The primary aim of this research is to develop a flexible and general Knowledge Graph Embedding (KGE) model that can effectively represent various relation patterns. The authors introduce OrthogonalE, a novel KGE model that addresses the limitations of existing rotation-based methods, such as RotatE and QuatE.
The key aspects of the OrthogonalE approach are:
Entity Representation: The model transforms entity vectors into matrices to regulate the entity dimension and avoid unnecessary expansion of the relation size.
Relation Representation: OrthogonalE employs block-diagonal orthogonal matrices for relations, which enhances the generality of the model by allowing for higher-dimensional rotations. This approach also captures several relation patterns, including Symmetry, Antisymmetry, Inversion, and Non-commutative Composition.
Optimization: The model separately optimizes relations using Riemannian optimization and entities using Stochastic Gradient Descent (SGD), leading to more effective training.
Experimental results on the WN18RR and FB15K-237 datasets demonstrate that OrthogonalE outperforms state-of-the-art KGE models in terms of link prediction accuracy while substantially reducing the number of relation parameters. The model's ability to adapt to datasets with varying complexities by adjusting the dimension of the block-diagonal matrices highlights its generality. Additionally, the authors provide visualizations and analyses to verify OrthogonalE's capability in capturing diverse relation patterns.
To Another Language
from source content
arxiv.org
Önemli Bilgiler Şuradan Elde Edildi
by Yihua Zhu, H... : arxiv.org 10-03-2024
https://arxiv.org/pdf/2401.05967.pdfDaha Derin Sorular