This is an implementation of the model from the paper Probabilistic Logic Neural Networks for Reasoning.
In our repo, four benchmark datasets are provided, including FB15k, FB15k-237, WN18, WN18RR. Those datasets are available in the data folder. The folder kge provides the codes for knowledge graph embedding, and the folder mln gives an implementation of the Markov logic network, in which four rule patterns are considered, including the composition rule, symmetric rule, inverse rule and subrelation rule.
Since the MLN module is written in C++, we need to compile the MLN codes before running the program. To compile the codes, we can enter the mln folder and execute the following command:
g++ -O3 mln.cpp -o mln -lpthread
Afterwards, we can run pLogicNet by using the script run.py in the main folder.
During training, the program will create a saving folder in record to save the intermediate outputs and the results, and the folder is named as the time when the job is submitted. For each iteration, the program will create a subfolder inside the saving folder. In each subfolder, the result of pLogicNet on validation set, the result of pLogicNet on test set and the result of pLogicNet* on test set are saved into result_kge_valid.txt, result_kge.txt and result_kge_mln.txt respectively. Based on the validation results, we can then pick up the best model, and use it for evaluation or apply it to other knowledge graphs for link prediction.
The knowledge graph embedding codes in the kge folder are from the nice repo KnowledgeGraphEmbedding, where many knowledge graph embedding algorithms are implemented.
Please consider citing the following paper if you find our codes helpful. Thank you!
@inproceedings{qu2019probabilistic,
title={Probabilistic Logic Neural Networks for Reasoning},
author={Qu, Meng and Tang, Jian},
booktitle={Advances in neural information processing systems},
year={2019}
}