Zack and Divyansh's work is excellent; I recommend you all go look at it. Our work provides a different, concurrent take on similar issues. Some key differences: https://twitter.com/zacharylipton/status/1247357810410762240

(1) We take a geometrical view of this problem, instead of a causality view. They are both interesting ways to view the problem, I think. Personally, I prefer the geometric view; it's simple, intuitive, and self-contained.

(2) We demonstrate the solution on a wider variety of NLP tasks, from long-standing problems like dependency parsing to newer reading comprehension datasets.

(3) Because we're looking at expert annotations here, we really don't think it's feasible to do this for training data on most problems that we care about, so we focused on evaluation.

One final note: Zack mentions that he thinks we misuse the word "concurrent" in our paper. Our paper was finished two months after theirs. Why is it showing up on arxiv six months later? Because of differences in publication norms across communities. This is a serious problem.

And while Zack thinks he knows when this project started because we invited him and Divyansh to collaborate on it when we were scaling our ideas to lots of datasets, he's just mistaken. "Concurrent" is indeed accurate; this project started long before their work was posted.

And I think that's fine. There's nothing wrong with independent groups coming to similar conclusions, and giving different takes on the same problem. It makes the case stronger that this is a good idea. I don't feel threatened by having similar work published close to mine.

That we often feel so defensive about other people's good ideas is a really big problem with incentives in academia. Our system is pretty broken.