In the active research area of employing embedding models for knowledge graph completion, particularly for the task of link prediction, most prior studies used two benchmark datasets FB15k and WN18 in evaluating such models. Most triples in these and other datasets in such studies belong to reverse and duplicate relations which exhibit high data redundancy due to semantic duplication, correlation or data incompleteness. This is a case of excessive data leakage---a model is trained using features that otherwise would not be available when the model needs to be applied for real prediction. There are also Cartesian product relations for which every triple formed by the Cartesian product of applicable subjects and objects is a true fact. Link prediction on the aforementioned relations is easy and can be achieved with even better accuracy using straightforward rules instead of sophisticated embedding models. A more fundamental defect of these models is that the link prediction scenario, given such data, is non-existent in the real-world. This paper is the first systematic study with the main objective of assessing the true effectiveness of embedding models when the unrealistic triples are removed. Our experiment results show these models are much less accurate than what we used to perceive. Their poor accuracy renders link prediction a task without truly effective automated solution. Hence, we call for re-investigation of possible effective approaches.

Incompleteness of large knowledge graphs (KG) has motivated many researchers to propose methods to automatically find missing edges in KGs. A promising approach for KG completion (link prediction) is embedding a KG into a continuous vector space. There are different methods in the literature that learn a continuous representation of KG (latent features of KG). The benchmark dataset FB15k has been widely employed to evaluate these methods. However, It has been noted that FB15k contains many pairs of edges in which a pair represents the same relationship in reverse directions. Therefore, the inverse of numerous test triples occurs in the training set. To address this problem, FB15k-237, a subset of FB15k, was created by removing those inverse-duplicate relations to form a more challenging, realistic dataset. There is not any study that investigates how the aforementioned bias in this widely used benchmark dataset affects the results of embedding-based knowledge graph completion methods and whether their promising results are largely due to the bias. Motivated by this question, we conducted extensive experiments and report the link prediction results on FB15K and FB15k-237 using several embedding-based methods. We compare the results of different methods to see how their performances change in absence of inverse relations. Our experiment results demonstrate that the performance of embedding models in link prediction task diminishes tremendously when the inverse relationships do not exist anymore.