arXiv submission

[pminervini]

There are a few things that still need to be done, that don't justify another paper, but would round things up and could be helpful in case of rejection.

What are these? For example, running on more datasets, and maybe more models.

[pminervini]

Experiments:

• The "triple classification" experiments on WN18 and FB122 in the KALE paper [1] (the code is already there).
• Injecting schema in some model without closed form solutions, such as Knowledge Vault's ER-MLP [2].
• Countries, for an analysis similar to the one in the HolE paper [3].

[1] https://www.aclweb.org/anthology/D/D16/D16-1019.pdf [2] https://www.cs.ubc.ca/~murphyk/Papers/kv-kdd14.pdf [3] https://cbmm.mit.edu/sites/default/files/publications/holographic-embeddings.pdf

Formalisation/Notation:

I'm having difficulties describing the "adversarial input set" (pg. 4, right column) as a set of pairs - can it be considered as a variable embedding layer ?

[riedelcastro]

For the arxiv submission, I'd be a little careful not to add too many new experiments (or any), really. Reasons:

1) In case the UAI paper gets accepted, it's a little risky to add more results in the camera-ready versions, as the reviewers haven't seen these.

2) This can go on indefinitely, and we never move on.

3) Things that don't justify a new paper are good for technical reports or journal papers.

So personally, I'd prefer focussing on polishing polishing polishing now, preparing a blog post/distil.pub or the like, and then do more experiments etc. when we hear from UAI.