PDFMathTranslate/PDFMathTranslate

What it solves

PDFMathTranslate solves the problem of translating scientific PDF documents while maintaining their original layout. It ensures that complex elements like mathematical formulas, charts, tables of contents, and annotations are preserved during the translation process, preventing the formatting breakage common in standard document translators.

How it works

The tool extracts text from PDFs and translates it using various supported translation services. It employs different processing modes to handle different PDF types:

• Normal Mode: Uses an ONNX-based layout model (DocLayout-YOLO) to identify document structure.
• Ultrafast Mode: Uses native PDF text positions for heuristic layout, skipping OCR and ONNX layout for faster processing of native-text PDFs.
• Experimental OCR: Automatically runs local OCR on image-only pages, regrouping words into paragraphs before translation.
• Typesetting: Translated text is dynamically scaled to fit original bounding boxes, starting at the median source font size.

Who it’s for

It is designed for researchers, students, and academics who need to translate scientific papers and technical documents without losing the structural integrity of the formulas and layouts.

Highlights

• Layout Preservation: Keeps formulas, charts, and annotations intact.
• Flexible Deployment: Available as a command-line tool, a browser-based GUI, a Docker container, and a Zotero plugin.
• Multiple Translation Backends: Supports various translation services, including Google and DeepL.
• Versatile Input: Handles both native-text PDFs and scanned documents via OCR.