pith. machine review for the scientific record. sign in

arxiv: 2604.06160 · v1 · submitted 2026-04-07 · 💻 cs.CV · cs.LG

Recognition: 2 theorem links

· Lean Theorem

The Character Error Vector: Decomposable errors for page-level OCR evaluation

Jonathan Bourne , Mwiza Simbeye , Joseph Nockels
Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:05 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords OCR evaluationCharacter Error RateDocument UnderstandingError decompositionPage-level metricsText extractionArchival documentsPipeline triage
0
0 comments X

The pith

The Character Error Vector evaluates page-level OCR by decomposing errors into parsing, OCR, and interaction components.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard Character Error Rate becomes unusable when text parsing is imperfect, which is common in real documents. This paper proposes the Character Error Vector as a bag-of-characters alternative that stays defined and can be broken down by error type. The breakdown separates mistakes from parsing the page layout, from the actual character recognition, and from how those two interact. Users can therefore direct improvement efforts at the weakest link in their document processing chain. Validation on archival newspaper images shows that the vector correlates with known metrics and reveals that traditional pipelines often outperform modern end-to-end models on degraded material.

Core claim

We introduce the Character Error Vector, a bag-of-characters evaluator for OCR that remains defined under page-parsing errors. The CEV decomposes into parsing error, OCR error, and interaction error components using methods such as SpACER or Jensen-Shannon character distribution distance. This allows direct measurement of page-level OCR quality and triage of which pipeline stage needs attention, as demonstrated by superior performance of traditional methods over end-to-end models on complex archival data.

What carries the argument

The Character Error Vector, a bag-of-characters representation of OCR output that is decomposed into additive parsing error, OCR error, and interaction error terms.

If this is right

  • Practitioners can identify whether parsing or OCR is the main bottleneck in a text extraction pipeline.
  • Simple thresholding on available values predicts the dominant error source with an F1 of 0.91.
  • Traditional pipeline approaches outperform state-of-the-art end-to-end models on degraded archival newspapers with complex layouts.
  • The CEV acts as a bridge connecting page-level parsing quality measures to local character accuracy metrics like CER.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same decomposition approach could be tested on handwritten or multilingual documents to check whether interaction terms behave similarly.
  • Reporting the three CEV components alongside any single score would give clearer diagnostics across different document collections.
  • Future pipelines might optimize parsing and recognition stages separately once the error contributions are routinely measured.

Load-bearing premise

The decomposition accurately separates the error sources without the specific implementations of SpACER or Jensen-Shannon distance introducing confounding effects.

What would settle it

A set of pages where the CEV attributes most error to one component, such as parsing, but manual inspection of the same pages shows that the dominant source is actually OCR recognition mistakes.

Figures

Figures reproduced from arXiv: 2604.06160 by Jonathan Bourne, Joseph Nockels, Mwiza Simbeye.

Figure 1
Figure 1. Figure 1: CDD decomposition when character position information is available. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of the impact of spatial uncertainty on the character position [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: We can see that SpACER has good correlation with CER at page [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: PPDoc-m and PPDoc-S, fail to correctly parse the images resulting in [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The figure shows that even conservative cutoffs for SpACER and [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
read the original abstract

The Character Error Rate (CER) is a key metric for evaluating the quality of Optical Character Recognition (OCR). However, this metric assumes that text has been perfectly parsed, which is often not the case. Under page-parsing errors, CER becomes undefined, limiting its use as a metric and making evaluating page-level OCR challenging, particularly when using data that do not share a labelling schema. We introduce the Character Error Vector (CEV), a bag-of-characters evaluator for OCR. The CEV can be decomposed into parsing and OCR, and interaction error components. This decomposability allows practitioners to focus on the part of the Document Understanding pipeline that will have the greatest impact on overall text extraction quality. The CEV can be implemented using a variety of methods, of which we demonstrate SpACER (Spatially Aware Character Error Rate) and a Character distribution method using the Jensen-Shannon Distance. We validate the CEV's performance against other metrics: first, the relationship with CER; then, parse quality; and finally, as a direct measure of page-level OCR quality. The validation process shows that the CEV is a valuable bridge between parsing metrics and local metrics like CER. We analyse a dataset of archival newspapers made of degraded images with complex layouts and find that state-of-the-art end-to-end models are outperformed by more traditional pipeline approaches. Whilst the CEV requires character-level positioning for optimal triage, thresholding on easily available values can predict the main error source with an F1 of 0.91. We provide the CEV as part of a Python library to support Document understanding research.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Character Error Vector (CEV), a bag-of-characters evaluator for page-level OCR that remains defined under parsing errors (unlike CER). It claims the CEV decomposes into parsing, OCR, and interaction error components, demonstrated with SpACER (spatially aware) and Jensen-Shannon character distribution methods. Validation shows relationships to CER and parse quality, direct page-level quality assessment, and F1=0.91 for predicting main error source via thresholding on available values. Analysis of degraded archival newspapers finds traditional pipelines outperform end-to-end models; a Python library is provided.

