Implementing Poincaré Embeddings (2017)

rare-technologies.com

74 points·bra-ket··8 comments

we had a similar adventure implementing the next thing which was lorentz embeddings. Some equations were wrong in the paper which AFAIK have not been fixed in subsequent revisions.

@dchatterjee172 found a lot of the inconsistencies in the work.

https://github.com/theSage21/lorentz-embeddings

Equations wrong in the original Poincare embeddings paper https://arxiv.org/abs/1705.08039 ? Do you recall which?

And thanks for sharing your Lorentz implementation!

we never implemented the poincare paper since by then Facebook had published it's own code, but for the lorentz paper there were a few which were wrong. I'll update the repo to point to those in a bit

There is an actual application of this on the frontpage right now:

https://news.ycombinator.com/item?id=20574182

Coincidence?

Why isn't this being used heavily in other NLP frameworks? All the newest Language Models seem to be throwing more compute at transformer models instead of moving away from Euclidean Space.

So its been ~2 years. Have the problems been sorted out? Have original authors published their implementation? Would enjoy an update on these efforts.

I really like how they describe the adventure of replicating results, rather than just taking the paper at face value.

Bummer that they could not replicate the results

Reply on news.ycombinator.com