Best Practices

Academic libraries face growing resource volumes and a shortage of experienced catalogers. To address this, the National University of Singapore Libraries initiated the AI-Powered Cataloging (AICAT) Project, exploring how large language models (LLMs) can augment bibliographic metadata creation while maintaining human oversight. This case study shares a practical implementation that generates draft MARC21 records compliant with RDA, ISBD, LCSH, and LCC standards. Utilizing engineered prompts with strict operational constraints—such as hallucination prevention and metadata validation—the project evaluates multiple LLMs to assess accuracy and standards adherence. The three-pronged methodology encompasses prompt engineering, comparative model benchmarking, and workflow automation testing. Results show that AI significantly reduces repetitive cataloging effort and improves scalability, while catalogers retain final editorial authority. Ultimately, effective AI deployment depends on aligning machine outputs with human values through standards-based validation and expert oversight.

The National Library of Korea (NLK) has been implementing the OAK (Open Access Korea) project since 2014 to promote the spread of open access. Through this initiative, the NLK has distributed repositories to universities, research institutes, and public institutions to support the expansion of Green OA (self-archiving). In parallel, the Korean Journal Copyright Information (KJCI) system has been operated to provide information on copyright policies and Creative Commons (CC) licenses of academic societies. As part of these efforts, JATS XML full texts and metadata have been built for domestic open-access journals and shared with academic societies and the National Research Foundation of Korea (NRF). More recently, the NLK has been engaging directly with academic societies to obtain consent for the use of JATS XML scholarly data as AI training data. The consented data will be labeled as AI training data on the NLK's OAK National Repository, and will also be made publicly available through the NLK's Open Data Library. In the long term, the NLK seeks to broaden the reach of Korean research and enhance the societal impact of OA journals by offering a copyright-cleared, openly accessible Korean scholarly dataset.

Since 2013, the National Library of Korea(NLK) has been operating Korean National Bibliography(KNB) LOD service, opening bibliographic and authority data based on KORMARC and MODS as RDF. This presentation reports on the methodology and interim outcomes of an ongoing study on the design and empirical implementation of a BIBFRAME conversion module, building on this operational experience. In particular, regarding the BIBFRAME transition, this study compares the LC approach with the NLK’s own proposed approach to examine which approach is more suitable for the KNB environment. Furthermore, it investigates how the structural characteristics of KORMARC influence the distinction between ‘Work’ and ‘Instance’ in BIBFRAME during this process. Rather than presenting a finalized transition result, this presentation aims to share key considerations and practical directions in the process of preparing for a BIBFRAME-based environment on the basis of experience gained from operating the KNB LOD.

The National Museum of Korea has led the standardization, integration, and opening of museum collection data in Korea through the establishment and dissemination of the K-Museum data standard, the release of public data, and the operation of eMuseum, an integrated public access platform. This presentation introduces the development and major achievements of museum data standardization in Korea through the case of the National Museum of Korea's eMuseum, and further explores possible directions for the future application of global data standards. In particular, it examines how the practical experience of reorganizing collection information that had previously been managed differently across institutions into a more consistent structure and linking it to integrated search and public access services can serve as a foundation for establishing and disseminating the K-Museum data standard. eMuseum is a representative platform where the achievements of data openness are brought together, allowing users to explore and use collection information from most museums in Korea within a single digital environment. This presentation examines the main components and implementation of the K-Museum data standard, and discusses how standardized data has supported the operation and enhancement of eMuseum as well as wider public access to museum collection information. Building on this domestic standardization experience, it also considers the practical challenges and institutional implications that need to be addressed when applying global data standards to the Korean museum environment.

