Towards the expressive power of translating approach for knowledge graph completion

Area of Science:

• Artificial Intelligence
• Data Science

Background:

• Previous research cast doubt on the expressive power of translation-based models for knowledge graph completion.
• These models are crucial for understanding and structuring complex data relationships.

Purpose of the Study:

• To investigate and reevaluate the limitations of translation-based approaches in knowledge graph completion.
• To propose a novel model that overcomes identified limitations and enhances performance.

Main Methods:

• Formulation of a new model, RosE, incorporating two degrees of freedom.
• Introducing vector operations that rotate entity and relation embeddings.
• Testing performance on standard datasets: FB15k, WN18, FB15k237, and WN18RR.

Main Results:

• The RosE model demonstrated superior performance compared to traditional translation-based models.
• Identified that intrinsic limitations are specific to training in real vector space.
• Showcased that alternative spaces (e.g., trigonometric, complex) do not exhibit these limitations.

Conclusions:

• The translation approach for knowledge graph completion remains promising when specific pitfalls are avoided.
• Limitations are conditional and tied to particular model implementations, not the entire research line.
• Further exploration in this area is encouraged, focusing on successful strategies and avoiding known limitations.