A Novel Embedding Model for Knowledge Graph Completion Based on Multi-Task Learning

Knowledge graph completion is the task of predicting missing relationships between entities in knowledge graphs. State-of-the-art knowledge graph completion methods are known to be primarily knowledge embedding based models, which are broadly classified as translational models and neural network models. However, both kinds of models are single-task based models and hence fail to capture the underlying inter-structural relationships that are inherently presented in different knowledge graphs. To this end, in this paper we combine the translational and neural network methods and propose a novel multi-task learning embedding framework (TransMTL) that can jointly learn multiple knowledge graph embeddings simultaneously. Specifically, in order to transfer structural knowledge between different KGs, we devise a global relational graph attention network which is shared by all knowledge graphs to obtain the global representation of each triple element. Such global representations are then integrated into task-specific translational embedding models of each knowledge graph to preserve its transition property. We conduct an extensive empirical evaluation of multi-version TransMTL based on different translational models on two benchmark datasets WN18RR and FB15k-237. Experiments show that TransMTL outperforms the corresponding single-task based models by an obvious margin and obtains the comparable performance to state-of-the-art embedding models.