ENTRANT: A Large Financial Dataset for Table Understanding.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Tabular data is crucial for organizing and comparing information in a two-dimensional matrix format.
• Training large models to understand structured tables is essential for knowledge transfer in various applications.
• Effective pre-training of these models necessitates large, well-formatted datasets that capture table and cell characteristics.

Purpose of the Study:

• To introduce ENTRANT, a novel financial dataset specifically curated for pre-training deep learning models on table understanding.
• To provide a machine-readable dataset with detailed table and cell information, including metadata, attributes, and hierarchical structure.
• To facilitate automated data processing and validation for robust dataset utility.

Main Methods:

• The ENTRANT dataset was created by transforming millions of financial tables.
• Data processing and curation were fully automated.
• Technical validation was performed using unit testing with high code coverage.

Main Results:

• The ENTRANT dataset contains millions of transformed tables with cell attributes, positional, and hierarchical information.
• The dataset is provided in a machine-readable format, including metadata.
• Automated processing and validation ensure data quality and usability.

Conclusions:

• ENTRANT is a valuable resource for advancing deep learning-based table understanding.
• The dataset effectively supports pre-training tasks for models designed to interpret structured tabular data.
• Demonstrated use in a pre-training task for cell classification highlights its practical applicability.