Differences with Lorentz Embeddings

[gokul-uf]

Hi,
Thanks a lot for open-sourcing both this and the original poincare embeddings paper.

The authors of the original paper (poincare embeddings) have another trained on the hyperboloid.

Could you let me know what the differences between this repo and those paper are?

Thanks

[mleimeister]

Hi Gokul,

thanks a lot for your interest in the code. It indeed implements the model that is described in the second paper of Nickel/Kiela that you mention. The vector class represents points on the hyperboloid model and the function geodesic_update computes the exponential function given a point and a tangent vector. This is the same as equation 9 in the paper.

There is a small difference in the initialization, where we are sampling a tangent vector at the base point uniformly and map it by the exponential function vs. sampling the first n dimensions and fixing the last one to lie on the hyperboloid for Nickel/Kiela.

Also, there is no burnin implemented. To achieve this, you can run the training twice, as described here.

Let me know if you have further questions or encounter any problems with the code.

Best regards,
Matthias

[gokul-uf]

Thanks for the clarification! I'm training with the poincare embeddings codebase now and it seems much faster than the gensim implementation :) I'm planning to train with lorentz embeddings this week. I was wondering if you folks had any benchmarks / results comparing with the numbers from the lorentz embeddings paper.

Thanks