vrdiu-track-a

This repository is for the IJCAI 2024 VRDIU (Visually Rich Form Document Intelligence and Understanding) challenge, sponsored by Google Research. We achieved 2nd place in Track A (Key Information Extraction) of the competition.

Environment Setting

Create conda environment (It may take some time)

conda create -n vrdiu python=3.8
conda activate vrdiu

Install packages

bash env.sh

Download Form-NLU[1] datasets

bash dataset.sh

Training

bash train.sh

Inference

Download checkpoint

gdown 1frEB4gJEBgdBLypVZXoXNAvlY-FX332A
unzip checkpoint.zip -d checkpoint

Inference

bash inference.sh

Post-processing for ensuring no overlapping keys and save it to csv file

python post_process.py
python submission.py

[1] Ding, Y., Long, S., Huang, J., Ren, K., Luo, X., Chung, H., & Han, S. C. (2023, July). Form-NLU: Dataset for the Form Natural Language Understanding. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (pp. 2807-2816).