A new method improves entity alignment in knowledge graphs by introducing Predicate Importance Estimation and Decoupled Rationale-Score Distillation. These modules enhance classification accuracy and enable human-in-the-loop verification by detecting uncertain predictions through a decoupled confidence-score estimation.