Embeddings as Probabilistic Equivalence in Logic Programs

Code to replicate the paper "Embeddings as Probabilistic Equivalence in Logic Programs", published at NeurIPS2025.

Installation

The code has only been tested on Linux with Python3.10.12. First, install GLOG (see also here).

git clone https://github.com/karmaresearch/glog-python.git
cd glog-python
mkdir build
cd build
cmake ..
make

Second, clone this repo and install the Python dependencies.

pip install -r requirements.txt

It's important that the folder of this repo and glog-python are in the same directory, as a relative import is used.

Usage

The data and configuration files are stored under /data. To start a knowledge graph experiment, run

python run_kg.py data/[KG]/config.json --eval_split=test

You can also use wandb to start a sweep over 10 random seeds.

wandb sweep data/[KG]/seed.yaml

Similarly, a hyperparameter sweep is started using

wandb sweep data/[KG]/sweep.yaml

To run the finite state machine experiment, use

python run_fsm.py [LANGUAGE] --seed 1

Paper

@inproceedings{
maene2025embeddings,
title={Embeddings as Probabilistic Equivalence in Logic Programs},
author={Jaron Maene and Efthymia Tsamoura},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
}