Arabic OCR with human review where it matters

Digitising Arabic print means handling ligatures, diacritics and degraded scans. One OCR engine is rarely good enough, and reviewing every page by hand is too slow.

I designed a multi-engine pipeline. Engines run in parallel. A language model scores perplexity on each output. Morphological validation with CAMeL Tools and Farasa flags impossible words. Acceptance-quality-level rules route only suspect pages to human reviewers in tiers.

A review process that spends human time where errors are most likely. This pipeline underpins large digitisation bids.