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

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Knowledge Representation

Background:

• Knowledge Graph Completion (KGC) is crucial for semantic search and question answering.
• Effective KGC integrates structural and descriptive information to address limitations like long-tail issues.
• Current methods integrate information at the embedding level but struggle with multi-semantic entities and knowledge transfer.

Purpose of the Study:

• To propose a novel framework, AKD-KGC, for enhancing knowledge graph completion.
• To improve the integration of structural and descriptive information by addressing the limitations of multi-semantic entities.
• To facilitate knowledge transfer from structural models to pre-trained language models (PLMs).

Main Methods:

• Developed AKD-KGC, a framework incorporating Adaptive Knowledge Distillation (AKD) for feature integration.
• Employed a teaching-learning procedure to transfer multi-semantic knowledge from structural models to PLMs.
• Integrated features at the embedding level while using structural models to guide the integration model's prediction behavior.

Main Results:

• AKD-KGC effectively transfers entity multi-semantic knowledge, enhancing the integration of structural and descriptive information.
• The framework demonstrated state-of-the-art performance on multiple datasets in both transductive and inductive settings.
• Achieved superior results by improving PLMs' learning of entity semantics beyond descriptions.

Conclusions:

• AKD-KGC significantly improves knowledge graph completion by enabling effective knowledge transfer and feature integration.
• The proposed method enhances the understanding of multi-semantic entities, leading to state-of-the-art performance.
• AKD-KGC offers a robust solution applicable to both transductive and inductive KGC tasks.