Feedback

Great summary! As I was reading, I thought the idea of the second paper, but worried the introduction of those new vectors per relation would lead to overfitting and to extremely high computational costs. It seems like the authors addressed the overfitting issue, which is great. Did they measure the increase of computational costs or address it in any way?

Hey Lucca, they have not directly addressed the issue of computational costs. Nonetheless, if you look at the last equation in the page, you can modify the number of vectors which you build your weights from (parameter b). In this way you can control the computational costs. However the authors didn't measure or add a section on that issue.