Data rescue of historical tables through semi-supervised table structure recognition.

Area of Science:

• Computer Science
• Digital Humanities
• Information Retrieval

Background:

• Tabular Structure Recognition (TSR) is vital for digitizing historical documents.
• Challenges include document degradation and inconsistent handwriting, complicating annotation for model training.
• Existing methods often require extensive labeled data, which is costly and time-consuming to produce for historical archives.

Purpose of the Study:

• To investigate if semi-supervised learning can reduce the need for expensive data annotations in TSR.
• To determine if semi-supervised training enhances model robustness for historical document analysis.
• To improve the accessibility and analysis of historical tabular data through efficient digitization.

Main Methods:

• A novel semi-supervised learning framework was developed and applied.
• The CascadeTabNet model was employed for Tabular Structure Recognition.
• Methodology was tested on historical (GloSAT, ICDAR-2019) and modern (PubTabNet) document datasets.

Main Results:

• Semi-supervised learning significantly increased TSR accuracy across datasets.
• The approach reduced dependency on large amounts of labeled data.
• Improved model robustness was observed, particularly for historical documents.

Conclusions:

• Semi-supervised learning offers a robust and efficient solution for large-scale digitization of historical documents.
• This framework enhances the preservation and accessibility of valuable historical data.
• The study demonstrates the effectiveness of semi-supervised approaches in overcoming annotation challenges in archival research.