GeoAlign

Source code for AKBC 2021 paper Manifold Alignment across Geometric Spaces for Knowledge Base Representation Learning.

Requirements

• python: 3.6
• torch: 1.0.0
• scikit-learn
• pandas
• cython
• tqdm
• numpy

The hyperbolic embeddings in this repository (the poincare-embeddings folder) is inherited from the poincare-embeddings repository.

Data

We construct two taxonomies (YAGOwordnet and wikiObjects) and one knowledge graph (YAGOfacts) from YAGO3. Please refer to our paper for the data construction details.

Taxonomy data

• taxonomy.csv: This file contains the edges (i.e., the hypernymy relations) of the taxonomy.
• full_taxonomy.csv: This file contains the edges of the full transitive closure of the taxonomy, which is used in our experiments.
• full_transitive.txt / basic_edges.txt: The edges in the full transitive closure of the taxonomy / the transitive reduction of the taxonomy.
• types.txt / entities.txt / taxonomy_nodes.txt: The types / entities / union of types and entities in the taxonomy.
• taxonomy: The folder contains the training set and test set under different training rates.

Knowledge graph data

• TransE_KG: The folder contains the pretrained TransE embeddings of YAGOfacts named 'TransE.ckpt'. Please obtain a pairwise distance matrix of the entity embeddings and store it in this folder before running the code.

Usage

To run the whole framework, run:

zsh run_all.sh

or

echo $Training_Rate"\n"$Data"\n"$Dimension"\n"$Hyperbolic_Model"\n" | xargs -L 4 -P $PARALLEL_R bash hyper_label_rate.sh

where $Training_Rate = {1, 2, 3, 4, 5}; $Data = {YAGOwordnet, wikiObjects}; $Dimension is the embedding dimension of the hyperbolic space; $Hyperbolic_Model = {lorentz, poincare}; $PARALLEL_R is the number of parallel programs. More parameters can be edited in constants.sh.

Citation

If you find this repository useful for your research, please kindly cite our paper:

@inproceedings{
xiao2021manifold,
title={Manifold Alignment across Geometric Spaces for Knowledge Base Representation Learning},
author={Huiru Xiao and Yangqiu Song},
booktitle={3rd Conference on Automated Knowledge Base Construction},
year={2021},
url={https://openreview.net/forum?id=TPymTKJR-Pi}
}