RulE: Neural-Symbolic Knowledge Graph Reasoning with Rule Embedding

This is an implementation of the RulE model for knowledge graph reasoning. RulE is a novel and principled framework to represent and model logical rules and triplets. It jointly represents entities, relations and logical rules in a unified embedding space. By learning rule embeddings, it can improve the generalization capability of KGE and improve logical rule reasoning.

Overview

This is the overall architecture of RulE, consisting of three components: pre-training, grounding, inference.

Datasets

Datasets used are contained in the folder data/. The format is as follows:

train.txt, valid.txt, test.txt: training, valid, test set with triplets; Format: (h,r,t).

mined_rules.txt: These rules are mined by RNNLogic. Format: [rule_head, rule_body_list]. For example, $r_1 \land r_2 \rightarrow r_3$ can be represented as $[r_3,r_1,r_2]$

Installation

First clone the repository, then install the package dependency in the requirement.txt with the following command:

conda create -n RulE python=3.8.0
pip install -r requirement.txt

Usage

The folder config/ contain the configuration file of each dataset and the folder src/ provides the implementation of RulE. You can edit the config file and enter the folder src/ to excute the following command:

python main.py --init ../config/umls_config.json

Results

Using the default configuration files, we can achieve the results reported on paper.