Right-To-Left Text Recognition

Right-to-left (RTL) text recognition is OCR and document understanding adapted for scripts that read right-to-left — principally Arabic and Hebrew, along with several less widely digitized scripts — where the directional assumption baked into most OCR tooling (built first and most thoroughly for left-to-right Latin script) needs explicit correction at multiple pipeline stages, not just a flag flipped at the end. Getting RTL wrong doesn't just misorder words; for Arabic specifically, it can misrender them entirely, because Arabic letterforms change shape depending on their position within a word (initial, medial, final, or isolated) and depending on which adjacent letters they connect to — a cursive-by-design script where character segmentation itself requires script-aware handling that Latin-tuned recognizers simply don't have.

The recognition and layout challenges compound at each stage. Detection and segmentation must respect the script's connected letterforms rather than assuming character-by-character separability the way Latin block letters allow. Recognition models need training specifically on Arabic or Hebrew script rather than adaptation of a Latin-trained model, since the visual vocabulary of connected, contextual letterforms is different enough that transfer learning alone rarely closes the gap to purpose-trained accuracy. Layout and reading-order detection must flow right-to-left at the line level while often still flowing top-to-bottom at the paragraph level, and — critically — must handle bidirectional text correctly: an Arabic or Hebrew document containing embedded Latin script (a company name, a number, an English phrase) requires the recognizer and the layout engine to switch directional context mid-line without breaking either script's internal ordering, a genuinely tricky rendering and recognition problem that shows up constantly in real bilingual documents like ID cards, contracts, and government forms across the Middle East and beyond.

Document AI serving Middle Eastern, and broader Arabic- and Hebrew-speaking markets treats RTL capability as a first-class requirement rather than an afterthought, because the alternative — bolting RTL support onto a Latin-first pipeline — reliably produces the subtle errors this glossary's multilingual entries warn about: text that displays plausibly but reads scrambled, or fields extracted with reversed digit sequences (numerals within RTL text are conventionally written left-to-right even mid-RTL-flow, a further wrinkle recognition and normalization must both get right). As with all script-specific capability, the honest evaluation is per-script accuracy on real documents, not an assumption that "multilingual support" implies RTL was tested with the same rigor as the scripts that came first.