Invited Talk 1: Metadata in the Age of AI

Last November, AI and large language models (LLMs) emerged from the shadows of academia with the promise, or threat, of disruptively changing how we find information, learn, and solve problems. How does this fit with the evolution of metadata technology since the arrival, thirty years ago, of the World Wide Web? Modern metadata began with RDF, a model for expressing data as "graphs" of related entities with global identity (URIs). RDF provided the basis for "ontologies" -- carefully engineered models of logically related classes in support of automatic reasoning. "Linked Data" and "Simple Knowledge Organization System" (SKOS) achieved adoption at scale by promoting easier methods for building "knowledge graphs". Google became the dominant search engine by leveraging full-text indexing and weblink statistics to answer queries with relevant Web "hits". In November 2022, ChatGPT raised the bar for search engines by answering queries not with Web links, but with (ideally) well-formulated explanations. This talk offers a few guesses on how generative AI will change the nature of metadata. Modern metadata in the style of Dublin Core uses a "language of description", with RDF or OWL properties and classes along with SKOS concepts -- all identified with URIs -- to describe things in the world (metadata "descriptions"). It uses ShEx or SHACL "data shapes" ("application profiles") to describe those descriptions. Seen from the standpoint of neural AI networks, the terms of this language now appear as tokens in layered models of statistically weighted vectors. Models that have traditionally required explicit engineering are now "learned" from training data. Creators of metadata models can increasingly become trainers of AI models. Curated models built around controlled vocabularies can potentially improve the focus and reduce the financial and environmental costs of LLMs. On the example of "NALT in the Machine Age", a project of the USDA National Agricultural Library, this talk argues that intelligent use of AI can potentially improve the quality and scalability of information management in an environment where the sheer mass of information continues to expand while budgets for labor-intensive manual curation of that information continue to contract.