Al-Warraq at AR-MS NAKBA-NLP 2026: Adapting Vision-Language and Transformer Models for Automatic Manuscript OCR/HTR

Abstract

We present our submission to the NAKBA NLP 2026 Automatic Manuscript OCR/HTR shared task on Arabic manuscripts. The task aims to transcribe manuscript line images into machine-readable Arabic text. Our approach followed an iterative pipeline including model selection, training, error analysis, test-time augmentation, and postprocessing. After evaluating several OCR/HTR models, we selected and trained the most suitable model on the provided manuscript line images and transcriptions. Error analysis showed better character-level performance than word-level performance, which motivated the use of test-time augmentation and text cleaning to improve robustness. The final system achieved a CER of 0.1142 and a WER of 0.378, placing fifth in the shared task. These results show that simple but targeted improvements can support effective Arabic manuscript transcription.

Resources

Details