The transition from record-based cataloguing to entity-based knowledge graphs is reshaping not only bibliographic models and technologies, but also the social and organizational foundations of metadata creation. In shared cataloguing environments, entities are no longer owned and maintained by a single institution; instead, they become community-managed resources that evolve through the contributions of multiple stakeholders. In this context it’s interesting to present the experience of the Share Family community in defining best practices for collaborative entity management within JCricket, an entity editor operating on a shared, multi-provenance knowledge graph. Drawing on several years of experimentation with linked data cataloguing, cooperative workflows, and cross-platform interoperability, the community has identified a set of governance challenges that cannot be addressed by traditional cataloguing rules alone. The contribution discusses how best practices are being developed around key areas such as entity creation, provenance management, selection of preferred labels, authority control, relationship management, entity merge and split operations, visibility policies, and integration with external authoritative sources. These practices are designed to ensure consistency, transparency, and trust in an environment where multiple institutions simultaneously enrich and curate the same network of entities. Rather than focusing on cataloguing standards themselves, the presentation argues that the next phase of metadata practice requires community-governed operational frameworks capable of supporting collaborative stewardship of shared knowledge graphs. The Share Family experience offers a practical case study of how libraries can move beyond the traditional dichotomy of local versus copy cataloguing toward new models of cooperative and shared cataloguing, where governance, interoperability, and collective responsibility become central components of metadata quality. The presentation aims to contribute to the broader DCMI discussion on metadata best practices by exploring how communities can establish sustainable governance mechanisms for entity-based description in the linked data era.

Metadata is increasingly recognized not merely as descriptive information, but as critical infrastructure enabling open knowledge ecosystems. As a foundation for interoperability, discovery, and reuse across institutional and national boundaries, open metadata supports more connected and efficient access to information across systems. Yet traditional record-based systems remain constrained by fragmentation, local variation, and limited interoperability, resulting in persistent data silos that hinder effective resource sharing. This presentation examines an ongoing, multi-phase pilot led by the Partnership for Academic Library Collaboration & Innovation (PALCI), a consortium of over 70 academic and research libraries across the Mid-Atlantic United States. The E-ILL Pilot is a proof-of-concept effort to understand how Linked Open Data (LOD) may support resource sharing within a consortial environment. This pilot, a collaboration with Share-VDE and ReShare, examines integrating LOD entities into the ReShare platform. It evaluates how entity models can improve discovery and secure inter-institutional e-book delivery. Moving beyond string-based bibliographic matching, which often fragments representations of the same work, the pilot employs entity-based clustering using the Share-VDE BIBFRAME ontology, a resource description model that reconciles IFLA LRM and RDA. This approach reconciles metadata variations and aggregates holdings at the work level, a capability especially important in a digital context where resources can have multiple formats and versions. To support emerging “e-first” interlibrary loan models, the project examines whether open, interoperable, entity-based metadata can identify related instances and improve matching across manifestations of the same work to support fulfillment. Framed within PALCI’s Digital Sharing Strategy, this work operationalizes metadata as infrastructure, emphasizing openness, reuse, and interoperability as essential to scalable, vendor-neutral systems. As an exploratory initiative, it advances understanding of how open metadata and LOD frameworks can inform interoperable, efficient, and scalable resource-sharing ecosystems across consortial and institutional boundaries.

This paper presents a case study of using AI-assisted development to redevelop iSearch, a legacy cataloging application used daily at the University of Toronto Libraries. Redevelopment has been deferred for six years due to competing institutional priorities. Emboldened by success using AI to develop Model Context Protocol (MCP) servers, the author produced a web-based prototype replacement in approximately six hours. This paper identifies the prerequisites that enabled rapid AI-assisted development and examines emerging implications for library practitioners considering using AI similarly. The paper argues that while AI-assisted development lowers the barrier to writing code, it does not reduce the need for preparation; rather, it changes what skills and knowledge practitioners must bring to the work, and what institutions must invest in to support them.

Now the requirements for academic discovery services (DSs) are dynamically changing in the era of Open Science and Artificial Intelligence. The emerging requirement is that DS is the source of evidence for faithful research activities and for Artificial Intelligence inference. The other important requirement is that DS is becoming a member of the big ecosystem of information processing by Artificial Intelligence. We are developing our service called CiNii, the discovery service for academic resources in Japan, towards this direction. In this talk, we report the current status of CiNii development, in particular, Knowlege Graph Service.

