RNNLogic

This is an implementation of the RNNLogic model for knowledge graph reasoning.

Introduction

RNNLogic focuses on knowledge graphs, which are collections of real-world facts, with each fact represented by a (h,r,t)-triplet. As collecting facts is expensive, knowledge graphs are imcomplete, and thus predicting missing facts in a knowledge graph is an important problem with growing attention. Such a problem is known as knowledge graph reasoning.

RNNLogic solves knowledge graph reasoning by learning logic rules, which have been proved to improve the interpretability and precision of reasoning. To do that, RNNLogic employs a rule generator and a reasoning predictor. The rule generator is parameterized by a RNN, which is able to model and generate chain-like rules. The reasoning predictor follows stochastic logic programming, which uses a set of logic rules as input to predict the answers of queries. Given a query, the rule generator generates a set of logic rules, which are fed into the reasoning predictor. The rule generator further applies the logic rules to the existing knowledge graph for predicting the answer.

To optimize the reasoning predictor and the rule generator, we propose an EM-based algorithm. At each iteration, the algorithm starts with generating a set of logic rules, which are fed into the reasoning predictor and we further update the reasoning predictor based on the training queries and answers.

Then in the E-step, a set of high-quality logic rules are selected from all the generated logic rules according to their posterior probabilities.

Finally in the M-step, the rule generator is updated to be consistent with the high-logic logic rules identified in the E-step.

Data

We provide four datasets for knowledge graph reasoning, and these datasets are FB15k-237, WN18RR, Kinship, UMLS. For FB15k-237 and WN18RR, there are standard splits for the training/validation/test sets. For Kinship and UMLS, we split the training/validation/test sets by ourselves, and the details are available in our paper.

Usage

We provide two versions of RNNLogic implemetations.

Version 1

In the folder codes_toy, we provide a toy implementation, which only implements the model RNNLogic w/o emb. This toy implementation is easy to understand, which can be used to reproduce the results of RNNLogic w/o emb reported in the paper.

To run this code, please first go to the directory codes_toy/pyrnnlogiclib and run the following command:

python setup.py install

After that, you can go back to the main directory and use the following commands to do knowledge graph reasoning:

python run.py --data_path ../data/wn18rr --num_generated_rules 200 --num_rules_for_test 200 --num_important_rules 0 --prior_weight 0.01 --cuda
python run.py --data_path ../data/kinship --num_generated_rules 2000 --num_rules_for_test 200 --num_important_rules 0 --prior_weight 0.01 --cuda
python run.py --data_path ../data/umls --num_generated_rules 2000 --num_rules_for_test 100 --num_important_rules 0 --prior_weight 0.01 --cuda

Version 2

In the folder codes, we provide the full implementation of both RNNLogic w/o emb and RNNLogic with emb. The codes can be used to reproduce the results of RNNLogic with emb reported in the paper. Note that in order to speed up training, this implementation trains a separate model for each relation, which allows us to deal with all the relations in parallel.

To run this code, you first need to download the pretrained entity and relation embeddings. You could do that through the link. Afterwards, please put the folder FB15k-237/RotatE_500 in data/FB15k-237 and put the folder wn18rr/RotatE_200 in data/wn18rr. Finally, please edit the script codes/run.py and use this script for model training.

Citation

Please consider citing the following paper if you find our codes helpful. Thank you!

@inproceedings{qu2020rnnlogic,
  title={RNNLogic: Learning Logic Rules for Reasoning on Knowledge Graphs},
  author={Qu, Meng and Chen, Junkun and Xhonneux, Louis-Pascal and Bengio, Yoshua and Tang, Jian},
  booktitle={International Conference on Learning Representations},
  year={2021}
}