BiQCap: A Biquaternion and Capsule Network-Based Embedding Model for Temporal Knowledge Graph Completion

Sensen Zhang, Xun Liang, Zhiying Li, Junlan Feng, Xiangping Zheng, Bo Wu

Abstract: Temporal Knowledge Graphs (TKGs) provide a temporal context for facts, capturing temporal information and the dynamic nature of actual world facts. However, typical TKGs often suffer from incomplete dynamics with missing facts in real-world scenarios. Temporal Knowledge Graph Embedding (TKGE) is one of the critical approaches to tackling the challenge. However, the existing TKGE models are weak in simultaneously representing hierarchical semantics and other relation patterns. Therefore, embedding TKGs in a single space, no matter the Euclidean space, or hyperbolic space, cannot capture the complex structures of TKGs accurately. In addition, few existing models have a “deep” architecture for modeling the entries in a quadruple at the same dimension. In this paper, we propose a new TKGE model, $$\textbf{BiQCap}$$ , which for the first time, combines biquaternion and capsule network in modeling to make up for the defects of existing TKGE models. BiQCap represents each temporal entity as a translation and each relation as euclidean rotation and hyperbolic rotation in biquaternions vector space. Further, we employ the embeddings of entities, relations, and temporal trained from biquaternions as the input to capsule networks. Experimental results on five well-known benchmark datasets show that our BiQCap achieves state-of-the-art performance.