WeatherGovPlus

Repository containing supplementary material for the DocEng 2023 paper "WEATHERGOV+". Please see paper for details on implementation and results.

**Download dataset files from: ** http://web.uvic.ca/~mcote/WeatherGovPlus/WeatherGovPlus.zip

Requires:

• Python 3.8
• Docker
• NVIDIA CUDA Docker Toolkit

This has been tested and run on Linux Mint 20 (Ubuntu 20.04) and CUDA 11.6. To use a different version of CUDA, you will need to build the docker containers for your system.

Download pretrained recognition model files

1. Go to https://github.com/hikopensource/DAVAR-Lab-OCR.git and download to DAVAR-Lab-OCR folder:

• maskrcnn-lgpma-pub-e12-pub.pth
• text_det_model.pth
• text_rcg_model_e8.pth

2. Go to https://github.com/JiaquanYe/TableMASTER-mmocr.git and download to TableMASER-mmocr/checkpoints folder:

• 200~epoch_16_0.7767.pth
• master_epoch_6.pth
• pse_epoch_600.pth

Build and run the Docker images

0. Install Docker with NVIDIA Toolkit

1. Go to the recognition model folder:

   cd table_recognition/[Model]

2. Build the docker image:

   nvidia-docker build -t [docker name] .

3. Run the inference.py script:
   docker run --rm -it --runtime=nvidia --gpus all --mount type=bind,source=../../evaluation/WEATHERGOV_PLUS,target=/data,readonly --mount type=bind,source=../../evaluation /results,target=/results [docker name] python inference.py

   • This will bind the evaluation folder to the docker image and save the results to evaluation/results/[lgpma,davar]

An example can be found in example.sh.

Run the NLP evaluation using MVP

0. Create a virtualenv and install the packages in evaluation/requirements

1. Unzip the folders in evaluations/WEATHER_PLUS and wg/

   • Your folder structure should be:
     • WEATHER_PLUS/TablesJPG
     • WEATHER_PLUS/TablesHTML_htags

2. Run the MVP NLP algorithm:

   python run_analysis.py --html-dir evaluation/[davar,lgpma] [--run-teds] [--run-rouge] [--run-bleu]

   • --run-teds: Run the TEDS algorithm on the table recognition results (this is rather slow)
   • --run-[rouge,bleu] : Run the BLEU/ROUGE NLP metrics on the table summarization results

Citation

If you use this code or dataset for your research, please cite:

@inproceedings{dash2023weathergov+,
  title={WEATHERGOV+: A Table Recognition and Summarization Dataset to Bridge the Gap Between Document Image Analysis and Natural Language Generation},
  author={Dash, Amanda and Cote, Melissa and Albu, Alexandra Branzan},
  booktitle={Proceedings of the ACM Symposium on Document Engineering 2023},
  pages={1--10},
  year={2023}
}