Significance. If the decomposition is robust, the CEV provides a practical bridge between parsing metrics and local metrics like CER, enabling targeted improvements in document understanding pipelines by identifying whether parsing or OCR is the dominant error source. The library release and empirical results on complex archival data strengthen utility and reproducibility.

major comments (2)
  1. [Validation and method sections (decomposition and error-source prediction)] The central claim of clean decomposability into parsing/OCR/interaction components (abstract and validation description) requires that component ratios remain stable under alternative character alignments; the paper demonstrates SpACER and Jensen-Shannon but does not report tests with standard Levenshtein or optimal transport, leaving open whether attribution is confounded by the specific matching procedure.
  2. [Validation process (page-level OCR quality and error-source prediction)] The reported F1=0.91 for main-error-source prediction via thresholding lacks details on threshold selection, cross-validation procedure, or sensitivity analysis; without these, it is unclear whether the triage guidance generalizes beyond the specific dataset and implementations.
minor comments (2)
  1. The abstract states the CEV is provided as part of a Python library; the manuscript should include an explicit link, GitHub reference, or installation instructions to support reproducibility.
  2. Consider reporting confidence intervals or error bars on the F1 score and other quantitative validation results (e.g., relationships with CER) to better convey uncertainty.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the robustness and reproducibility of our proposed Character Error Vector (CEV). We address each major comment point by point below and will revise the manuscript accordingly to strengthen the validation sections.

read point-by-point responses

  1. Referee: [Validation and method sections (decomposition and error-source prediction)] The central claim of clean decomposability into parsing/OCR/interaction components (abstract and validation description) requires that component ratios remain stable under alternative character alignments; the paper demonstrates SpACER and Jensen-Shannon but does not report tests with standard Levenshtein or optimal transport, leaving open whether attribution is confounded by the specific matching procedure.

    Authors: We appreciate the referee's emphasis on verifying the stability of the decomposition. The CEV is defined as a bag-of-characters vector difference between predicted and ground-truth counts, with the decomposition into parsing, OCR, and interaction terms derived directly from this vector (independent of any particular alignment). SpACER and Jensen-Shannon were chosen as representative methods because they explicitly handle spatial and distributional mismatches common in page-level OCR. However, we acknowledge that reporting results under standard Levenshtein alignment and optimal transport would provide stronger evidence that component ratios are not artifacts of the chosen matching procedure. In the revised manuscript, we will add a new subsection in the validation experiments that recomputes the CEV decomposition using Levenshtein-based alignment on the same archival newspaper dataset and reports the resulting component ratios and error-source predictions for comparison. revision: yes

  2. Referee: [Validation process (page-level OCR quality and error-source prediction)] The reported F1=0.91 for main-error-source prediction via thresholding lacks details on threshold selection, cross-validation procedure, or sensitivity analysis; without these, it is unclear whether the triage guidance generalizes beyond the specific dataset and implementations.

    Authors: We agree that the current description of the F1=0.91 result is insufficiently detailed for full reproducibility and assessment of generalizability. The reported score was obtained by thresholding on easily computable page-level statistics (parse quality proxy and aggregate error magnitude) to classify the dominant error source. In the revised manuscript, we will expand the relevant validation subsection to specify: the exact threshold values and selection method (grid search over a held-out validation split of the archival dataset), the cross-validation procedure used to compute the F1, and a sensitivity analysis plotting F1 as a function of small perturbations around the chosen thresholds. These additions will clarify the scope of the triage guidance while preserving the original claim that thresholding on readily available values can achieve high predictive performance. revision: yes

