Source: TableFormer (Nassar et al., 2022) and OTSL (Lysak et al., 2023)

IBM Research papers on table structure recognition: TableFormer predicts table structure plus cell boxes with transformer decoders; OTSL is a 5-token table language that halves sequence length versus HTML.

Contents
  • What: two research papers (arXiv:2203.01017, CVPR 2022; arXiv:2305.03393, ICDAR 2023)
  • Author/publisher: Ahmed Nassar, Nikolaos Livathinos, Maksym Lysak, Peter Staar (TableFormer); Maksym Lysak, Ahmed Nassar, Nikolaos Livathinos, Christoph Auer, Peter Staar (OTSL); IBM Research
  • Date: 2022-03-02 (TableFormer); 2023-05-05 (OTSL)
  • Original: TableFormer, OTSL

Summary

Table structure recognition turns a table image into its logical grid (rows, columns, spans, headers). TableFormer improves the earlier PubTabNet encoder/dual-decoder design by using transformer decoders and adding a cell-detection decoder, so cell text can be taken straight from a programmatic PDF instead of being transcribed by a custom OCR decoder. OTSL then replaces HTML as the target sequence with a minimal table language, which shortens outputs and guarantees well-formed structures.

Key claims

  • TableFormer is evaluated with TEDS (tree-edit-distance-based similarity) and reports improvement from 91% to 98.5% on simple tables and from 88.7% to 95% on complex tables versus the prior state of the art.
  • Predicting cell bounding boxes lets the system copy exact text from the PDF layer for digital documents.
  • OTSL needs 5 tokens versus 28+ for HTML, roughly halves sequence length, halves inference time, and always yields syntactically valid tables.
  • These models are the table component of Docling.

Feeds

Document processing

Suggest changes →