Answers are grounded in this paper and cite their sources.
Shows effective diagnosis of defect records in bridge inspections, indicating improved model transparency.
Natural language processing application in bridge inspection is growing, yet "black-box" opacity hinders engineering trust. This paper proposes a global interpretability framework using attribute attribution trees to diagnose error-prone samples by analyzing information interaction patterns. By converting attention weights into attribution matrices, a specificity score is introduced to quantify data quality. Results show that model information acquisition aligns with bridge domain knowledge. Crucially, a negative correlation exists between sample specificity and prediction performance, with a maximum accuracy differential of 36%. By elucidating decision preferences, this work provides actionable insights for optimizing defect descriptions at the source, contributing to the improvement of reliability and maintenance efficiency of intelligent bridge management systems.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: