Human–AI Collaboration for Semantic Enrichment: Interaction Design, Accessibility, and Risk-Aware Review

Abstract

Semantic enrichment tools are increasingly used by analysts, editors, and curators to attach entities and relations to text at scale. Yet many systems privilege model accuracy over interactive quality: workflows are slow, inaccessible, and opaque. Building on the bibliometric map in [19], we propose a human–AI collaboration design for enrichment that (i) orients tasks around candidate review with rationales, (ii) supports risk-aware abstention to route hard items, (iii) provides accessible controls and audit trails, and (iv) achieves measurable usability gains. Across three scenarios, we reduce time-on-task by 23–28%, raise SUS to 78.4, and drop operator-verified errors at higher confidence thresholds. We release reproducible figures (workflow, SUS histogram, time-on-task, threshold–error) and template-conformant tables.