PaperlayerView exact revision on arXiv

Source-linked AI summary

Mining the Nineteenth Century for the Thirteenth: Register-Scale Forced Alignment and HTR Bootstrapping via Legacy Regesta and Antiquarian Editions

Tamas Kovacs, Daniel Luger

arXiv:2610.36956v1cs.CVcs.CL

TL;DR

The registers of Honorius III hold more than 5,000 letters and lack a modern edition; how antiquarian editions differ from modern transcriptions and what that costs a recognizer had not been measured. The study audits expert transcriptions against legacy texts and evaluates alignment, spelling projection, recognizer adaptation, vocabulary support, and entry structuring using the overlapping folios of Reg. Vat. 9. −1.50 WER points pooled over all year-2 sides with projection and replay of modern source lines; legacy vocabulary covers 82.2% of OOV words of unedited entries and identifies recognition errors with 66.7% precision.

  • Problem

    The registers of Honorius III hold more than 5,000 letters and lack a modern edition; how antiquarian editions differ from modern transcriptions and what that costs a recognizer had not been measured.

    Source: §1 · p. 1

  • Method

    The study audits expert transcriptions against legacy texts and evaluates alignment, spelling projection, recognizer adaptation, vocabulary support, and entry structuring using the overlapping folios of Reg. Vat. 9.

    Source: §1 · p. 2, §2 · p. 2

  • Results

    −1.50 WER points pooled over all year-2 sides with projection and replay of modern source lines; legacy vocabulary covers 82.2% of OOV words of unedited entries and identifies recognition errors with 66.7% precision.

    Source: §8 · p. 9

  • Takeaways & Limitations

    Modern transcriptions remain the primary training data, while legacy texts can adapt the recognizer, check vocabulary, and structure its output.

    Source: §1 · p. 2, §8 · p. 9

    Measurements come from one register volume; the year-2 test set has 13 pages, legacy-text coverage beyond fol. 203 is not known in advance, and linking depends on uncorrected Internet Archive OCR.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Handwritten text recognition (HTR) for the registers of Pope Honorius III depends on modern expert transcriptions, which exist only for regnal years 1-2 of Reg. Vat. 9. For the remaining registers, nineteenth-century scholarship offers Pressutti's regesta and partial editions. On fol. 1-203, where both layers exist, we measure what this legacy layer can contribute, on transcriptions that we first audit against it, and find a division of labor across three tiers. At the orthographic tier, raw editions are at best neutral for adapting a recognizer to the second regnal year and harmful without replay, whereas projection onto curial spelling with replay of modern source lines turns them into zero-label adaptation targets: pooled over all 134 year-2 sides, the WER falls by 1.50 points (95% CI -1.84 to -1.15), closing 33% of the gap to an oracle; the spelling changes at the year split, through the scribe or the transcriber, whereas the hand changes visibly only at fol. 191r. At the lexical tier, the corpus-wide legacy lexicon contains 82.2% of the out-of-vocabulary words of unedited entries and flags recognition errors with 66.7% precision, and labels from pre-trained medieval NER models, distilled into the recognizer's NER head, reach F1 82.1% on an expert validated benchmark, level with the best tagger applied after recognition. At the topological tier, Pressutti's regests, the closing datum clauses, and the regests' dates converted to the chancery's Roman form find 73.9% of entry starts in the recognized line stream, and edition agreement estimates a CER of 7.80% beyond fol. 203. Antiquarian scholarship thus cannot replace modern transcription, but it can steer, check, and segment HTR across the untranscribed registers.

1 Introduction

Modern expert transcriptions cover only the first two regnal years of Honorius III’s registers, while nineteenth-century regesta and editions provide a potential resource for the remaining folios. The paper measures how this legacy material can support recognition, identifying different roles for it in adapting, checking, and structuring HTR.

