Human–AI Semantic Identification of Components in Traditional Chinese Floor-Standing Interior Timber Screens: Expert Effort and Record Agreement

Semantic documentation of traditional Chinese interior timber screens requires integrating local form, spatial position, and constructional relations. We compared human-only labeling (H), comprehensive AI assistance (HA), and risk-oriented selective review (RH) for components in five floor-standing screens at three heritage sites. Among 188 candidate masks, 30 measured agreement among six independent experts, and 158 entered a counterbalanced crossover experiment. Machine labels and self-reported review flags were frozen before expert assessment. Complete agreement on the shared set was 90.00% (Fleiss’ κ = 0.9396). Matching labels from two H experts defined an internal operational consensus for 154 formal candidates. Agreement with this reference was 64.94% for AI-only, 97.08% for HA, and 75.32% for RH. HA exceeded AI-only by 32.14 percentage points (95% CI, 24.68–39.61). RH routed 20.89% of formal candidates to experts and reduced recorded expert decision time by 81.54% relative to H in this experiment. The frozen review flag captured 33.33% of machine–consensus disagreements, while experts realigned 88.89% of routed disagreement records. For these screen components, comprehensive assistance achieved the highest consensus agreement, whereas selective review under the tested routing strategy reduced expert effort but left most machine–consensus disagreements unreviewed.