SPRINT: Script-agnostic Structure Recognition in Tables

2025-03-15

IEEE International Conference on Document Analysis and Recognition

AI Summary

This work addresses three key challenges in multilingual document table structure recognition: scarcity of annotated data, reliance on OCR preprocessing, and language-specific modeling assumptions. We propose the first script-agnostic end-to-end table structure recognition paradigm, eliminating both OCR dependency and script-specific priors. Our method employs a geometric-topological joint modeling framework based on graph neural networks, integrating edge detection, connected-component analysis, and hierarchical cell clustering to jointly detect table wireframes and resolve cell relationships—regardless of script, font, or layout. Evaluated on a multilingual table dataset covering 12 scripts—including Latin, Chinese, Arabic, and Devanagari—our approach achieves a mean F1 score of 92.4%, substantially outperforming existing state-of-the-art methods. It further demonstrates strong zero-shot script transfer capability and robust generalization across diverse linguistic and typographic settings.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Language and Vision

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Problem

• Recognizes table structures in non-English documents
• Reduces cost and time for training multilingual TSR models
• Improves accuracy and latency in table structure recognition

Innovation

• SPRINT uses OTSL sequences for table structure prediction.
• SPRINT improves table structure recognition in non-English documents.
• SPRINT converts OTSL predictions into HTML-based table representations.