SimRAG as Traceable Governance for Multilingual Thesaurus Localization: Evidence From AAT

Multilingual thesauri support cross-lingual access and semantic interoperability, but localization remains difficult to scale because it requires stable concept boundaries, traceable decisions, and accountable maintenance. This exploratory design study treats multilingual thesaurus localization as both a translation problem and a knowledge organization governance problem. It proposes Simulated Retrieval-Augmented Generation (SimRAG) as a traceable design for examining artificial intelligence (AI)-assisted localization before a full retrieval-augmented generation environment is available. Using the Art & Architecture Thesaurus (AAT) as the empirical setting, the study examines 30 concepts across seven facets and focuses on two governance-critical tasks: term alignment and scope note translation. SimRAG constructs a controlled evidence setting using curated snippets from authoritative sources so that candidate terms, evidence links, risk notes, and editorial decisions can be reviewed together. In the internally adjudicated review setting, preferred-term status was stable across all 30 sampled concepts. Non-preferred-term alignment was more variable, with evidence-backed candidates showing higher observed alignment than model-inferred candidates, suggesting that candidate provenance is a useful cue for editorial prioritization. For scope notes, AI-drafted translations received higher adjudicated consensus ratings than existing translations on semantic accuracy, fluency, and cultural adaptation. These findings should be interpreted as design-oriented observations under controlled evidence conditions, without external validation of model quality. Qualitative analysis identified two recurring risks: semantic compression and pragmatic overextension. SimRAG’s main contribution is a traceable editorial governance design that links evidence, candidate provenance, risk judgment, and knowledge organization system (KOS) status assignment. The paper contributes to knowledge organization by showing how AI-assisted multilingual thesaurus work can be organized as a reviewable, warrant-based decision process.