RConE: Rough Cone Embedding for Multi-Hop Logical Query Answering on Multi-Modal Knowledge Graphs

The repository consists of code for our RConE paper.

Below is the description of each model. Readme is provided in folders to run the code.

Folders

fcsgg

The folder consists of Scene Graph Generation module for our model.

baselines

Contains preprocessing file for datasets to be compatible for both the baselines (ConE and BetaE).

dataset

Will contain dataset files (see installation step - Prepare dataset).

model

RConE model

Installation

Prepare dataset

• Download the dataset folder from link - https://drive.google.com/file/d/1UknhjOn86U0SsPwfmHZHt2XwOb-GG66H/view?usp=sharing
• unzip the folder and extract as the dataset folder (replace it with current empty dataset folder).

Prepare RConE module

Move to model folder, and follow the instructions in its README.md file to prepare our model.

Prepare baselines

Move to baselines folders

Download the models from

• ConE - https://github.com/MIRALab-USTC/QE-ConE/tree/main
• BetaE - https://github.com/snap-stanford/KGReasoning
• run preprocess.py for each dataset based on the updated location in the file (preprocess.py)

Code Execution

After following all the installation steps, follow following commands for query generation and to run the model.

Query Generation

Scene Graph Generation

Generate scene graphs for fb-15k dataset by following steps in the fcsgg folder and copy the results in the model/results folder with architecture config 32 or 48, as currently present

ComplEx

Move to model/complex folder. Uncomment the desired dataset command in the following script files and then run.

sh script_preprocess.sh
sh script_main.sh
sh script_post.sh 

Create Queries

Move to model folder. Run command

sh query_scriptfb.sh

for fb15k dataset (similar scripts are there for other datasets).

RConE

Scene Graph Generation

Generate scene graphs for fb-15k dataset by following steps in the fcsgg folder and copy the results in the model/results folder with architecture config 32 or 48, as currently present

Preprocess

Move to model folder. Run command

python process.py
python transform_preprocess.py
cd complex
sh script_inrun.sh

Training

Move to model folder. Uncomment the dataset command you want to train the model onin the script file and than execute.

scripts.sh

The trained model will be in log folder with desired dataset name

Testing

After model is trained (in the model folder). uncomment the dataset command you want to train the model on in the script file and than execute.

scripts_test.sh

The results will be in log folder with desired dataset name under file test.log

Acknowledgement

• https://github.com/MIRALab-USTC/QE-ConE/tree/main
• https://github.com/snap-stanford/KGReasoning
• https://github.com/liuhengyue/fcsgg
• https://github.com/ttrouill/complex