Robust Retrieval Across Orthographic and Editorial Variants

Develop retrieval methods that remain effective under orthographic and editorial variation between manuscript witnesses and source texts, thereby improving text matching and source attribution in Arabic manuscript transcription.

Background

The retrieval-augmented correction ablation evaluated queries against transcriptions aligned to the same manuscripts and achieved substantial recall improvements when using multiple visual candidates. However, the authors caution that these results do not establish performance when source witnesses differ in orthography or editorial treatment. Reliable retrieval across such variation is important because Athar uses retrieved parallels as evidence for correction and attribution, while explicitly distinguishing ambiguous matches from unique attributions.