Open-World Taxonomy and Knowledge Graph Co-Learning (AKBC'22)

Data and code of AKBC'22 paper "Open-World Taxonomy and Knowledge Graph Co-Learning". For any suggestion/question, please feel free to create an issue or drop an email @ ([email protected] and [email protected]).

Table of Contents

• Datasets
• Prerequisites
• Reproduce Results

Datasets

The six datasets used in paper can be downloaded from https://figshare.com/articles/dataset/Taxo-KG-Bench/16415727.
After downloading and decompressing them under ./data/ directory, the directory looks like:

📁 ./data/CGC-OLP-BENCH
|-- 📁 MSCG-OPIEC
|-- 📁 MSCG-ReVerb
|-- 📁 SEMedical-OPIEC
|-- 📁 SEMedical-ReVerb
|-- 📁 SEmusic-OPIEC
|-- 📁 SEMusic-ReVerb
|-- CHANGELOG

And each subdirectory represents one of the six datasets, for instance:

📁 ./data/CGC-OLP-BENCH/MSCG-OPIEC
|-- oie_triples.train.txt
|-- oie_triples.dev.txt
|-- oie_triples.test.txt
|-- cg_pairs.train.txt
|-- cg_pairs.dev.txt
|-- cg_pairs.test.txt

where oie_triples.*.txt indicate the OpenKG triples after alginment, and cg_pairs.*.txt indicate the AutoTAXOnomy pairs after aligment.

Prerequisites

python>=3.7.10
conda>=4.10.3

The packages used in this project is listed in ./taxoKG_conda_env.yml. You can re-create and/or reactivate it by the following scripts:

# Re-create the environment. This will install the env *taxoKG* under your default conda path
conda env create --file taxoKG_conda_env.yml
# Or if you want to specify the env path
conda env create --file taxoKG_conda_env.yml -p /home/user/anaconda3/envs/env_name

# Reactivate the environment
conda activate taxoKG

Reproduce Results

The porposed HAKEGCN model checkpoints can be downloaded from https://figshare.com/articles/software/HakeGCN_Checkpoints/17108342.
Please decompress them under ./checkpoints folder.

Examples of getting the reported numbers for the main experiments (general, medical and music domain TaxoKG completion result tables):

sh run_reproduce.sh

Notes for using run_reproduce.sh:
L2-L4 just contains the set-ups for SLURM. I add it because I am working on a SLURM controled GPU server. You don't need to change anything to run it.

#SBATCH --job-name=Reproduce-TaxoKG 
#SBATCH --gres=gpu:1 
#SBATCH --output=logs/slurm_reproduce_HAKEGCN 

Citing Our Work

If you use our data or code in a scientific publication, please cite the following paper:

@inproceedings{lu22:hakeGCN,
  author     = {Jiaying Lu and
                Carl Yang},
  title      = {Open-World Taxonomy and Knowledge Graph Co-Learning},
  year       = {2022},
  Booktitle  = {4th Conference on Automated Knowledge Base Construction},
  Series = {AKBC 2022},
}