Humanist Discussion Group, Vol. 40, No. 118.
        Department of Digital Humanities, University of Cologne
                      Hosted by DH-Cologne
                       www.dhhumanist.org
                Submit to: humanist@dhhumanist.org




        Date: 2026-08-07 14:35:22+00:00
        From: VLADIMIR BESKOROVAINYI <admin@besk.tech>
        Subject: Fond 555 (Tsiolkovsky papers): open catalogue, and measuring HTR accuracy without ground truth

Dear colleagues,

I would like to report an open dataset and a measurement that may be of use
beyond the archive it came from.

The personal archive of Konstantin Tsiolkovsky (1857–1935) is held as fond
555 of the Archive of the Russian Academy of Sciences. It has been scanned 
and published — 2,019 files, 51,008 images — but without a queryable 
catalogue, full-text search or dataset. I have built and released, under CC0:

- a machine-readable catalogue of all 2,019 files, resolved to their real
  archival   numbers rather than the portal's identifiers, which do not coincide;
- a dating for 1,969 files, collected from the archive's own cards, preserving 
  the  archivists' flag for conjectural dates;
- a page-level classification of all 51,008 scans;
- a corpus of machine transcriptions, currently 359 files and 7,932 scans,
  with uncertainty marked and released only for files transcribed in full.

https://github.com/beskvladimir-create/tsiolkovsky-papers
https://beskvladimir-create.github.io/tsiolkovsky-papers/

The measurement may be the more transferable part. Evaluating handwritten
text recognition in application is hampered by the absence of ground truth, and
existing ground-truth-free approaches rely on intrinsic proxies for
plausibility, which a fluent invention satisfies. Personal archives of the 
typewriter era, however, frequently preserve one text twice: the author's 
manuscript and a typed copy made from it, filed together. Transcribing both 
isolates the reading error, since the source text and the pipeline are identical 
and only the difficulty of the page differs.

Across 294 such pairs from 27 files, two readings of one handwritten page
agree on a median 37% of words, with a longest verbatim run of 12 words. 
On two files that also have a published edition the estimate can be checked 
against ground truth: it is unbiased to within a percentage point and ranks
pages as the truth does, at a rank correlation of 0.92 where the edition is a 
faithful witness.

Two negative results are reported alongside. The page classifier, validated
on 19 hand-labelled pages and correct on all of them, proves 80% accurate 
when scored against an independent signal over 4,675 sheets. And a 
text-critical comparison of two variants of one work cannot be made at this 
recognition quality: they share 19% of their words, below the rate at which 
two readings of a single page agree. The tool now refuses to report 
differences that fail to clear the measured floor.

A preprint describing the method is at arXiv:2608.03617. I would be glad of
comment, particularly on how this sits with established textual-critical
practice, and on whether the duplicate-holdings approach has been used 
elsewhere and I have simply not found it.

With thanks,
Vladimir Beskorovainyi
admin@besk.tech
https://vladimir.besk.tech


_______________________________________________
Unsubscribe at: http://dhhumanist.org/Restricted
List posts to: humanist@dhhumanist.org
List info and archives at at: http://dhhumanist.org
Listmember interface at: http://dhhumanist.org/Restricted/
Subscribe at: http://dhhumanist.org/membership_form.php