Circularity Check

0 steps flagged

No load-bearing circularity; CEV decomposition defined independently and validated externally

full rationale

The paper introduces the CEV as a bag-of-characters evaluator for page-level OCR that handles parsing errors where CER is undefined. It defines a decomposition into parsing, OCR, and interaction components, then demonstrates two implementations (SpACER and Jensen-Shannon character distribution) and validates the metric against independent external quantities: relationship to CER, parse quality, and direct page-level OCR quality. Thresholding on available values predicts error source with reported F1 0.91. No equations or claims reduce the decomposition to a fitted parameter renamed as prediction, no self-definitional loops appear, and no uniqueness theorems or ansatzes are imported via self-citation as load-bearing premises. The central claim remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the definition of CEV and its three-way decomposition, which uses standard distance measures but introduces new error-component categories without external grounding beyond the validation experiments.

axioms (1)
  • domain assumption Jensen-Shannon distance is a suitable measure for comparing character distributions in OCR output
    Invoked for one of the two CEV implementations described in the abstract.
invented entities (1)
  • Character Error Vector (CEV) no independent evidence
    purpose: Decomposable bag-of-characters error metric for page-level OCR evaluation
    Newly defined construct that enables the reported decomposition and validation.

pith-pipeline@v0.9.0 · 5597 in / 1218 out tokens · 58687 ms · 2026-05-10T19:05:49.800294+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

48 extracted references · 30 canonical work pages · 3 internal anchors

  1. [1]

    Qwen2.5-VL Technical Report

    Shuai Bai et al.Qwen2.5-VL Technical Report. arXiv:2502.13923 [cs]. Feb. 2025.doi:10.48550/arXiv.2502.13923.url:http://arxiv.org/abs/ 2502.13923(visited on 04/04/2026). 22

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.13923.url:http://arxiv.org/abs/ 2025

  2. [2]

    Reading the unreadable: creating a dataset of 19th century English newspapers using image-to-text language models

    Jonathan Bourne. “Reading the unreadable: creating a dataset of 19th century English newspapers using image-to-text language models”. In: Digital Scholarship in the Humanities(Dec. 2025), fqaf151.issn: 2055- 7671.doi:10.1093/llc/fqaf151.url:https://doi.org/10.1093/ llc/fqaf151(visited on 02/04/2026)

    work page doi:10.1093/llc/fqaf151.url:https://doi.org/10.1093/ 2025

  3. [3]

    Scrambled text: fine-tuning language models for OCR error correction using synthetic data

    Jonathan Bourne. “Scrambled text: fine-tuning language models for OCR error correction using synthetic data”. en. In:International Journal on Document Analysis and Recognition (IJDAR)28.4 (Dec. 2025), pp. 741– 755.issn: 1433-2825.doi:10.1007/s10032-025-00522-0.url:https: //doi.org/10.1007/s10032-025-00522-0(visited on 02/06/2026)

    work page doi:10.1007/s10032-025-00522-0.url:https: 2025

  4. [4]

    arXiv:2603.12718 [cs]

    Jonathan Bourne, Mwiza Simbeye, and Ishtar Govia.The COTe score: A decomposable framework for evaluating Document Layout Analysis models. arXiv:2603.12718 [cs]. Mar. 2026.doi:10 . 48550 / arXiv . 2603 . 12718. url:http://arxiv.org/abs/2603.12718(visited on 03/16/2026)

    work page arXiv 2026

  5. [5]

    Flexible character accuracy measure for reading-order-independent eval- uation

    Christian Clausner, Stefan Pletschacher, and Apostolos Antonacopoulos. “Flexible character accuracy measure for reading-order-independent eval- uation”. In:Pattern Recognition Letters131 (Mar. 2020), pp. 390–397. issn: 0167-8655.doi:10.1016/j.patrec.2020.02.003.url:https:// www.sciencedirect.com/science/article/pii/S0167865520300416 (visited on 04/04/2026)

    work page doi:10.1016/j.patrec.2020.02.003.url:https:// 2020

  6. [6]

    “Q i-jtb the Raven

    Ryan Cordell. ““Q i-jtb the Raven”: Taking Dirty OCR Seriously”. In: Book History20 (2017), pp. 188–225.issn: 1098-7371.url:https : / / www.jstor.org/stable/48554671(visited on 04/07/2026)

    work page arXiv 2017

  7. [7]

    PaddleOCR 3.0 Technical Report

    Cheng Cui et al.PaddleOCR 3.0 Technical Report. arXiv:2507.05595 [cs]. July 2025.doi:10.48550/arXiv.2507.05595.url:http://arxiv.org/ abs/2507.05595(visited on 04/04/2026)

    work page internal anchor Pith review doi:10.48550/arxiv.2507.05595.url:http://arxiv.org/ 2025

