|
|
 |
| |
|
|
|
A Statistical Characterization of Arabic Historical Manuscript Datasets: Resources, Annotations, Scripts, and Accessibility |
|
|
|
PP: 1613-1626 |
|
|
doi:10.18576/jsap/150607
|
|
|
|
Author(s) |
|
|
|
Amani Jamal,
|
|
|
|
Abstract |
|
|
| Benchmark datasets are the ground on which computational work on Arabic historical manuscripts stands, yet that ground has never been measured. Following PRISMA 2020, we searched Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library (471 records; 27 studies retained), added 8 studies through citation searching, and coded 35 dataset records (2010–2026) twice against their primary papers. Scale is heavily right-skewed (authentic subset: median 300 pages, mean 3,516) and has not grown over time (Spearman ρ = 0.195, p = 0.268). Fifteen records (42.9%) report a public download, but only eight endpoints delivered open files in October 2026; five (14.3%) name a dataset license and four (11.4%) have a DOI. On a 13-item reporting score, the median record documents 5 items, with no improvement since 2010 (ρ = 0.13, p = 0.44). Maghrebi hands are the most frequently named script (7 records) yet supply about 1% of authentic pages. Only two records test exclusively on manuscripts unseen during training; where one study reports both protocols, writer-identification accuracy falls from 98.1% to 60.1% (WrIHAM) and from 94.95% to 81% (WAHD) once test manuscripts are kept out of training. We propose a minimum reporting checklist and recommend manuscript-disjoint splits, DOI-backed versioned releases with explicit licenses, and larger Maghrebi collections.
|
|
|
|
|
 |
|
|