1 Introduction

  • The registers of Honorius III contain more than 5,000 letters but have no modern edition, and expert transcriptions cover only the first two regnal years.The recognizer’s held-out performance on the same volume was 12.5% WER, or 8.7% after re-alignment, bidirectional decoding with BPE targets, and language-model rescoring.Source: §1 · p. 1
  • Pressutti’s regesta and nineteenth-century full-text editions provide legacy texts for the remaining folios, but they modernize spelling, abridge formulae, expand initials, and contain OCR noise.Full texts include Horoy’s Opera omnia, Rodenberg’s MGH selection, Raynaldi’s Annales ecclesiastici, and Theiner’s regional collections.Source: §1 · p. 1
  • The paper asks how far antiquarian editions differ from modern transcriptions and what that distance costs a recognizer.Earlier work aligned images with existing editions and found that models reproduce their editorial conventions; the papal registers had instead used crowd-sourced or expert line transcriptions.Source: §1 · p. 1
  • The study measures legacy alignment and spelling projection, vocabulary and name support, and entry segmentation against modern transcriptions where both layers exist.Its proposed division of labor is for modern transcriptions to train the recognizer, while legacy sources adapt it, check its vocabulary, and structure its output.Source: §1 · pp. 1–2

2 Data, Recognizers, and Legacy Sources

The study combines expert transcriptions, recognizer outputs, Pressutti’s regesta, and nineteenth-century editions for Reg. Vat. 9. An audit uses agreement between editions and an independent HTR reading to correct some transcription words before experiments.

2 Data, Recognizers, and Legacy Sources

  • An automatic audit flagged 419 of 84,660 words where the projected edition and an out-of-fold HTR reading agreed against the transcription; 355 words (0.42%) with letter-for-letter agreement were corrected.Source: §2 · p. 2
  • The data comprise 394 pages, 860 expert-transcribed entries, and 13,122 modern silver lines divided across 315 training, 40 development, and 39 test pages.The recognizers include a Conformer-CTC model, a hybrid with attention decoding and an NER head, and an R2L-BPE hybrid with language-model rescoring.Source: §2 · p. 2
  • Pressutti’s OCR was parsed into 5,059 regesta with entry metadata; 626 of the 860 transcribed entries were linked through Pressutti’s numbering.Normalization enabled recovery of 3,845 regest addresses and 3,928 letter numbers; 46.7% of regesta name no printed edition, an upper bound because citations are incomplete.Source: §2 · p. 2
  • Table 1 covers transcribed Reg. Vat. 9 entries from years 1–2 linked to a nineteenth-century full text, grouped by the edition that matched best.The supplied caption establishes the table’s scope, not its cell values.Source: Table 1 · p. 2
  • Text matching linked 509 entries to nineteenth-century full texts and recovered 249 of the 267 entries for which Pressutti cites Horoy.For 358 entries, the edition span was within distance 0.30 of the expert text; most remaining span mismatches came from running heads and footnotes at page breaks.Source: §2 · p. 3

3 Alignment to Nineteenth-Century Texts

Legacy editions can be aligned to recognized lines, and curial projection reduces spelling differences between the editions and modern transcriptions. The study also tests entry anchoring, formula bridging, and rules learned from aligned examples.

3 Alignment to Nineteenth-Century Texts

  • Entry anchoring assigns lines in reading order to non-decreasing entries, while formula bridging inserts likely continuations after etc.; curial projection maps targets toward register spelling.The baseline semi-global aligner accepts distances ≤0.35 and treats distances ≤0.25 with at least 15 characters as silver lines.Source: §3 · p. 3
  • After curial projection, legacy targets have 3.13% reading-layer CER and 10.8% WER, compared with 3.40% CER and 12.5% WER raw.Projection lowers strict-layer CER from 3.86% to 3.21%; 98.9% of legacy lines are placed on the same text as the modern target.Source: §3 · p. 3, Table 3 · p. 4
  • Projection removes classicizing spelling, while edition emendations and abridgements, OCR noise, and line-edge cuts account for most remaining differences.Anchoring adds few lines at slightly lower fidelity, and bridging applies to only 10 lines because the register itself abridges most formulae.Source: §3 · p. 3
  • Learned from 3,171 year-1 silver lines, 46 projection rules lower strict-layer CER to 3.38% on held-out year-1 pages and 3.34% on year 2.The table describes comparisons of hand-written and learned projection on all strict-link pairs, held-out year-1 pages, and year 2, with some rules optionally reader-gated.Source: §3 · pp. 3–4

4 Training on Antiquarian Lines Only

Training on antiquarian lines alone yields a usable but weaker recognizer than training on modern lines. Curial projection improves the legacy targets, and a small modern seed helps further.

