ChatRule

Mining Logical Rules with Large Language Models for Knowledge Graph Reasoning with 1 dollar.

Official Implementation of "ChatRule: Mining Logical Rules with Large Language Models for Knowledge Graph Reasoning".

Logical rules are essential for uncovering the logical connections between relations, which could improve reasoning performance and provide interpretable results on knowledge graphs (KGs). In this paper, we propose a novel framework, ChatRule, unleashing the power of large language models (LLMs) for mining logical rules over knowledge graphs with less than 1 dollar. The final rules can be used to conduct reasoning over KGs without additional model training.

Requirements

pip install -r requirements.txt

Set your OpenAI API key in .env file

Mining Rules with ChatRule

Please check examples of different datasets and LLMs in here.

1. Sample Paths

python path_sampler.py --dataset ${DATASET} --max_path_len 3 --anchor 100 --cores 8

2. Generate Rules

python chat_rule_generator.py --dataset ${DATASET} --model_name gpt-3.5-turbo -f 50 -l 10

3. Collect and Clean Rules

python clean_rule.py --dataset ${DATASET} -p gpt-3.5-turbo --model none

4. Rank Rules

python rank_rule.py --dataset ${DATASET} -p clean_rules/${DATASET}/gpt-3.5-turbo-top-0-f-50-l-10/none

5. Evaluate Completion

python kg_completion.py --dataset ${DATASET} -p ranked_rules/${DATASET}/gpt-3.5-turbo-top-0-f-50-l-10/none/all

Reproduce KGC results with mined rules.

python kg_completion.py --dataset family -p FinalRules/family
python kg_completion.py --dataset umls -p FinalRules/umls
python kg_completion.py --dataset wn-18rr -p FinalRules/wn-18rr
python kg_completion.py --dataset yago -p FinalRules/yago

Results

Bibinfo

If you found this repo helpful, please help us by citing this paper:

@article{luo2023chatrule,
  title={Chatrule: Mining logical rules with large language models for knowledge graph reasoning},
  author={Luo, Linhao and Ju, Jiaxin and Xiong, Bo and Li, Yuan-Fang and Haffari, Gholamreza and Pan, Shirui},
  journal={arXiv preprint arXiv:2309.01538},
  year={2023}
}

Acknowledgement

The code of KGC reasoning in this work is mainly based on NCRL with a bug in ranking function fixed. We thank the authors for their great works.