  8. [8]

    arXiv:2603.24373 [cs]

    Cheng Cui et al.PP-OCRv5: A Specialized 5M-Parameter Model Rivaling Billion-Parameter Vision-Language Models on OCR Tasks. arXiv:2603.24373 [cs]. Mar. 2026.doi:10.48550/arXiv.2603.24373.url:http://arxiv. org/abs/2603.24373(visited on 04/03/2026)

    work page doi:10.48550/arxiv.2603.24373.url:http://arxiv 2026

  9. [9]

    Shuaiqi Duan et al.GLM-OCR Technical Report. en. arXiv:2603.10910 [cs]. Mar. 2026.doi:10.48550/arXiv.2603.10910.url:http://arxiv. org/abs/2603.10910(visited on 04/04/2026)

    work page doi:10.48550/arxiv.2603.10910.url:http://arxiv 2026

  10. [10]

    A genetic sparse distributed mem- ory approach to the application of handwritten character recognition

    Kuo-Chin Fan and Yuan-Kai Wang. “A genetic sparse distributed mem- ory approach to the application of handwritten character recognition”. In:Pattern Recognition30.12 (Dec. 1997), pp. 2015–2022.issn: 0031- 3203.doi:10 . 1016 / S0031 - 3203(97 ) 00017 - 4.url:https : / / www . sciencedirect.com/science/article/pii/S0031320397000174(vis- ited on 03/26/2026)

    1997

  11. [11]

    Drew Hemment and Cory Kommers.Doing AI differently: rethinking the foundations of AI via the humanities. Ed. by Ruth Ahnert et al. London: The Alan Turing Institute, July 2025.doi:10.5281/zenodo.1642129. 23

    work page doi:10.5281/zenodo.1642129 2025

  12. [12]

    arXiv:2507.23295 [cs]

    Inbum Heo et al.LED Benchmark: Diagnosing Structural Layout Errors for Document Layout Analysis. arXiv:2507.23295 [cs]. July 2025.doi:10. 48550/arXiv.2507.23295.url:http://arxiv.org/abs/2507.23295 (visited on 12/13/2025)

    work page arXiv 2025

  13. [13]

    Cross.Tracking Amendments to Leg- islation and Other Political Texts with a Novel Minimum-Edit-Distance Algorithm: DocuToads

    Henrik Hermansson and James P. Cross.Tracking Amendments to Leg- islation and Other Political Texts with a Novel Minimum-Edit-Distance Algorithm: DocuToads. arXiv:1608.06459 [cs] version: 1. Aug. 2016.doi: 10 . 48550 / arXiv . 1608 . 06459.url:http : / / arxiv . org / abs / 1608 . 06459(visited on 04/04/2026)

    work page arXiv 2016

  14. [15]

    July 2025.url:https://huggingface

    IBM.Granite Docling Model Card. July 2025.url:https://huggingface. co/ibm-granite/granite-docling-258M(visited on 03/27/2026)

    2025

  15. [16]

    original-date: 2020-03-14T11:46:39Z

    JaidedAI.EasyOCR. original-date: 2020-03-14T11:46:39Z. Mar. 2026.url: https://github.com/JaidedAI/EasyOCR(visited on 03/27/2026)

    2020

  16. [17]

    Transkribus - A Service Platform for Transcription, Recognition and Retrieval of Historical Documents

    Philip Kahle et al. “Transkribus - A Service Platform for Transcription, Recognition and Retrieval of Historical Documents”. In:2017 14th IAPR International Conference on Document Analysis and Recognition (IC- DAR). Vol. 04. ISSN: 2379-2140. Nov. 2017, pp. 19–24.doi:10.1109/ ICDAR . 2017 . 307.url:https : / / ieeexplore . ieee . org / document / 8270253(v...