4 Training on Antiquarian Lines Only

  • 14.5% WER for the hybrid trained on as many modern lines compares with 20.1% for the projected-legacy hybrid, so editorial divergence costs more than halving modern data.Projected legacy lines yield 20.1% WER for the hybrid and 23.6% for C, versus 23.3% and 26.9% on raw lines; adding a small modern seed brings the hybrid to 18.1%.Source: §4 · p. 4
  • Projection improves the hybrid by 3.2 WER points and C by 3.4 points; adding a small modern seed improves the hybrid by another 2.0 points.The experiments train on raw or projected nineteenth-century lines, with a separate run combining projected lines and 1,052 random modern lines.Source: §4 · p. 4
  • Clause-final words have 24–35% WER, where the editions abridge closing formulae.Table 4’s caption specifies strict-layer CER and WER comparisons for hand-written and learned projection across all strict-link pairs and held-out year-1 and year-2 pages.Source: §4 · p. 4, Table 4 · p. 4

5 Transfer to the Second Regnal Year

Projected legacy lines with modern source-line replay improve year-2 recognition without modern year-2 labels, while raw lines can harm performance without replay. The benefit tracks the spelling shift across year 2; the later hand change adds no measurable gain of its own.

5 Transfer to the Second Regnal Year

  • Year 2 differs in spelling from the year split, while the hand changes only on fol. 191–203; the legacy lines adapt to the spelling shared across both segments.The authors report no measurable additional gain from the new hand, while modern labels yield more gain there.Source: §5 · pp. 4–5
  • Without replay, raw legacy lines raise pooled WER by 2.2 points; replay helps, and projection adds further improvement over raw lines.Classicizing Horoy raw lines worsen WER by 0.7 points, whereas raw MGH and Theiner lines improve it by 0.5 and 0.8 points, respectively.Source: §5 · p. 5
  • −1.50 WER points (95% CI −1.84 to −1.15): across 134 year-2 sides, projected legacy lines with replay lower WER from 17.0% to 15.5%, closing 33% of the gap to the 12.5% oracle.This pooled cross-validation estimate is the one the authors rely on.Source: §5 · pp. 4–5, Table 6 · p. 5

Beyond fol. 203

  • Beyond fol. 203, edition agreement on covered lines yields a calibrated estimate of 7.80% CER, with 48% line coverage and higher agreement CER than on transcribed pages.The block shows no hand change; its only visual boundary is before fol. 191r.Source: §5 · pp. 5–6

6 Pressutti as an Entry Segmenter

Pressutti’s regesta, closing datum clauses, and converted Roman dates provide complementary signals for segmenting recognized lines into entries. Combining the signals finds more starts than either Pressutti or the datum baseline alone, including on entries without editions.

6 Pressutti as an Entry Segmenter

  • Combining Pressutti with closing datum clauses finds 68.7% of starts (F1 63.7%), compared with F1 58.5% for the datum baseline.Pressutti alone finds 36.3% of starts within one line, at 60.2% precision.Source: §6 · p. 6
  • The monotone dynamic-programming alignment does not add recall, locating the bottleneck in matching regests to lines rather than in their order.Replacing OCR regest strings with expert address and incipit text raises union recall only to 68.5%.Source: §6 · p. 6
  • 73.9% of entry starts are found with the Roman-date anchor added to the union (F1 65.6%, 95% CI 62.5–68.5); recall reaches 73.4% on entries without an edition.The date anchor adds 5.1 recall points to the union.Source: §6 · pp. 6–7
  • The date anchor converts Pressutti’s modern dates into the chancery’s Roman form; a recognized regest date marks the letter’s closing line, so the next line starts an entry.Dates for 877 of 970 regests agree with Pressutti’s printed Roman dates in 89% of the 779 cases where he provides one.Source: §6 · p. 6

7 Legacy Vocabulary and Neural NER

The legacy layer is more useful for lexical coverage, error review, and entity supervision than for steering recognition. Distilled medieval NER labels improve the recognizer’s entity tagging, while regests also support review and sigla expansion.

