conjuncts/gmft

What it solves

It solves the difficulty of reliably extracting tables from PDF documents, which often lack a definitive extraction method. It specifically targets the challenge of maintaining table structure and alignment when converting PDF data into usable digital formats.

How it works

The tool uses Microsoft's Table Transformer (TATR) model, pretrained on the PubTables-1M dataset, to detect and recognize table structures. To maximize speed, it leverages existing text positional data within PDFs rather than relying on OCR by default. It uses PyPDFium2 for high-throughput document processing and can export results into various formats, including Pandas dataframes, Markdown, LaTeX, HTML, CSV, and JSON.

Who it’s for

It is designed for researchers, data scientists, and developers who need to extract tabular data from scientific papers or PDFs and require a lightweight, high-performance solution that can run on a CPU without needing a GPU.

Highlights

• High Performance: Approximately 10x faster on CPU than several other popular extraction alternatives.
• Lightweight: No GPU required and minimal external dependencies (no need for Tesseract or Poppler).
• Versatile Export: Supports multiple output formats and can provide cropped images of tables for use with vision-based LLMs like GPT-4 Vision.
• Advanced Structure Support: Handles implicit table structures, multiple-column headers, spanning cells, and rotated tables.