RecLink

Revisiting Document-Level Relation Extraction with Context-Guided Knowledge Graph Link Prediction

Acknowledgments

This repository contains code adapted from the following research papers for the purpose of document-level relation extraction. We extend our gratitude to the authors for generously sharing their clean and valuable code implementations.

1. Discriminative Reasoning for Document-level Relation Extraction

• Paper: Discriminative Reasoning for Document-level Relation Extraction
• Authors: Wang Xu, Kehai Chen, Tiejun Zhao
• Year: 2021

Code implementation can be found here: GitHub - xwjim/DRN

2. Modeling Relational Data with Graph Convolutional Networks

• Paper: Modeling Relational Data with Graph Convolutional Networks
• Authors: Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, Max Welling
• Year: 2017

Code implementation can be found here: GitHub - JinheonBaek/RGCN

3. Multi-Hop Knowledge Graph Reasoning with Reward Shaping

• Paper: Multi-Hop Knowledge Graph Reasoning with Reward Shaping
• Authors: Xi Victoria Lin, Richard Socher, Caiming Xiong
• Year: 2018

Code implementation can be found here: GitHub - kingsaint/InductiveExplainableLinkPrediction

Environment

This code is lastly tested with:

• Python 3.7.x
• PyTorch 1.7.x
• CUDA (11.0)
• torch_geometric 1.7.x, with torch_scatter 2.0.6 and torch_sparse 0.6.9

Libraries

• numpy (1.19.1)
• matplotlib (3.3.1)
• torch (1.7.1)
• transformers (4.1.1)
• scikit-learn (0.23.2)
• wandb (0.10.12)
• tqdm (4.9.0)

Directory structure

In this directory structure, you have a folder named "Reclink" containing a subdirectory "code." Within the "code" directory, there are several files and subdirectories:

• link prediction: Files for link prediction module
• `reasoning_path``: files for generating explanations.
• Context: Files for creating context.
• checkpoint/: Directory to store model checkpoints.
• logs/: Directory to store logs related to the code.
• models/: Directory to store trained models.
• config.py: Configuration file for the code.
• data.py: File containing code related to data processing.
• test.py: File for testing the code.
• train.py: File for training the code.
• utils.py: Utility functions used in the code.

Datasets

This project utilizes the following datasets:

• DocRED Dataset: The DocRED dataset can be accessed here place in data/docred directory

• Re-DocRED Dataset: The Re-DocRED dataset is available here place in data/REDocred directory

• DWIE Dataset: The DWIE dataset's repository can be found here place in data/DWIE directory

  To convert DWIE dataset into docred style: follow the guideline from this code: https://github.com/rudongyu/LogiRE

  Please make sure to review the terms and conditions of each dataset before use.

Training

Follow the steps below to start the training process:

1. Train reasoning module: Navigate to the code directory using the following command:

   cd code ./runBERT.sh gpu_id

2. Train link prediction module: python3 RGCN/main.py

Testing

Navigate to the code directory using the following command:

cd code ./evalBERT.sh gpu_id

Explanation

Navigate to the reasoning_path folder And download the respective dataset models from the link: https://drive.google.com/drive/folders/1j0ArOF9mJDvsYCgLmr022VRb51YxJ9np Get the explanation using: For DocRED dataset: ./experiment-rs.sh configs/DOCRED-rs.sh --inference --save_beam_search_paths For DWIE dataset: ./experiment-rs.sh configs/DWIE-rs.sh --inference --save_beam_search_paths

Context creation

Navigate to context directory: Select appropriate datasets and files:

1. Create REBEL triples using REBEL.py

2. Create triples of entity context using entity_context.py (give entity file here)

3. Create triples of context path using context_path.py (provide both entities to calculate path)

4. You can check these files align with dataset using similarity.py and remove noise.

Citation

Please cite:

@article{Jain_Mutharaju_Kavuluru_Singh_2024,
    title={Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction},
    volume={38},
    url={https://ojs.aaai.org/index.php/AAAI/article/view/29792},
    DOI={10.1609/aaai.v38i16.29792},
    number={16},
    journal={Proceedings of the AAAI Conference on Artificial Intelligence},
    author={Jain, Monika and Mutharaju, Raghava and Kavuluru, Ramakanth and Singh, Kuldeep},
    year={2024},
    month={Mar.},
    pages={18327-18335}
}