arabic-extract 0.1.0

Clean Arabic text extraction from PDFs and scanned images — OCR + visual-order repair in one pipeline

arabic-extract

Clean Arabic text extraction from PDFs and scanned images — one call, clean output.

Combines PDF text extraction, image OCR, and arabic-repair into a single pipeline. Handles the visual-order problem that breaks standard Arabic NLP pipelines.

The problem it solves

Arabic PDFs and scanned documents store text in visual order with presentation-form characters. Standard tools (NFKC, CAMeL Tools) remove the presentation forms but cannot restore the reversed word order — retrieval recall stays broken at ~27%. arabic-extract applies arabic-repair automatically, restoring both letter forms and word order before the text reaches your NLP pipeline.

Install

pip install arabic-extract[pdf]          # PDF text-layer extraction
pip install arabic-extract[tesseract]    # + image OCR via Tesseract (needs binary)
pip install arabic-extract[easyocr]      # + image OCR via EasyOCR (pure Python, ~200 MB)
pip install arabic-extract[pymupdf]      # + scanned PDF rendering via PyMuPDF
pip install arabic-extract[all]          # everything

Tesseract binary (for the tesseract extra):

• Windows: download from https://github.com/UB-Mannheim/tesseract/wiki — install the Arabic language pack
• Linux: sudo apt install tesseract-ocr tesseract-ocr-ara
• macOS: brew install tesseract && brew install tesseract-lang

Quick start

import arabic_extract as aocr

# PDF — auto-detects text layer vs scanned, repairs each page
result = aocr.extract("document.pdf")
print(result.text)           # clean logical Arabic, all pages joined
print(result.pages)          # per-page breakdown
print(result.contamination)  # how many words needed repair

# Scanned image
result = aocr.extract("scan.jpg")
print(result.text)

# Explicit PDF extraction
result = aocr.extract_pdf("document.pdf", engine="tesseract")

# Explicit image extraction
result = aocr.extract_image("scan.png", engine="easyocr")

# Chain into CAMeL Tools (normalize=True is the default)
result = aocr.extract("document.pdf", normalize=True)

How it works

Input PDF or image
    │
    ├─ PDF with text layer  → pdfplumber extracts text (visual order)
    │                                     ↓
    ├─ Scanned PDF          → render page as image → OCR engine
    │                                     ↓
    └─ Image file           → OCR engine (Tesseract or EasyOCR)
                                          ↓
                               arabic-repair (de-shape + restore order)
                                          ↓
                               NFKC / CAMeL Tools normalization
                                          ↓
                               Clean logical Arabic text

A single PDF can have mixed pages — some with a text layer, some scanned. Each page is handled correctly.

Per-page results

result = aocr.extract("document.pdf")

for page in result.pages:
    print(f"Page {page.page_number} [{page.method}]: {page.text[:80]}")
    # method: "text_layer" | "ocr" | "text_layer_empty"

Ecosystem

Package	Role
arabic-rt	Core shaping / fix / unfix engine
arabic-repair	Detect and repair visual-order contamination
arabic-extract	Full PDF + image extraction pipeline
arabic-benchmark	Benchmark proving the reordering gap

License

MPL-2.0