CRL-MM: Context-Aware Relational Learning and Multidimensional Matching for Few-Shot Knowledge Graph Completion.

Area of Science:

• Artificial Intelligence
• Data Science

Background:

• Few-shot knowledge graph completion (FKGC) models often struggle with the diverse semantics of relations in real-world knowledge graphs.
• Existing models primarily focus on aggregating entity representations, limiting their ability to capture context-dependent relational information.

Purpose of the Study:

• To propose a novel FKGC model, context-aware relational learning and multidimensional matching (CRL-MM), designed to address the limitations of existing models in handling multi-semantic relations.
• To enhance the representation of task relations by incorporating semantic information from different contexts and improve entity encoding through a unified contextual approach.

Main Methods:

• CRL-MM enhances task relation representations by leveraging semantic similarity between task and background relations.
• It encodes entities by treating the entity pair and its neighborhood as a unified whole, using adaptive task relations and paired entity awareness.
• A multidimensional matching network is employed, considering both entity pair similarity and triple rationality for improved generalizability.

Main Results:

• CRL-MM demonstrates superior performance compared to state-of-the-art methods on public benchmark datasets.
• Ablation experiments confirm the effectiveness of individual modules within the CRL-MM model.
• The proposed model successfully captures diverse semantic information of relations in different contexts.

Conclusions:

• The CRL-MM model offers a significant advancement in few-shot knowledge graph completion by effectively handling multi-semantic relations.
• Its context-aware approach and multidimensional matching contribute to improved accuracy and generalizability in knowledge graph inference.