Dataset Card for SynthTabNet_OTSL

Dataset Summary

This dataset is a conversion of the original SynthTabNet into the OTSL format presented in our paper "Optimized Table Tokenization for Table Structure Recognition". The dataset includes the original annotations amongst new additions.

SynthTabNet is organized into 4 parts of 150k tables (600k in total). Each part contains tables with different appearances in regard to their size, structure, style and content. All parts are divided into Train, Test and Val splits.

Appearance style	Records
Fintabnet	150k
Marketing	150k
PubTabNet	150k
Sparse	150k

Dataset Structure

• cells: origunal dataset cell groundtruth (content).
• otsl: new reduced table structure token format
• html: original dataset groundtruth HTML (structure).
• html_restored: generated HTML from OTSL.
• cols: grid column length.
• rows: grid row length.
• image: PIL image

OTSL Vocabulary:

OTSL: new reduced table structure token format More information on the OTSL table structure format and its concepts can be read from our paper. Format of this dataset extends work presented in a paper, and introduces slight modifications:

• "fcel" - cell that has content in it
• "ecel" - cell that is empty
• "lcel" - left-looking cell (to handle horizontally merged cells)
• "ucel" - up-looking cell (to handle vertically merged cells)
• "xcel" - 2d span cells, in this dataset - covers entire area of a merged cell
• "nl" - new line token

Data Splits

The dataset provides three splits

• train
• val
• test

Additional Information

Dataset Curators

The dataset is converted by the Deep Search team at IBM Research. You can contact us at deepsearch-core@zurich.ibm.com.

Curators:

• Maksym Lysak, @maxmnemonic
• Ahmed Nassar, @nassarofficial
• Christoph Auer, @cau-git
• Nikos Livathinos, @nikos-livathinos
• Peter Staar, @PeterStaar-IBM

Citation Information

@misc{lysak2023optimized,
      title={Optimized Table Tokenization for Table Structure Recognition}, 
      author={Maksym Lysak and Ahmed Nassar and Nikolaos Livathinos and Christoph Auer and Peter Staar},
      year={2023},
      eprint={2305.03393},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}```