Commentaries are texts produced through the interpretation and annotation of earlier classical works, representing one of the most important mechanisms of knowledge production and cultural transmission in traditional Chinese scholarship. However, the interpretive knowledge embedded in commentary document is characterized by complex, multi-layered, and linear structures, making comprehensive understanding and cross-textual synthesis challenging for both scholars and general readers. This paper presents a framework for GenAI-assisted deep interpretation based on Commentary Knowledge Graphs (CKGs). The proposed CKGs integrates interpretive relations, citation, provenance information, and entity alignments extracted from documents into a unified semantic representation, providing structured interpretive evidence for large language models (LLMs). Building upon this knowledge infrastructure, an LLM-based agent architecture supports deep interpretation through a series of coordinated tasks, including user intent recognition, key parameter extraction, resource retrieval and ranking, context-aware reasoning, and role-based multi-agent collaboration. By combining knowledge graph retrieval with generative AI, the framework enables multi-level interpretation, evidence-grounded reasoning, and the discovery of scholarly connections across texts. A prototype system was developed using Wenfu and its commentary documents as a case study. The system supports functionalities such as conversational interpretation and multi-agent scholarly debate, allowing users to explore classical texts through interactive and explainable AI-assisted workflows. The proposed framework illustrates how cultural heritage knowledge can be integrated with generative AI technologies to enhance the accessibility, interpretability, and reuse. It offers an approach to the digital preservation, exploration, and revitalization of intellectual heritage, while opening new possibilities for explainable AI-assisted research and next-generation digital humanities applications.

Digital game archives require more than descriptive cataloging; they call for a metadata ecosystem that supports both conceptual rigor and practical reusability, underpinning computational use. This paper reports on metadata practices developed at the Ritsumeikan Center for Game Studies (RCGS), organized into three interrelated layers: modeling, linking, and augmenting. These practices are implemented in the RCGS’s online catalog service, the RCGS Collection. At the modeling layer, we develop an extended model based on the IFLA Library Reference Model (LRM) to represent the complex relationships inherent in digital games, including works, variations, packages, and individual items. This approach enables consistent description across heterogeneous archival materials while maintaining compatibility with established bibliographic standards. At the same time, for entities that require subjective interpretation—such as works, genres, franchises, and characters—we design the model to incorporate external authority data. In addition, for the implementation of an online catalog, we employ a “dumb-down” strategy (flattening structured metadata) to improve accessibility for general users. At the linking layer, we implement a Linked Open Data (LOD) approach that connects local entities to external authority sources such as Wikidata and the Media Arts Database (MADB). This alignment supports the persistence of identifiers, enhances interoperability, and enables data enrichment through the incorporation of collectively curated external knowledge. At the augmenting layer, we develop a dataset construction pipeline based on text extraction and image embeddings, transforming digitized materials—including packages, manuals, and gameplay images—into research-ready corpora. Through multimodal feature extraction and automated text structuring, archival data can be repurposed as a resource for computational analysis and metadata generation. These practices are implemented through a SPARQL-enabled LOD infrastructure and a discovery interface based on Omeka S, achieving a balance between a robust identifier infrastructure, machine-readability, and user accessibility. Rather than proposing a prescriptive model, this paper reflects on design decisions and implementation strategies to share insights for building sustainable metadata ecosystems in game archives and related domains.

