This repository provides an implemenation of Lorentz embeddings as described in "Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry" (Nickel and Kiela, 2018). Note that this is not their official implementation but an attempt to recreate and visualize the results presented in the paper.

• Python 3.5
• PyTorch 0.4

• Implement the basic model
• Get model to converge on WordNet mammals
• Visualize results in Poincaré space
• Train the model on WordNet noun hierarchy
• Obtain same mAP/MR on WordNet nouns (currently 8.97 MR and 0.7315 mAP with d == 5)
• Obtain same mAP/MR on WordNet verbs
• Obtain same mAP/MR on EuroVoc Data
• Obtain same mAP/MR on ACM data
• Obtain same mAP/MR on MeSH data

Training visualization of 3D Lorentz embeddings trained on Wordnet mammals projected into the 2D Poincaré disk.

These heatmaps show how the distance between points scales as a point approaches the edge of the manifold.