    2017

  17. [18]

    eScriptorium: An Open Source Platform for Historical Document Analysis

    Benjamin Kiessling et al. “eScriptorium: An Open Source Platform for Historical Document Analysis”. In:2019 International Conference on Doc- ument Analysis and Recognition Workshops (ICDARW). Vol. 2. Sept. 2019, pp. 19–19.doi:10 . 1109 / ICDARW . 2019 . 10032.url:https : / / ieeexplore . ieee . org / abstract / document / 8893029(visited on 04/03/2026)

    2019

  18. [19]

    Saving the Bits

    William Kilbride. “Saving the Bits”. en. In: (Nov. 2015). Book Title: A New Companion to Digital Humanities ISBN: 9781118680605, pp. 408– 419.doi:10.1002/9781118680605.ch28.url:https://onlinelibrary. wiley.com/doi/10.1002/9781118680605.ch28(visited on 04/07/2026)

    work page doi:10.1002/9781118680605.ch28.url:https://onlinelibrary 2015

  19. [20]

    Loghi: An End-to-End Framework for Mak- ing Historical Documents Machine-Readable

    Rutger van Koert et al. “Loghi: An End-to-End Framework for Mak- ing Historical Documents Machine-Readable”. en. In:Document Analysis and Recognition – ICDAR 2024 Workshops. Ed. by Harold Mouch` ere and Anna Zhu. Cham: Springer Nature Switzerland, 2024, pp. 73–88.isbn: 978-3-031-70645-5.doi:10.1007/978-3-031-70645-5_6

    work page doi:10.1007/978-3-031-70645-5_6 2024

  20. [21]

    Document Language Models, Query Models, and Risk Minimization for Information Retrieval

    John Lafferty and Chengxiang Zhai. “Document Language Models, Query Models, and Risk Minimization for Information Retrieval”. In:SIGIR Forum51.2 (Aug. 2017), pp. 251–259.issn: 0163-5840.doi:10 . 1145 / 3130348.3130375.url:https://dl.acm.org/doi/10.1145/3130348. 3130375(visited on 02/07/2026). 24

    work page doi:10.1145/3130348 2017

  21. [22]

    Binary codes capable of correcting deletions, in- sertions, and reversals

    Vladimir Levenshtein. “Binary codes capable of correcting deletions, in- sertions, and reversals”. In:Soviet physics doklady10 (1966), pp. 707– 710

    1966

  22. [23]

    Trocr: Transformer-based optical character recognition with pre-trained models, 2022

    Minghao Li et al.TrOCR: Transformer-based Optical Character Recogni- tion with Pre-trained Models. arXiv:2109.10282 [cs]. Sept. 2022.doi:10. 48550/arXiv.2109.10282.url:http://arxiv.org/abs/2109.10282 (visited on 03/27/2026)

    work page arXiv 2022

  23. [24]

    dots.ocr: Multilingual Document Layout Pars- ing in a Single Vision-Language Model.arXiv preprint arXiv:2512.02498, 2025

    Yumeng Li et al.dots.ocr: Multilingual Document Layout Parsing in a Single Vision-Language Model. arXiv:2512.02498 [cs]. Dec. 2025.doi:10. 48550/arXiv.2512.02498.url:http://arxiv.org/abs/2512.02498 (visited on 03/27/2026)

    work page arXiv 2025

  24. [25]

    25 Zhiqiu Lin, Siyuan Cen, Daniel Jiang, Jay Karhade, Hewei Wang, Chancharik Mitra, Tiffany Ling, Yuhan Huang, Sifan Liu, Mingyu Chen, et al

    Zhang Li et al.MonkeyOCR: Document Parsing with a Structure-Recognition- Relation Triplet Paradigm. arXiv:2506.05218 [cs]. Feb. 2026.doi:10 . 48550/arXiv.2506.05218.url:http://arxiv.org/abs/2506.05218 (visited on 04/04/2026)

    work page arXiv 2026

  25. [26]

    Said Gurbuz, and Peter W

    Nikolaos Livathinos et al.Advanced Layout Analysis Models for Docling. arXiv:2509.11720 [cs]. Sept. 2025.doi:10 . 48550 / arXiv . 2509 . 11720. url:http://arxiv.org/abs/2509.11720(visited on 02/04/2026)

    work page arXiv 2025

  26. [27]

    Diverging Diver- gences: Examining Variants of Jensen Shannon Divergence for Corpus Comparison Tasks

    Jinghui Lu, Maeve Henchion, and Brian Mac Namee. “Diverging Diver- gences: Examining Variants of Jensen Shannon Divergence for Corpus Comparison Tasks”. eng. In:Proceedings of the Twelfth Language Re- sources and Evaluation Conference. Ed. by Nicoletta Calzolari et al. Mar- seille, France: European Language Resources Association, May 2020, pp. 6740– 6744....

