conceptual framework, since it provides a shared vocabulary for representing CH data across museums, archives, and libraries [5, 13].

CIDOC-CRM is particularly well suited to CHExGraph because it models heritage information in terms of entities and events rather than relying on a single fixed record structure. This event-centric perspective makes it possible to align heterogeneous source data while preserving the relationships between people, objects, documentary evidence, and the historical activities that connect them. In our case, this is essential for integrating the person-centric structure of CAC, the bibliographic records of JL, and the object-centered descriptions of MUJ within a common linked data environment.

However, the base CIDOC-CRM ontology alone is not sufficient to meaningfully capture all distinctions present in the source data. Thus, CHExGraph introduces a set of custom classes and properties to represent recurring domain-specific patterns in the AUJ, JL, and MUJ datasets². These include, among others, NE1_Educational_Activity and NE2_Occupational_Activity as subclasses of E7_Activity, as well as controlled types for academic degrees, fields of study, professional roles, occupational activity types, educational activity types, and creation roles. These extensions are paired with CRM-compatible properties that specialize P2_has_type, including, e.g., NP3_has_associated_ academic_degree, and NP5_has_professional_role. This modeling strategy preserves the semantics of the original records while keeping the integrated KG aligned with CIDOC-CRM and interoperable with external linked data resources. Figure 1 presents an excerpt of CHExGraph demonstrating how entities from different institutional sources are integrated through the shared CIDOC-CRM-based representation and the extended classes and properties.

The conceptual model was guided by a set of competency questions capturing the intended cross-collection uses of CHExGraph. These questions focus on tasks that cannot be answered easily from a single source system, such as connecting person-centred biographical records with object-centred bibliographic and museum metadata. The same questions are later operationalized as SPARQL queries for Sect. 5.2.

• CQ1: Person timeline. Can educational activities, occupational activities, creation events, and production events be combined into a single timeline for a selected person?
  
• CQ2: Place context for given person. Which places are associated with events involving a selected person?
  
• CQ3: Persons as subjects of works. Which persons are the subjects of linguistic objects and which of them are the most popular ones?
  
• CQ4: Persons who are both actors and subjects. Which persons are both represented as subjects of works and as actors in creation or production events?
  
• CQ5: Aggregated event distribution over time. What is the decade-level distribution of educational activities, occupational activities, creation events, and production events?
  

To assess the scope and coverage of CHExGraph, we computed statistics across the main classes in the KG. The graph contains 5,227 distinct E21_Person instances, representing individuals connected with the university's historical, academic, and cultural activities. These persons are further contextualized through 386 E74_Group instances, which represent institutional and social entities such as universities, organizations, associations, and other collective actors.

The spatial dimension of the graph is represented by 1,831 E53_Place instances, including locations associated with people, events, educational activities, professional activities, and cultural objects. The graph contains 12,781 NE1_Educational_Activity instances and 28,770 NE2_Occupational_Activity instances, which extend CIDOC-CRM's event-based modeling approach to represent recurring academic and professional trajectories. These entities enable the representation of relationships between persons, institutions, fields of study, academic degrees, and professional roles that are not explicitly captured by the base ontology.

At the object and documentary level, CHExGraph contains 10,407 E22_Human-Made_Object instances and 9,455 E33_Linguistic_Object instances. The former represents CH objects described in the museum collections, together with their production, material, technique, dimensions, and related events, while the latter captures textual and bibliographic resources.

At the moment, we have conducted a supervised entity linking process aligning 612 persons and 237 places with corresponding entities in Wikidata³. Although the current alignment is limited to these two entity types, it already demonstrates CHExGraph's integration into the linked data ecosystem. By connecting core entities to widely adopted external identifiers, the knowledge graph becomes interoperable with other CH and scholarly datasets published as linked data, facilitating data enrichment, federated querying, and reuse beyond the JU collections. The current version of CHExGraph contains a total of 862,877 RDF triples.

We used the Ontology Pitfall Scanner (OOPS!)⁴ [14] to assess the quality of the ontology extension, focusing on structural aspects of the ontology such as class and property modelling, reuse of existing vocabularies, and possible design weaknesses. The identified warnings were reviewed manually⁵ and addressed. Finally, OOPS! successfully validated our ontology extension.

The competency questions defined in Sect. 3.3 were operationalized as SPARQL queries⁶ and executed against the public endpoint (Sect. 4). Their successful execution confirms that the conceptual model is not only structurally expressive, but also practically queryable: the required entities and relations are present in the graph and can be retrieved through stable modeling patterns. The evaluation demonstrates that CHExGraph supports cross-collection queries that would be difficult to answer using the original institutional systems separately.

In particular, the queries combine person-centred archival data with object-centred bibliographic and museum-related metadata. For example, the person timeline query integrates educational and occupational activities with creation and production events, while the place-context query retrieves spatial traces across birth, death, education, occupation, and object-related events. The results also show that the graph can support research-oriented exploration rather than simple record lookup. The returned data can be used to reconstruct individual trajectories, study spatial contexts of historical actors, identify prominent persons in the corpus of works, and observe long-term distributions of academic, professional, and cultural production events. This confirms that CHExGraph provides an integrated evidential layer for cross-collection analysis of JU heritage.