Predicate Importance Estimation and Decoupled Rationale-Score Distillation for Entity Alignment

The article introduces two novel modules for enhancing entity alignment (EA) in knowledge graph (KG) integration: Predicate Importance Estimation (PIE) and Decoupled Rationale-Score Distillation (DRSD). PIE utilizes a compact embedding-based method to create predicate-aware entity embeddings by aggregating subjectless triples with learnable predicate-importance weights, while DRSD trains a distilled small language model using pseudo-answers from a teacher LLM, converting binary EA labels into text-based supervision. These advancements improve EA classification accuracy and facilitate a human-in-the-loop verification process by distinguishing between confidence scores and decision rationales, which is critical for practitioners working with heterogeneous KGs in LLM applications.