    2020

  27. [28]

    Application of artificial neural network model for optical character recognition

    N. Mani and B. Srinivasan. “Application of artificial neural network model for optical character recognition”. In:Computational Cybernetics and Sim- ulation 1997 IEEE International Conference on Systems, Man, and Cy- bernetics. Vol. 3. ISSN: 1062-922X. Oct. 1997, 2517–2520 vol.3.doi:10. 1109 / ICSMC . 1997 . 635312.url:https : / / ieeexplore . ieee . org ...

    1997

  28. [29]

    Word ranking in a single document by Jensen–Shannon divergence

    Ali Mehri, Maryam Jamaati, and Hassan Mehri. “Word ranking in a single document by Jensen–Shannon divergence”. In:Physics Letters A 379.28 (Aug. 2015), pp. 1627–1632.issn: 0375-9601.doi:10 . 1016 / j . physleta . 2015 . 04 . 030.url:https : / / www . sciencedirect . com / science/article/pii/S0375960115003722(visited on 02/07/2026)

    2015

  29. [30]

    NLS.The Spiritualist – Data Foundry. en-US. 2019.doi:https : / / doi. org/10.34812 /3b0g- 3j88.url:https:// data.nls. uk/data/ digitised-collections/spiritualist-newspapers/(visited on 04/07/2026)

    2019

  30. [31]

    2026.url:https: //huggingface.co/datasets/NationalLibraryOfScotland/Spiritualist_ Newspaper(visited on 04/07/2026)

    Joseph Nockels.Spiritualist Newspaper Transcription. 2026.url:https: //huggingface.co/datasets/NationalLibraryOfScotland/Spiritualist_ Newspaper(visited on 04/07/2026). 25

    2026

  31. [32]

    2025.url:https://github.com/VikParuchuri/ surya

    Vikas Paruchuri and Datalab Team.Surya: A lightweight document OCR and analysis toolkit. 2025.url:https://github.com/VikParuchuri/ surya

    2025

  32. [33]

    MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers

    Krishna Pillutla et al. “MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers”. en. In: ()

  33. [34]

    The PAGE (Page Analysis and Ground-Truth Elements) Format Framework

    Stefan Pletschacher and Apostolos Antonacopoulos. “The PAGE (Page Analysis and Ground-Truth Elements) Format Framework”. In:2010 20th International Conference on Pattern Recognition. Aug. 2010, pp. 257–260. doi:10.1109/ICPR.2010.72.url:https://ieeexplore.ieee.org/ document/5597587(visited on 02/11/2026)

    work page doi:10.1109/icpr.2010.72.url:https://ieeexplore.ieee.org/ 2010

  34. [35]

    olmocr: Unlocking trillions of tokens in pdfs with vision language models.arXiv preprint arXiv:2502.18443,

    Jake Poznanski et al.olmOCR: Unlocking Trillions of Tokens in PDFs with Vision Language Models. arXiv:2502.18443 [cs]. July 2025.doi:10. 48550/arXiv.2502.18443.url:http://arxiv.org/abs/2502.18443 (visited on 03/27/2026)

    work page arXiv 2025

  35. [36]

    Andrew Prescott and Lorna Hughes.Why Do We Digitize? The Case for Slow Digitization - Archive Journal. en. 2018.url:http : / / www . archivejournal.net/?p=7962(visited on 03/26/2026)

    2018

  36. [37]

    Reading order detection in visually-rich documents with multi-modal layout-aware relation prediction

    Liang Qiao et al. “Reading order detection in visually-rich documents with multi-modal layout-aware relation prediction”. In:Pattern Recognition 150 (June 2024), p. 110314.issn: 0031-3203.doi:10.1016/j.patcog. 2024.110314.url:https://www.sciencedirect.com/science/article/ pii/S0031320324000657(visited on 04/04/2026)

    work page doi:10.1016/j.patcog 2024

  37. [38]

    Reading order detection on hand- written documents

    Lorenzo Quir´ os and Enrique Vidal. “Reading order detection on hand- written documents”. en. In:Neural Computing and Applications34.12 (June 2022), pp. 9593–9611.issn: 1433-3058.doi:10.1007/s00521-022- 06948 - 5.url:https : / / doi . org / 10 . 1007 / s00521 - 022 - 06948 - 5 (visited on 04/04/2026)

    work page doi:10.1007/s00521-022- 2022

  38. [39]

