Dataset Card for MUSTARD

Dataset Details

Dataset Description

MUSTARD (Multilingual Scanned and Scene Table Structure Recognition Dataset) is a diverse dataset curated for table structure recognition across multiple languages. The dataset consists of tables extracted from magazines, including printed, scanned, and scene-text tables, labeled with Optimized Table Structure Language (OTSL) sequences. It is designed to facilitate research in multilingual table structure recognition, particularly for non-English documents.

• Curated by: IIT Bombay LEAP OCR Team
• Funded by: IRCC, IIT Bombay, and MEITY, Government of India
• Shared by: IIT Bombay LEAP OCR Team
• Language(s) (NLP): Hindi, Telugu, English, Urdu, Oriya, Malayalam, Assamese, Bengali, Gujarati, Kannada, Punjabi, Tamil, Chinese
• License: MIT

Dataset Sources

• Repository: GitHub Repository
• Paper: SPRINT: Script-agnostic Structure Recognition in Tables (ICDAR 2024)
• Dataset Download: Hugging Face Link

Uses

Direct Use

MUSTARD is primarily intended for training and evaluating table structure recognition models, especially those dealing with multilingual and script-agnostic document analysis.

Out-of-Scope Use

The dataset should not be used for tasks unrelated to table structure recognition. Additionally, any application involving sensitive data extraction should ensure compliance with relevant legal and ethical guidelines.

Dataset Structure

The dataset consists of:

• 1428 tables across 13 languages.
• Labels provided in OTSL format.
• A mixture of printed, scanned, and scene-text tables.

Dataset Creation

Curation Rationale

The dataset was created to address the lack of multilingual table structure recognition resources, enabling research beyond English-centric datasets.

Source Data

Data Collection and Processing

• Tables were sourced from various magazines.
• Labeled using OTSL sequences to provide a script-agnostic representation.
• Ground truth annotations were validated for accuracy.

Who are the source data producers?

The dataset was curated by researchers at IIT Bombay, specializing in OCR and document analysis.

Annotations

Annotation Process

• Tables were manually labeled using OTSL sequences.
• Verification was performed to ensure consistency.
• Annotations were aligned with HTML-based table representations for interoperability.

Who are the annotators?

Annotations were performed by research scholars and experts in OCR and document processing at IIT Bombay.

Personal and Sensitive Information

The dataset does not contain personally identifiable or sensitive information.

Bias, Risks, and Limitations

• Bias: The dataset is derived primarily from magazines, which may not fully represent all document styles.
• Limitations: The dataset size is limited (1428 tables), and performance may vary on unseen data sources.
• Risks: Use in sensitive domains should be accompanied by proper validation and legal compliance.

Recommendations

Users should be aware of dataset limitations and biases when applying models trained on MUSTARD to other real-world scenarios.

Citation

If you use this dataset in your research, please cite it as:

@InProceedings{10.1007/978-3-031-70549-6_21,
author="Kudale, Dhruv and Kasuba, Badri Vishal and Subramanian, Venkatapathy and Chaudhuri, Parag and Ramakrishnan, Ganesh",
editor="Barney Smith, Elisa H. and Liwicki, Marcus and Peng, Liangrui",
title="SPRINT: Script-agnostic Structure Recognition in Tables",
booktitle="Document Analysis and Recognition - ICDAR 2024",
year="2024",
publisher="Springer Nature Switzerland",
address="Cham",
pages="350--367",
isbn="978-3-031-70549-6",
url = "https://arxiv.org/abs/2503.11932"
}

More Information

For further details, refer to:

• SPRINT Model: GitHub Repository
• Pretrained Models: Model Releases
• Dataset Download: Hugging Face Dataset

Dataset Card Authors

• Badri Vishal Kasuba
• Dhruv Kudale

Dataset Card Contact

For queries, contact the authors via their respective institutional affiliations.

License

The dataset is licensed under the MIT License, allowing for free use and modification with proper attribution.