Universal Embeddings for Knowledge Graphs

This repository contains code related to the scientific article Universal Knowledge Graph Embeddings.

Code structure

The code of the Universal Embeddings approach comprises this main repository Universal_Embeddings and two sub-repositories: dice-embeddings and embeddings.cc. They are briefly introduced in the following and documented in the repositories themselves.

• Main repository: Universal_Embeddings
  This repository contains two notebooks: merge.ipynb and Visualize_Embeddings.ipynb. The first notebook implements the merge (fusion) of DBpedia and Wikidata via sameAs links. The second notebook visualizes the embeddings of the same entities across different knowledge graphs. The computation of universal embeddings is based on the dice-embeddings framework. We also develop the embeddings.cc repository to provide universal embeddings as a service.
• Sub-repository: dice-embeddings
  We use the dice-embeddings repository to compute universal knowledge graph embeddings. It implements several knowledge graph embedding models, including ConEx, QMult, ComplEx and DistMult.
• Sub-repository: embeddings.cc
  The embeddings.cc repository was implemented on top of our computed Universal Embeddings for the artivle Universal Knowledge Graph Embeddings. It implements the online API for providing the embeddings data. The API is deployed at https://embeddings.cc.

Resulting data

On https://embeddings.cc and also in a dump file on Zenodo (DOI: 10.5281/zenodo.7566020) we provide roughly 180 million entity emebddings from the merge of DBpedia (DBpedia Snapshot 2022-09) and Wikidata (Wikidata dumps). These embeddings are computed using the ConEx embedding model which was chosen based on its performance on our experiments in article table 3.

Installation

• Start with cloning this repository:

git clone --recursive https://github.com/dice-group/Universal_Embeddings.git

Note the flag --recursive which is used to clone the submodules. This is necessary to ensure that the code runs as each of the submodules have undergone refactoring.

• Make sure Ananconda is install on your machine.
  Create a working environment and activate it:

conda create -n unikge python=3.9.12
conda activate unikge

• Install all dependencies in requirements.txt:

pip install -r requirements.txt

Training Universal Embeddings

Reproducing the reported paper results (Table 3 in the article)

1. Download the evaluation datasets zip file and extract it.

2. Make sure dice-embeddings is cloned and switch to the CLF branch using git checkout CLF then git checkout .

3. To reproduce evaluation results (Table 3) in the paper, enter in the dice-embeddings repository with cd dice-embeddings and run the following command: python main.py --path_dataset_folder {path_to_kg_folder} --model {model_name} --batch_size 8192 --embedding_dim 32 --eval train_test --num_epochs 500
   Inside the command, {path_to_kg_folder} is the path to DBpedia, DBpedia+, Wikidata, or Wikidata+ which are downloaded above. {model_name} is the embedding model name, i.e., ConEx, DistMult, ComplEx, or QMult.`

Remark: It is also possible to run all experiments with a single bash script. For this, just execute ./execute_training from the parent directory Universal_Embeddings

Computing universal embeddings on other knowledge graphs

1. Merge your knowledge graphs using Algorithm 1 in the paper (page 7). For large knowledge graphs, we recommend saving the merged KG in parquet format to save memory and enable fast reading and processing with libraries such as polars.

2. Run python main.py --path_dataset_folder {path_to_kg_folder} --model {model_name} --batch_size {e.g. 1024} --embedding_dim {e.g. 32} --eval train_test --num_epochs {e.g. 200}

3. Call .get_embeddings() on the trained model to obtain entity and relation embeddings.