Handwriting as Signal: Toward Scribe-based Join Retrieval in the Cairo Genizah

The Cairo Genizah manuscripts are spread across more than seventy libraries globally, often with fragments of a single original held in different institutions. Recovering fragments from the same work, known as join retrieval, remains a core challenge: manual reconstruction is slow, expert-dependent, and hence yields limited labeled data. Handwriting style, unlike background appearance, degradation and scanning conditions, is specific to the manuscript’s scribe and serves as our primary signal for an automated computational join-suggestion approach.

We fine-tune a transformer-based vision model using contrastive self-supervised learning on character-sequence crops without join labels, introducing corpus-aware augmentations: parchment texture replacement, ink bleed-through simulation, and background substitution, to suppress non-scribal visual variation.

We evaluate the model's learned representations on 234 fragments from 86 known-join clusters across ten libraries. Preliminary results demonstrate improved fragment-level retrieval, whereas cross-library join retrieval remains a key focus of our ongoing efforts to make Genizah reconstruction more accessible to scholars.