This repository contains the code used in the research paper titled "Empowering Small-Scale Knowledge Graphs: A Strategy of Leveraging General-Purpose Knowledge Graphs for Enriched Embeddings" authored by Albert Sawczyn, Jakub Binkowski, Piotr Bielak, and Tomasz Kajdanowicz. The paper was accepted for the LREC-COLING 2024 conference.
You can find the paper here - TBA.
TBA- Configured Python 3.10 environment.
- Installed Poetry.
To install all dependencies, run:
poetry install The repository uses DVC to manage the data. To download the data, run:
dvc pullThe repository uses DVC to manage the experiments.
dvc.yamlcontains the stages used before the training and after the training. Each stage is described in the comments.experiments/configs/training/dvc.yamlcontains the training stages.
To reproduce all stages run:
dvc reproThe repository uses WandB to log the experiments. Before running the experiments, change the
wandb_entity and wandb_project in
experiments/configs/training/{WN18RR.yaml,FB15k237.yaml,WD50K.yaml} or set the environment
variable WANDB_MODE=offline.
The logs can be found here.
Keep in mind that the metric values calculated during training can differ from those calculated after training is completed. This is due to the fact that evaluation during training is done without restricting entities to DKG.