Against the backdrop of generative AI's rapid penetration into the library sector, how to make AI reliably and efficiently applicable to professional metadata production has become a shared concern for the global library community. Based on the Fourth-generation Intelligent System for Acquisition, Classification, and Cataloging introduced in 2024, the Capital Library of China has innovatively proposed a "Multi-Source Data Integration" technical roadmap, repositioning AI as an "intelligent hub" for coordinating, validating, and decision-making based on data from multiple sources, rather than a sole data producer. This study systematically integrates data from three distinct levels: (1) the Fundamental Data Layer, relying on CIP data from the multidimensional bibliographic database provided by China National Archives of Publications and Culture; (2) the Physical Data Layer, comprising physical metadata precisely collected by sensors in the intelligent processing system (e.g., millimeter-level dimensions, weight, ISBN images); and (3) the Intelligent Parsing Layer, extracting deep semantic information from book scans using multimodal large models, computer vision, and OCR technologies. Supported by four functional engines (multimodal recognition, semantic understanding, multi-source data integration, and automatic MARC field generation), the system automatically resolves data conflicts based on the rule of "priority to benchmarks, fact-checking, and intelligent supplementation," generating standard CNMARC records. From its trial operation in September 2025 to March 2026, the AI Cataloging System has stably processed over 220 books, with an average processing time of approximately 6 minutes per book and automation accuracy exceeding 98% for core descriptive fields. This practice has reshaped the human-machine collaboration paradigm in cataloging: the AI engine undertakes rule-based and repetitive information extraction and integration, while catalogers transition into roles as system trainers, arbiters for complex cases, and auditors of final quality, realizing the complementary strengths of human and AI. This study offers a reusable framework for global peers: by constructing a hybrid intelligent system centered on "multi-source data integration" and safeguarded by "human-machine collaboration," libraries can responsibly and efficiently propel core operations from labor-intensive to knowledge-intensive development.

The long-term sustainability of Linked Open Data (LOD) services remains a significant challenge. Many projects rely on complex server-side infrastructures, SPARQL endpoints, or custom applications that become difficult to maintain once project funding or dedicated technical support ends. This presentation introduces a practical approach to sustainable LOD publishing based on ttl2html, a static site generator that transforms RDF/Turtle datasets into human-readable and machine-accessible web resources. Using this approach, we have developed and continuously maintained an educational LOD infrastructure in Japan since its public launch in 2017. The platform interlinks educational resources such as textbooks, curriculum guidelines, instructional units, and large-scale educational assessments, providing persistent identifiers and Linked Data access through a lightweight and maintainable architecture. We discuss the design principles behind this approach, including static content generation, reduced operational complexity, and compatibility with Linked Data best practices. We report on long-term maintenance experience and examine how minimizing technical dependencies has contributed to the sustainability of the infrastructure. We also demonstrate how the same framework has been reused in other contexts, including student-led initiatives and library-related LOD projects, illustrating its adaptability and low barrier to adoption. By sharing lessons learned from a decade of operation, this presentation highlights how reducing technical complexity through static site generation can improve the sustainability of LOD publishing. Our experience further demonstrates that lightweight publishing infrastructures can be readily reused across projects and organizations, providing a practical foundation for the long-term development and preservation of open knowledge infrastructures.

Metadata for discovery, security and management of collections is a key priority at the University of Oxford, with our institutional strategy stating that to: ‘widen access to our outstanding cultural and scientific collections…. We will aim to have 100% of our collection recorded online… by 2030.’ Delivering this target requires creative approaches to funding, partnerships and technology. This presentation will explore a range of initiatives currently underway, including the automatic creation and enrichment of metadata within our institutional repository; the repurposing and delivery of metadata from legacy finding aids; and the development of pilot workflows for the curation, creation and automated generation of textual transcripts for digitised Special Collections materials. These projects feed into our work at a national level, where we are leading funded research to explore how metadata from different content types can be brought together through minimum viable metadata to enable researchers to engage with collections at scale. We are also one of five testing partners for the UK implementation of BlueCore, the LD4P/Mellon-funded platform for linked open bibliographic metadata. This collaboration represents an important first step in exploring the potential of BIBFRAME for large-scale implementation across the UK. We are also working in partnership with commercial suppliers and other higher education colleagues to test the use of large language models and model context protocols as a route to discovery of our collections. The metadata available to these tools and products is crucial and will be a key element in our evaluation and examination of the effectiveness of the tools and when considering possible future developments. As a next step we plan to develop a metadata strategy, which will identify synergies and areas for collaboration and bring a user focused approach, in order to connect our collections to each other as well as to other collections globally.