dice-group / TemporalFC

This open-source project contains the Python implementation of our approach TemporalFC. This project is designed to ease real-world applications of fact-checking over knowledge graphs and produce better results.

TemporalFC: A Temporal Fact-Checking Approach for Knowledge Graphs

Time predication component

This open-source project contains the Python implementation of time predication component our approach TemporalFC. This project is designed to ease real-world applications of fact-checking over knowledge graphs and produce better results. With this aim, we rely on:

1. PytorchLightning to perform training via multi-CPUs, GPUs, TPUs or computing cluster,
2. Pre-trained-TKG-embeddings to get pre-trained TKG embeddings for knowledge graphs for knowledge graph-based component,
3. Elastic-search to load text corpus (wikipedia) on elastic search for text-based component, and
4. Path-based-approach to calculate output score for the path-based component.

This project consists of 2 independent components:

1. Fact-checking component.
2. Time-point prediction component.

Installation

First clone the repository:

git clone https://github.com/factcheckerr/TemporalFC.git

cd TemporalFC

Reproducing Results

There are two options to repreoduce the results. (1) using pre-generated data, and (2) Regenerate data from scratch. Please chose any 1 of the these 2 options.

1) Re-Using pre-generated data

download and unzip data and embeddings files in the root folder of the project.

pip install gdown

wget https://zenodo.org/records/8197242/files/data_TP.zip
wget https://zenodo.org/records/7552968/files/dataset.zip
wget https://zenodo.org/records/7552968/files/Embeddings.zip

unzip data.zip

Note: if it gives permission denied error you can try running the commands with "sudo"

2) Generating data from scratch

To regenerate data from scratch, you need to run embedding algorithm again and put the generated SPO embedding in data/dataset_name/temporal_embeddings folder, and dataset in data/dataset_name/train and data/dataset_name/test foder.

Running experiments

Install dependencies via conda:

#setting up environment
#creating and activating conda environment

conda env create -f environment.yml

conda activate hfc2

#If conda command not found: download miniconda from (https://docs.conda.io/en/latest/miniconda.html#linux-installers) and set the path: 
#export PATH=/path-to-conda/miniconda3/bin:$PATH

start generating results:

# Start training process, with required number of hyperparemeters. Details about other hyperparameters is in main.py file.
python run.py             --dataset DBPedia5             --model sFourDETim             --rank 100             --regularizer N3             --reg 0.00000000001             --optimizer Adagrad             --max_epochs 500             --patience 15             --valid 10             --batch_size 100             --neg_sample_size -1             --init_size 0.001             --learning_rate 0.1             --gamma 0.0             --bias learn             --dtype single             --double_neg             --cuda_n 2             --dataset_type quintuple_Tim
# computing evaluation files from saved model in "dataset/Hybrid_Stroage" directory
python evaluate_checkpoint_model.py        --dataset DBPedia5             --model sFourDETim             --rank 100             --regularizer N3             --reg 0.00000000001             --optimizer Adagrad             --max_epochs 500             --patience 15             --valid 10             --batch_size 100             --neg_sample_size -1             --init_size 0.001             --learning_rate 0.1             --gamma 0.0             --bias learn             --dtype single             --double_neg             --cuda_n 2             --dataset_type quintuple_Tim

comments:

1. To reproduc exact results you have to use exact parameters as listed above.

2. For other datasets you need to change the parameter in front of --dataset

3. Use GPU for fast processing. Default parameter is set to 2 GPUs that we used to generate results.

4. For different embeddings type(emb_type) or model type(model), you just need to change the parameters.

Available embeddings types: dihedron, T-TransE, T-ComplEx.

Available models: temporal-model

Note: model names are case-sensitive. So please use exact names.

Fact checking part:

Fact checking part of TemporalFC is available in a second project: TemporalFC-FactCheck.

Future plan:

As future work, we will exploit the modularity of TemporalFC by integrating time-period based fact checking.

Acknowledgement

The work has been supported by the EU H2020 Marie Skłodowska-Curie project KnowGraphs (no. 860801)).

Authors

• Umair Qudus (DICE, Paderborn University)
• Michael Röder (DICE, Paderborn University)
• Sabrina Kirrane (WU, Vienna University)
• Axel-Cyrille Ngonga Ngomo (DICE, Paderborn University)