Towards the expressive power of translating approach for knowledge graph completion

The pessimistic conclusions from previous research on the Expressive Power of translating approaches for knowledge graph completion are investigated and rethought. To this end, a novel model RosE is formulated by introducing two degrees of freedom and outperforms traditional translation-based models on widely used datasets such as FB15k, WN18, FB15k237, and WN18RR. Every new freedom is a vector in the model, the operation of which multiplies with entity and relation embeddings, rotating them to a new position. Consequently, the head entity and relation embedding are equal to the tail entity. Fortunately, the intrinsic limitations merely exist in this research line when the model is trained in real vector space, not in other spaces such as trigonometric functions and complex. The experimental and theoretical results, together with the newly proposed model RosE, also confirm this conclusion. Therefore, the findings in this work do not discourage further exploration in this research line, but rather avoid those with discouraging outcomes. In short, this paper clarifies that the limitations of the translation approach for knowledge graph completion are specific conditions that only involve partial models. That is, the research line of translation approach is still promising when certain known pitfalls are avoided.