7 Appendix

To obtain the edge precision, we compare the latent graphAij with the ground truth adjacency matrixBij. Weuse the Adam optimizer with alearning rate of 0.001 to minimize the cross-entropy loss. We use a randomized grid search training on paths of length 1 and validating hitset10 on paths of length2 for the L2 regularization of the entities and the relations between1e 20,...,1e 5 and for the dropout probabilities for the subject, object and relationsbetween0,...,0.8,respectively. We use discretization after every single relation. Hence, we obtain the logits for all possibleentities.