3D Topological Positioning in Generative Table Recognition

This repository contains the implementation and evaluation code for our paper "Evaluating Three-Dimensional Topological Positioning in Generative Table Recognition". It enables researchers to replicate our experiments and results.

⚙️ Requirements

• Linux environment
• 4+ GPUs with 32GB+ VRAM (for training)
• 1 GPU with 32GB+ VRAM (for evaluation)
• Python 3.11+

🛠️ Setup Instructions

1. Download and Preprocess Data

Run the following scripts to download and preprocess the data:

./get_web_data.sh
./preprocess_data.sh

The preprocessing script will automatically train tokenizers and set up the required directory structure.

🚀 Training Models

Navigate to the scripts directory and run the appropriate training script:

cd proj/scripts
# Choose one or more of the following based on your model of interest
python3 Train_Model-3D_TML.py      # 3D Emb. TML model
python3 Train_Model-Pos_HTML.py    # Pos. Enc. HTML model
python3 Train_Model-3D_HTML.py     # 3D Emb. HTML model
python3 Train_Model-Pos_TML.py     # Pos. Enc. TML model

Note: All training scripts require at least 4 GPUs with 32GB+ VRAM.

📊 Evaluation

1. Generate Model Outputs

Run the corresponding evaluation script:

cd proj/scripts
# Choose the appropriate evaluation scripts
python3 Evaluate_Model-3D_TML.py      # 3D Emb. TML model
python3 Evaluate_Model-Pos_HTML.py    # Pos. Enc. HTML model
python3 Evaluate_Model-3D_HTML.py     # 3D Emb. HTML model
python3 Evaluate_Model-Pos_TML.py     # Pos. Enc. TML model

Note: Evaluation requires at least 1 GPU with 32GB+ VRAM.

Evaluation outputs are saved in JSON format in the aux/outputs/ directory, organized by model and number of training steps:

aux/outputs/
├── [model_name]/
│   ├── [model]-[step]-output.json 
...

2. TEDS Evaluation

To compute Tree Edit Distance Scores, clone PubTabNet's, for the TEDS script, repository and run the evaluation:

cd aux
git clone https://github.com/ibm-aur-nlp/PubTabNet
python3 Eval_Outputs.py

Note: The Eval_Outputs.py script will automatically produce the scores for all files the generated in the previous step.

📂 Results

Evaluation scores are saved in JSON format under:

aux/outputs/
├── [model_name]/
│   └── evals/
│       ├── [model]-[step]-output-all.json       # Complete evaluation
│       └── [model]-[step]-output-struct.json    # Structure-only evaluation
│       ...
...

Example structure:

aux/outputs/3D_TML/evals/
├── model_3D_TML-STEP_12000-output-all.json
├── model_3D_TML-STEP_12000-output-struct.json
├── ...
└── model_3D_TML-3_EPOCHS-output-struct.json

How to cite

@article{Lima:25,
  author={Lima, João Paulo Paiva and Pereira, Denilson Alves},
  title={Evaluating Three-Dimensional Topological Positioning in Generative Table Recognition},
  volume={7},
  DOI={10.33969/ais},
  number={1},
  journal={Journal of Artificial Intelligence and Systems},
  year={2025},
  month={Oct},
  pages={61–75}
}