    Computer Vision Intelligence Test Modeling and Gener- ation: A Case Study on Smart OCR

    Jing Shu et al. “Computer Vision Intelligence Test Modeling and Gener- ation: A Case Study on Smart OCR”. In:2024 IEEE International Con- ference on Artificial Intelligence Testing (AITest). arXiv:2410.03536 [cs]. July 2024, pp. 21–28.doi:10 . 1109 / AITest62860 . 2024 . 00011.url: http://arxiv.org/abs/2410.03536(visited on 04/04/2026)

    work page arXiv 2024

  39. [40]

    An Overview of the Tesseract OCR Engine

    R. Smith. “An Overview of the Tesseract OCR Engine”. In:Ninth Interna- tional Conference on Document Analysis and Recognition (ICDAR 2007). Vol. 2. Sept. 2007, pp. 629–633.doi:10 . 1109 / ICDAR . 2007 . 4376991. url:https : / / ieeexplore . ieee . org / document / 4376991(visited on 02/03/2025)

    2007

  40. [41]

    METAe—Automated Encoding of Digitized Texts

    Birgit Stehno, Alexander Egger, and Gregor Retti. “METAe—Automated Encoding of Digitized Texts”. In:Literary and Linguistic Computing18.1 (Apr. 2003), pp. 77–88.issn: 0268-1145.doi:10.1093/llc/18.1.77. url:https://doi.org/10.1093/llc/18.1.77(visited on 02/11/2026). 26

    work page doi:10.1093/llc/18.1.77 2003

  41. [42]

    Ting Sun et al.PP-DocLayout: A Unified Document Layout Detection Model to Accelerate Large-Scale Data Construction. en. Mar. 2025.url: https://arxiv.org/abs/2503.17213v1(visited on 02/04/2026)

    work page arXiv 2025

  42. [43]

    Thai OCR: a neural network ap- plication

    C. Tanprasert and T. Koanantakool. “Thai OCR: a neural network ap- plication”. In:Proceedings of Digital Processing Applications (TENCON ’96). Vol. 1. Nov. 1996, 90–95 vol.1.doi:10.1109/TENCON.1996.608717. url:https://ieeexplore.ieee.org/abstract/document/608717(vis- ited on 03/26/2026)

    work page doi:10.1109/tencon.1996.608717 1996

  43. [44]

    Interpreting textual artefacts: cognitive insights into expert practices

    S. Tarte. “Interpreting textual artefacts: cognitive insights into expert practices”. English. In: (2012).url:https://ora.ox.ac.uk/objects/ uuid:c1c39da4-fa0d-4644-bc80-1672b89d1d95(visited on 03/26/2026)

    2012

  44. [45]

    TEI.TEI: Guidelines for Electronic Text Encoding and Interchange. Apr. 2025.url:https://tei- c.org/release/doc/tei- p5- doc/en/html/ index.html

    2025

  45. [46]

    Melissa Terras.Image to Interpretation: An Intelligent System to Aid His- torians in Reading the Vindolanda Texts. en. Google-Books-ID: wpREAAAQBAJ. OUP Oxford, Oct. 2006.isbn: 978-0-19-152544-5

    2006

  46. [47]

    InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Zilong Wang et al. “LayoutReader: Pre-training of Text and Layout for Reading Order Detection”. In:Proceedings of the 2021 Conference on Em- pirical Methods in Natural Language Processing. Ed. by Marie-Francine Moens et al. Online and Punta Cana, Dominican Republic: Association for Computational Linguistics, Nov. 2021, pp. 4735–4744.doi:10.18653/v1/ 2021....

    work page doi:10.18653/v1/ 2021

  47. [48]

    Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks

    Haoran Wei et al.General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model. arXiv:2409.01704. Sept. 2024.doi:10.48550/arXiv. 2409 . 01704.url:http : / / arxiv . org / abs / 2409 . 01704(visited on 10/14/2024)

    work page internal anchor Pith review doi:10.48550/arxiv 2024

  48. [49]

    2024 , eprint =

    Zhiyuan Zhao et al.DocLayout-YOLO: Enhancing Document Layout Anal- ysis through Diverse Synthetic Data and Global-to-Local Adaptive Percep- tion. arXiv:2410.12628. Oct. 2024.doi:10.48550/arXiv.2410.12628. url:http://arxiv.org/abs/2410.12628(visited on 11/14/2024). 27

    work page doi:10.48550/arxiv.2410.12628 2024