Knowledge graph embedding aims to learn representations of entities and relations in a knowledge graph. Recently, QuatE has introduced the graph embeddings into the quaternion space. However, there are still challenges in dealing with complex patterns, including 1-N, N-1, and multiple-relations between two entities. Since the learned entity embeddings tend to overlap with each other in the first two cases, and the learned relation embeddings tend to overlap with each other in the last case. To deal with these issues, we propose QuatSE, a novel knowledge embedding model that adjusts graph embeddings via spherical linear interpolation (Slerp) of entities and relations. For a triple (head entity, relation, tail entity), QuatSE calculates Slerp between each entity and its relation, and adds the normalized interpolation to the corresponding entity. The operation avoids the problem of embedding overlap and ensures the information of original entity is not missed. We further compare the effect of interpolation using different normalization strategies (L1 or L2 ) for Slerp. Several experiments suggest that QuatSE works well in 1-N, N-1 and multiple-relations pattern. QuatSE outperforms the existing quaternion-based models.
CITATION STYLE
Li, J., Su, X., Ma, X., & Gao, G. (2022). QuatSE: Spherical Linear Interpolation of Quaternion for Knowledge Graph Embeddings. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 13551 LNAI, pp. 209–220). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-17120-8_17
Mendeley helps you to discover research relevant for your work.