Folio gazetteer

  • The folio gazetteer changes name WER only from 29.9% to 29.5%, and contextual biasing leaves it at 29.7%.The attention beam lacks correct name forms for many misreads, while regest summaries are too sparse to steer decoding.Source: §7 · p. 7, Table 8 · p. 7

Corpus-wide legacy lexicon

  • 82.2% of training-OOV words in test entries without printed editions occur in the corpus-wide lexicon in exactly the same form.Coverage rises to 93.5% by stem; the lexicon includes the regesta and eleven edition volumes.Source: §7 · p. 7, Table 9 · p. 7
  • 66.7% precision for words outside the leak-free legacy lexicon makes absence a useful error flag, while the flag catches 30.9% of wrong words.Combining it with the hybrid posterior routes 1.2% of words to review at 93.0% precision.Source: §7 · p. 7

Legacy lexicon in decoding

  • Legacy language models and post-correction lower test WER from 10.0% to 9.8%, but the total −0.16-point change is not distinguishable from zero.Name WER does not improve, so the authors position the lexicon as a review signal rather than a decoding aid.Source: §7 · p. 8

Neural NER

  • 82.1% F1 with distilled ensemble labels matches Span-NER’s 81.7% F1 post hoc, while keeping recognition WER effectively unchanged at 11.1% versus 11.0%.The result is on an expert-validated benchmark; the distilled head adds only a linear layer to decoding.Source: Table 10 · p. 8, §7 · p. 8

Sigla

  • The sigla layer proposes full-name candidates without changing recognized text, covering 83% of sigla on fol. 204–287.It is intended as a candidate list for expert review, not automatic resolution.Source: §7 · p. 8

8 Discussion

The legacy layer contributes across orthographic, lexical, and topological tasks, but it does not replace modern transcription. Its value depends on projection, replay, and the distinct roles of editions and regesta.

Orthographic tier

  • −1.50 WER points pooled over 134 year-2 sides result when projected legacy editions and replay of modern source lines provide zero-label adaptation targets.Raw editions are at best neutral for an adapted recognizer and harmful without replay; the gain closes 33% of the oracle gap.Source: §8 · p. 9
  • Legacy text cannot replace modern training data: its WER is 20.1% against 14.5% with as many modern lines.Where projected editions and an independent reading agree against a transcription, 88% of decidable corrections in a judged sample hold.Source: §8 · p. 9

Lexical tier

  • The corpus-wide lexicon covers 82.2% of OOV words in unedited entries and flags recognition errors with 66.7% precision; distilled NER reaches 82.1% F1.The distilled head is level with the best post-hoc tagger on the expert-validated benchmark; regesta also propose sigla expansions for review.Source: §8 · p. 9

Topological tier

  • Pressutti’s regests, closing datum clauses, and Roman-form dates identify 73.9% of entry starts, while edition agreement estimates 7.80% CER beyond fol. 203.The reported misses in entry finding lie in regest parsing and folio reconciliation, not OCR of the regests.Source: §8 · p. 9

Limitations

The evidence is bounded by one register volume and a small year-2 test set, with additional uncertainty in source coverage, OCR, annotation, and palaeographic interpretation.

Scope

  • All measurements come from one register volume, and the year-2 test set contains 13 pages.Transfer gain varies by about a point across training seeds; the cross-validated estimate over 134 sides is half the 13-page gain.Source: Limitations · p. 9

Legacy sources

  • Nineteenth-century texts cover about half of transcribed entries, and their coverage beyond fol. 203 is not known in advance.Linking and alignment depend on Internet Archive OCR that was not corrected.Source: Limitations · p. 9

Benchmark and palaeography

  • The NER benchmark’s unchecked agreed labels retain an estimated 11% error, and its entity conventions may differ from another annotator’s.The year split also confounds hand and transcriber; the inferred boundaries await palaeographic checking.Source: Limitations · p. 9
Loading 2610.36956v1…