ByT5 Reduces Uzbek OCR Errors Using Homoglyph Noise Matrix
November 11, 2026
A byte-level ByT5 model with a custom noise generator lowers character error rates from 0.0578 to 0.0369 for mixed-script Uzbek text. The method uses sampling weights from 359 manual error classifications and Levenshtein-bounded decoding to fix visual character confusion between Cyrillic and Latin scripts.
HOW THIS AFFECTS YOU
●
builderYou can improve text extraction quality for Uzbek documents by implementing byte-level models rather than standard token-based ones.
●
researcherYou can use this homoglyph noise sampling approach to improve OCR correction in other mixed-script languages.