To extract the entities and understand the relations in knowledge graphs is probably one of the challenging research focuses recently in machine learning community. In this paper, we aim to study one of the narrow-down problem of learning the embedding of entities and relations in knowledge graphs. From using a 3D rotation transformation in high dimensional vector spaces, we present a new method for knowledge graph embedding named Rotat3D. More specifically, the 3D-valued embeddings will be used to represent for the entities in the graphs, in which the rotations are modeled as popular rotation formulation in 3D vector spaces. Experimental results, carried out on four common benchmark datasets for link prediction, have shown that our proposed Rotat3D method is able to infer the common relation patterns in a graph more easily, and also has a critical improvement compared with some other state-of-the-art methods.