This paper proposes an event-centric data modeling workflow for contemporary art events in South Korea. Although GLAM has invested substantial effort in event-centric conceptual modeling, such as CIDOC-CRM (Doerr et al. 2007), contemporary art events are still published as simple lists rather than managed as interoperable data. As Wildenhaus (2019) notes for exhibitions, art events are relational entities that are time- and site-specific; yet their documentation remains dispersed across each institution's webpages, making it difficult to reconstruct an integrated view of the art field across institutional boundaries. Extending this insight beyond exhibitions, this study positions art events—including not only exhibitions but also education and cultural programs—as activity-bearing units through which creation, curation, education, critique, and research can be contextualized. Structuring and integrating such events into a relational database (RDB) and a semantic database with an event-centric schema can support data accessibility and reusability, while enabling role-relevant discovery by making cross-institutional connections among art activities easier to trace—without requiring users to manually assemble context from scattered webpages.

To address this gap, the study constructs an activity-oriented art event corpus of South Korean contemporary art events (2000-present) by integrating data from public contemporary art institutions—National Museum of Modern and Contemporary Art, Korea (MMCA), Seoul Museum of Art (SeMA), and National Asian Culture Center (ACC)—with open call announcements from ARKO Artnuri. South Korea provides a suitable testbed because public funding opportunities are especially prominent in shaping art activities, enabling an examination of event-centric integration across institutions. Concretely, the paper contributes: (1) an activity-oriented cross-institutional art event corpus; (2) a common-core schema with event-type extensions, formalized as a lightweight event ontology and serialized as Linked Art-compatible JSON-LD; and (3) ontology-guided semantic enrichment via OntoChatGPT (ontology-driven prompting) to generate ontology-consistent conceptual annotations for each event.

In the data collection phase, data were harvested through an automated web-harvesting pipeline implemented in Python with Playwright from the institutions' websites (www.mmca.go.kr, www.sema.seoul.go.kr, www.acc.go.kr), complemented by ARKO Artnuri, an open call portal operated under the Arts Council Korea (www.artnuri.or.kr). Because (1) exhibitions, (2) education programs, (3) cultural programs, and (4) funding/support open calls are curated for different purposes, they are published with specific metadata fields that vary both within and across the websites. Accordingly, text-based pages are processed via rule-based DOM extraction for stable fields such as title, dates, venue, and descriptions, alongside each site's internal resource identifiers. For open calls—where essential details are frequently embedded in poster images or attached files—the pipeline applies OCR as a fallback to recover key textual fields. To support verification and re-harvesting, each page is also archived via the Internet Archive's Wayback Machine (web.archive.org), with the source URL, capture timestamp, raw HTML, and resulting snapshot URL stored as provenance. Importantly, we retain each source's native event subtypes (e.g., Film & Videos, Performance Arts, Conservations) in their original form as provenance-rich descriptors. Together, these harvested records are prepared as inputs for the subsequent data structuring and organizing phase.

Given that exhibitions, education programs, cultural programs, and open calls are designed for different purposes, their descriptive fields should not be forced into a single flat schema. The proposed model therefore defines a common-core event schema capturing identifiers and provenance, titles, time span, venue, and overview descriptions, and adds three event-type extensions to preserve what is meaningfully specific. For instance, exhibition extensions capture exhibition fields and genre; program extensions—covering education and cultural subtypes—capture target audience and delivery format; and open-call extensions capture eligibility and application periods. Subtypes native to each source are preserved as categories. Related entities such as agents, organizations, and venues are extracted and normalized from event records into separate tables, forming a relational database (RDB). The integrated schema is also expressed as a lightweight event ontology in which Event–Agent connections are reified through a Participation record, allowing each agent's role and activity context to be captured at the relationship level. Each Participation record stores not only a role label (e.g., artist, curator, educator, researcher) but also an activity type (e.g., creation, curation, education, research), indicating how a contribution becomes meaningful in context. This enables an activity-oriented perspective, such as decomposing a single event into participation records aligned with distinct activity types—creation (artists), curation (curators), and support (organizing institutions or sponsors)—and distinguishing the same person's engagements as different activities across events. The design is aligned with established cultural heritage models and vocabularies (CIDOC-CRM, LIDO, Getty AAT) and serialized as Linked Art–compatible JSON-LD for interoperable publication and cross-institution querying.

Building on the structured event records, we conduct an ontology-aligned concept annotation phase that adds explicit, reusable semantics to each event. Because institutional overview descriptions (and open call texts recovered from poster-image OCR) vary widely in length and specificity, the consistency of concept annotations depends on how extraction is constrained. To align with established cultural heritage vocabularies and support cross-institutional linkability, we adopt an ontology-driven prompting approach that bounds concept extraction to Getty AAT, rather than relying on unconstrained methods such as frequency-based methods (e.g., TF–IDF) or free-form LLM output. Specifically, we adapt the ontology-driven prompting principle of OntoChatGPT (Palagin et al., 2023)—its facet-bounded structured-prompt design—without implementing the full meta-ontology dialogue pipeline. This approach constrains LLM outputs by providing facet-specific instructions and requiring a fixed JSON schema aligned to the target ontology and vocabularies. In our setting, the input is each event’s descriptive text, and the output is a set of facet-specific concept annotations (e.g., practice/format, medium, topic/discourse, and support-related categories). In particular, Getty AAT hierarchies are used to bound concept extraction by prompting the LLM to select an appropriate AAT facet/branch first and then propose narrower terms within that scope, rather than emitting free-form keywords. For example, ontology fields can be aligned with AAT facets—role labels with Agents, activity types with Activities, and topic/discourse with Associated Concepts—so each annotation is drawn from a facet-bounded concept space and recorded as a typed value in the corresponding field. These annotations make cross-institution linking and querying feasible at the level of activity-oriented concepts, and can subsequently support further analysis and application. As a minimal quality-control protocol, generated annotations are screened with schema-level constraint checks (facet admissibility, type compatibility, and controlled-vocabulary alignment), with non-conforming outputs rejected. To validate the pipeline on a held-out sample, we further measure (i) the first-pass AAT validity rate—whether LLM-generated terms exist in the Getty AAT before any rejection, testing for hallucination in vocabulary selection—and (ii) inter-event annotation consistency, the rate at which thematically similar events receive overlapping AAT terms. A manual spot-check against the source descriptions/OCR text provides qualitative verification.

Taken together, the workflow demonstrates how dispersed contemporary art event listings can be transformed into interoperable, activity-oriented records with controlled vocabulary annotations, enabling cross-institution comparison and reuse without manual context assembly. The Linked Art (JSON-LD) corpus will be published as a demo website, illustrating how event-centric modeling and annotations restrained with controlled vocabulary can underpin faceted exploration, and querying of art events.