Papers: AI Knowledge and Practice

This paper calls for a rethinking of metadata literacy in the age of AI. Previous discussions often defined metadata literacy as knowledge of descriptive structures or skills in creating and using metadata records. But this understanding is no longer enough – though still necessary – when AI systems create, transform, rank, summarise and recommend information at scale. In such environments, metadata is not simply about post hoc description of resources; it structures provenance, visibility, accountability, cultural representation, and the conditions under which machine outputs can be interpreted and trusted. Using metadata, metadata instruction, AI literacy, and information literacy scholarship, this conceptual paper presents a reconstructed model of metadata literacy for LIS education. The model is built on five dimensions: understanding of infrastructure, contextual description and representation, provenance and disclosure, evaluation of algorithmically mediated outputs and intervention in terms of ethics and governance. The paper also gives examples of learning tasks and assessment evidence to illustrate how the model can be applied. It argues for viewing metadata literacy not as a narrow technical specialisation but as a fundamental educational response to AI-mediated knowledge environments.

This paper presents a literature-based synthesis of motivations for participating in biomedical ontology communities, viewed as metadata infrastructures. As generative AI transforms ontology curation into human-in-the-loop workflows, human engagement becomes essential for ensuring metadata quality. Using Self-Determination Theory and Activity Theory, it identifies four themes—intrinsic motivation, extrinsic motivation, community aspects, and human AI collaboration—and analyzes their impact on autonomy, competence, and relatedness. Based on these themes, the study proposes practical implications for provenance-enhanced verification, quality-based incentives, and collaborative environments to sustain metadata quality and ongoing contributions.

There is a wide variety of date metadata practices employed by professional, paraprofessional, and volunteer catalogers and metadata creators. The lack of standardization in date metadata inhibits discovery. While normalization routines can be created to improve metadata, the sheer scale of date values and types of date formats can make that time consuming and onerous. As artificial intelligence (AI) is employed in metadata work, this paper addresses its potential usefulness in standardizing date metadata in more frequently used scenarios. In this exploratory study, the researchers employed several freely available AI chatbots: ChatGPT, Claude Sonnet, and Gemini.

The rapid expansion of digital repositories and scholarly resources has increased the demand for scalable and intelligent metadata management systems. Traditional metadata creation methods, which rely on manual cataloguing by information professionals, struggle to keep pace with the growing volume and heterogeneity of digital content. This study develops and evaluates an AI-guided framework, Metadata-Building-AI-Guidance V.1.0 that combines document ingestion, chunking, OpenAI text-embedding-3-small embeddings, and the gpt-4o-mini large language model into a single Streamlit application supporting both automated Dublin Core-aligned metadata extraction and embedding-based semantic retrieval. We position the system as a design-science artefact in which retrieval is not a side feature but a feedback loop: the same vector index that powers similarity search is also used to surface the contextual chunks from which structured metadata are extracted and to support retrieval-augmented question answering over uploaded collections. The system was evaluated on a purposively sampled corpus of 30 open-access academic documents and 20 expert queries, using (i) field-level F1 against librarian-curated ground truth with Cohen's κ for inter-annotator agreement and (ii) Precision@5 and Mean Reciprocal Rank with binary relevance judgements from two librarians. Field-level F1 ranged from 0.67 to 0.73 for dc:title, dc:creator, dc:subject, and dc:description, with overall κ = 0.83 (almost perfect agreement); the lowest F1 was 0.57 for dc:type, traced to a fixed generic prompt output rather than to a model-capability limitation. Semantic retrieval reached Precision@5 = 0.61 and MRR = 0.69 with κ = 0.66 (substantial agreement) across the 20 queries. We discuss the limits of LLM-only evaluation—including the absence of a head-to-head comparison with established non-LLM extractors such as GROBID—and identify controlled baseline comparison together with a refined dc:type prompt as the immediate next steps. Prompts, JSON schema, library versions, sampling log, and evaluation queries are released to support replication. The contribution is a reproducible reference implementation that aligns AI-assisted metadata extraction with Dublin Core Terms and the FAIR principles for digital libraries, archives, and cultural-heritage repositories.

Automated subject indexing remains a critical challenge for digital libraries and knowledge organization systems. To address this issue, this study develops a knowledge-augmented domain adaptation framework that aligns general-purpose language models with the hierarchical logic of the Chinese Library Classification (CLC). A supervised fine-tuning (SFT) strategy is proposed to resolve domain knowledge drift and deep-category recognition bottlenecks in large language model (LLM)-based Chinese book indexing, using dual bibliographic and category data. Three evaluation experiments were conducted to assess the effectiveness of the proposed techniques for quantifying ontological contributions: hyperparameter sensitivity analysis for baseline establishment, backbone model comparison for architectural fitness, and knowledge injection ablation for. Results demonstrate that dual-data fine-tuning significantly enhances precision for long-tail and fine-grained categories. While ensuring high-quality output, the solution features low computational thresholds, robust local deployment, and high scalability, effectively internalizing knowledge organization systems within LLMs. By bridging the gap between classical theory and generative AI, this work provides a high-accuracy, institutionally autonomous solution for automated indexing, offering substantial theoretical and practical significance for the intelligent transformation of digital libraries.

Large language models (LLMs) applied to subject cataloguing using Library of Congress Subject Headings (LCSH) tend to generate headings and strings that are syntactically plausible but policy-invalid, bypassing the controlled vocabularies and governing rules that subject collocation depends on. This paper describes a two-track experiment to address the lack of grounding in standards and policy. The first track is lc-vocabularies-mcp, a Model Context Protocol (MCP) server connecting an LLM to live Library of Congress (LC) linked data APIs, enabling real-time authority validation for LCSH and related controlled vocabularies. The second track is a conversion of the Subject Headings Manual (SHM) from PDF to structured DITA (Darwin Information Typing Architecture), designed as a machine-actionable retrieval-augmented generation (RAG) corpus. Together, the two tracks support a five-part subject cataloguing workflow in which non-parametric knowledge is supplied to Claude at each part where parametric knowledge alone is insufficient. The paper reports on the architecture of lc-vocabularies-mcp, the DITA conversion methodology, and an evaluation design that tests whether policy-grounded retrieval improves AI-assisted subject cataloguing.

AIs are increasingly used to support metadata processing and investigation, which depends on whether AI-encoded meaning aligns with human meaning. However, AI encodes word meaning through distributional and contextual representations, and it remains unclear whether such representations preserve the meaning value of human system. We answer this question through Zipf’s meaning law, which links word frequency to number of word meanings. We compare multiple AI-induced meaning estimates with human-measured meaning. To quantify alignment, we propose Meaning-Zipf Deviation (MZD), which covers continuous meaning distributions and measures their divergence with reliability adjustment. Extensive experiments show that human words consistently follow Zipf’s meaning law. AI-encoded meanings also exhibit Zipfian regularities, inheriting part of the statistical structure of human language. However, AI meaning distributions remain flatter than human distributions, with lower scaling exponents and non-negligible MZD values. Larger models do not reduce this gap. AI tends to bind words to context-conditioned senses rather than preserve their broader polysemous potential.

Applying the theory of representation, in this paper it is argued that the limitations of AI generated solutions are due to AI’s lack of multifaceted cultural, historical and situational contexts and beliefs that constitute the fabric of the authoritative knowledge. This paper suggests that the problem of AI is not so much of the scale of the data on which it is trained as it is the lack of depth and quality of data that represents the contextual and situated knowledge produced by human experience and cognition.