NCRL

Introduction

The official Pytorch implementation of the paper Neural Compositional Rule Learning for Knowledge Graph Reasoning

KG Data:

• entities.txt: a collection of entities in the KG
• relations.txt: a collection of relations in the KG
• facts.txt: a collection of facts in the KG
• train.txt: the model is trained to fit the triples in this data set
• valid.txt: create a blank file if no validation data is available
• test.txt: the learned ryles is evaluated on this data set for KG completion task

Usage

For example, this command train a NCRL on family dataset using gpu 0

  python main.py --train --test --data family --max_path_len 4 --model family --gpu 0 --get_rule --topk 500

Each parameter means:

• --train: train the model
• --test: assign score to each rule in the rule space
• --max_path_len: the maximum length of paths observed during training
• --get_rule: output the learned rules
• --data: dataset
• --topk: number of the output rules
• --model: where do we save our model