Relational Knowledge Prediction via Dynamic Bi-Mode Embedding

Knowledge graphs are a crucial concept in artificial intelligence with a wide spectrum of real-life applications. Nonetheless, they are currently suffering from the incompleteness issue, i.e., relational knowledge in the graphs may not yet meet the practical needs. To address this issue, mainstream solutions propose to predict links by using compositional models or translation models. However, the prediction accuracy is still of particular concern. In this paper, we propose a new method, namely, Bi-Mult, which combines the advantages of compositional models and translation models. Bi-Mult is based on the compositional model, such that an entity (resp. relation) embedding is decomposed into two parts, one is to represent intraentity (resp. relation) state and the other is for inter-entity (resp. relation) state, and we call such embedding as bi-mode embedding. In addition, the bi-mode relation embedding enhances relation's interaction with entities, resulting its improvement on handling antisymmetric relations. Moreover, we incorporate mapping matrices in translation models through bi-mode entity embedding to construct dynamic embeddings for expressing complex relations, such as “1-to-N”, “N-to-1,” and “N-to-N” relations. In experiments, we evaluate our method on the benchmark data sets and the task of link prediction, and our method is demonstrated to outperform state-of-the-art methods consistently and significantly.