Improving Event Representation Learning via Generating and Utilizing Synthetic Data

We are pleased to release the official implementation of our paper titled "Improving Event Representation Learning via Generating and Utilizing Synthetic Data", which was submitted to the journal of Information Processing & Management.

News

• Dec 12 2024, the code, dataset and checkpoints are coming soon!
• Dec 14 2024, the code has been released, the dataset and checkpoints are coming soon!
• Jan 27 2025, the paper has been accepted by Information Processing & Management! 🎉🎉🎉

Quick Start

Installation

To run a docker container:

docker run ubuntu:22.04

To install pip requirements:

pip3 install \
  texar-pytorch \
  torch==1.13.1 \
  tensorflow==2.14 \
  numpy==1.26.4 \
  nltk \
  faiss-cpu \
  tiktoken \
  jupyter \
  matplotlib \
  openai \
  scipy \
  scikit-learn

Synthetic

python3 syn-dat.py \
  --anchor /data/train.json \
  --atomic /data/atomic/v4_atomic_all.csv \
  --output /data/out \
  --api-key API-KEY \
  --prompt analogy-reasoner.j2

Train

python3 main.py \
  --do-train \
  --output-dir /data/out

Test

python3 main.py \
  --do-eval \
  --checkpoint /data/checkpoint.pt

Acknowledgement

The code is developed based on SWCC. We appreciate all the authors who made their code public, which greatly facilitates this project.

Citation

@article{feng2025improving,
  title = {Improving event representation learning via generating and utilizing synthetic data},
  author = {Yubo Feng and Lishuang Li and Xueyang Qin and Beibei Zhang},
  journal = {Information Processing & Management},
  volume = {62},
  number = {4},
  pages = {104083},
  year = {2025},
  issn = {0306-4573},
  doi = {https://doi.org/10.1016/j.ipm.2025.104083},
  url = {https://www.sciencedirect.com/science/article/pii/S0306457325000251},
}