Capsule neural tensor networks with multi-aspect information for Few-shot Knowledge Graph Completion.

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Few-shot Knowledge Graph Completion (FKGC) aims to expand relation coverage in knowledge graphs with limited data.
• Current FKGC methods often rely on random one-hop neighbor selection, overlooking rich multi-aspect entity information.
• Existing approaches using Long Short-Term Memory (LSTM) for relation embedding are sensitive to input order.

Purpose of the Study:

• To propose InforMix-FKGC, a Capsule Neural Tensor Network approach leveraging multi-aspect information for improved FKGC.
• To address limitations of random neighbor selection and input order sensitivity in prevailing FKGC methods.

Main Methods:

• InforMix-FKGC utilizes a value-based one-hop neighbor selection strategy.
• It encodes multi-aspect entity information, including neighbors, attributes, and literal descriptions.
• A capsule network integrates the support set for relation embedding, and a neural tensor network matches query and support sets.

Main Results:

• InforMix-FKGC demonstrates superior performance in Few-shot Knowledge Graph Completion.
• Experiments on NELL-One and Wiki-One datasets show significant improvements over ten state-of-the-art methods.
• The approach enhances FKGC accuracy by learning few-shot relation embeddings more precisely.

Conclusions:

• InforMix-FKGC effectively addresses limitations in current FKGC techniques.
• The proposed method offers a more precise way to learn few-shot relation embeddings.
• This advancement contributes to more accurate and comprehensive knowledge graph completion.