Awesome

KG-FIT

This repository contains the code for the paper "KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge" (NeurIPS 2024). Paper Link

alt text

Data Preparing & Precompute:

To enable precompute, you need to put a file named "openai_api.key" (with your OpenAI API key in there) under code/precompute, then run the following command with a specified dataset (FB16K-237 in this case):

cd code/precompute
python cluster.py --dataset FB15K-237 --output_dir ../../processed_data  # precomputation for seed hierarchy
cd llm_refine
python llm_refine.py --dataset FB15K-237  --model gpt-4o-2024-05-13 # LLM-Guided Hierarchy Refinement (LHR)
cd ..
python cluster.py --dataset FB15K-237  --output_dir ../../processed_data # precomputation for llm hierarchy

where the first call of cluster.py is used to build seed hierarchy; llm_refine.py is used to refine the seed hierarchy with LLM; The second call of cluster.py is used to build the final hierarchy with LLM.

KG-FIT Training & Evaluation:

Use the scripts runs_xxx.sh to run the experiments for all the models. For example:

bash runs_rotate.sh
bash runs_tucker.sh

We provide several variants of KG-FIT model under the `code` folder:

File	KG-FIT with KGE base models	Text and Hierarchical Constraints	Text Embedding within Entity Embedding
`model_common.py`	All models except TuckER and ConvE	On negative batches	Frozen
`model_flex.py`	All models except TuckER and ConvE	On negative batches	On Fire
`model_p_anc.py`	All models except TuckER and ConvE	On both positive and negative batches	Frozen
`model_tucker_conve.py`	KG-FIT-TuckER and KG-FIT-ConvE	On both positive and negative batches	Frozen

Cite KG-FIT

@inproceedings{jiang2024kgfit,
  title = {KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge},
  author = {Jiang, Pengcheng and Cao, Lang and Xiao, Cao and Bhatia, Parminder and Sun, Jimeng and Han, Jiawei},
  year = {2024},
  booktitle = {The Thirty-Eighth Annual Conference on Neural Information Processing Systems},
}

Thank you for your interest in our work!