Table Structure Recognition

A document table has both visual boundaries and relational structure. Table detection answers where the table is; structure recognition reconstructs its rows, columns, and cell boundaries. Functional analysis may further label headers or other roles, while text extraction supplies the words inside cells. A system that finds the outer rectangle but loses a merged header or shifts text into the wrong column has not completed useful table extraction. Microsoft’s Table Transformer repository describes an object-detection model for extracting tables from PDFs and images and links it to the PubTables-1M dataset and GriTS metric. The repository notes that its inference pipeline needs text from OCR or directly from a PDF as a separate input to include content in HTML or CSV. PubTables-1M includes annotated pages, tables, cell locations, and text. The paper also addresses oversegmentation in earlier annotations by canonicalizing table structure, which matters because inconsistent ground truth can distort evaluation. Evaluate detection, structure, and text alignment separately. Test tables with and without visible rules, spanning cells, nested or multi-level headers, and diverse document styles. Compare extracted grids with source images and verify totals and key values. A clean CSV can still encode the wrong relationships. Keep the source document and provenance, and route uncertain or high-impact tables to a person. Dataset benchmark scores are tied to their test sets and do not promise accuracy on every organization’s PDFs.