Adaptive knowledge distillation based structure-text embedding integrating for knowledge graph completion

Structure-text integration for KGC: Knowledge graph completion requires effective integration of structural and description information to improve downstream utility and address complementary weaknesses.

Multi-semantic entity transfer gap: Existing embedding-level coupling approaches do not transfer multi-semantic entity knowledge learned from structure models to PLMs, limiting integration quality.

Structure-text integration for KGC: Knowledge graph completion requires effective integration of structural and description information to improve downstream utility and address complementary weaknesses.

Multi-semantic entity transfer gap: Existing embedding-level coupling approaches do not transfer multi-semantic entity knowledge learned from structure models to PLMs, limiting integration quality.

AKD-KGC: AKD-KGC adds an Adaptive Knowledge Distillation teaching-learning procedure during feature integration to transfer structural multi-semantic knowledge and enhance structure-text integration for KGC.

Structural-model-guided integration: The framework integrates structural and textual features at the embedding level while using structural models to guide the integration model’s prediction behavior and supervise PLM weighting.

AKD-KGC: AKD-KGC adds an Adaptive Knowledge Distillation teaching-learning procedure during feature integration to transfer structural multi-semantic knowledge and enhance structure-text integration for KGC.

Structural-model-guided integration: The framework integrates structural and textual features at the embedding level while using structural models to guide the integration model’s prediction behavior and supervise PLM weighting.

State-of-the-art KGC performance:

Code and datasets released:

State-of-the-art KGC performance: