📄 Doaa Sulaiman at AR-MS NakbaNLP 2026: Faithful Diplomatic Transcription of Arabic Manuscripts Using a Human-Centred Annotation Framework

Summary

Automatic recognition of historical Arabic handwriting still trails Latin-script results. Two obstacles are a shortage of carefully annotated reference data and the absence of shared transcription conventions. This system paper describes one entry in Subtask 1 (Human Transcription Track) of the NAKBA-NLP 2026 shared task. The task material comes from the Omar Al-Saleh Memoir Collection, handwritten memoirs recording Palestinian Nakba narratives. The author manually transcribed a batch of 500 cropped line images using a strictly diplomatic approach. Spelling, punctuation, diacritics, numeral systems and layout cues were kept exactly as in the source, with no normalisation or correction. The organisers supplied no guidelines, so the author built a 26-convention framework in three layers. It was paired with a four-step self-review procedure meant to offset having only one annotator. The submission covered every line and scored CER 0.02, WER 0.01 and accuracy 0.98. The work offers a documented, reusable convention set for producing gold-standard data to train and evaluate Arabic HTR. Inter-annotator agreement and transfer to other collections remain untested.

Key points

• Data: The Omar Al-Saleh Memoir Collection contains 16 documents from 1951–1965, about 6,395 pages and about 1.6M words. Each team received roughly 500 PNG line images, full-page context images and an annotations.csv template with fields for each entry's filename and text, the source_image it came from, and its page, line and year.
• Framework layers:
  • 11 editorial-structural symbols, such as [...] for truncation, (?) for an illegible word, * for an unclear line and {...} for a transcriber insertion.
  • 12 faithful-copying rules, covering hamzah errors, tāʾ marbūṭa/hāʾ, Eastern vs. Western numerals and tanwīn, among others.
  • 3 documentation labels: Decision, Interpretive Note and Note.
• Quality control: Transcription took about three days. The four steps were periodic batch review, cross-document consistency checks, anomaly detection (duplicate IDs, empty fields, odd symbol combinations, outlier line lengths) and logging of every ambiguous reading.
• Results: Coverage was 100% (500/500 lines), with CER 0.02, WER 0.01 and accuracy 0.98. The paper cites automated HTR systems for context, such as HATFormer (8.6% CER) and Invizo (7.91% CER on handwritten text).
• Limitations and recommendations: There was a single annotator, so no inter-annotator agreement was computed, and the framework is tailored to one collection. The author recommends that future tasks provide full-page context, explicit scoring criteria for guidelines and some multi-annotator data.