662a2e96162905620397b19c9d249781-Supplemental.pdf

However,itseffectonknowledgegraph completion task remains unknown. We further compare the performance of ConE with one that does not use cone restricted rotation for modeling hierarchical relations, which we name asRotC. ConE w/o rotation is the model that applies restricted rotation in the whole embedding space for hierarchical relations. Due to larger number ofdimensions used persubspace, weuseoverlapping subspace strategytoassign relation-specific subspaces. One of the main benefits of learning embeddings in hyperbolic space is that it can model well even in low embedding dimensionalities.