pith. sign in

arxiv: 2606.24984 · v1 · pith:UAYN7DRRnew · submitted 2026-06-23 · 💻 cs.LG · cs.CL· cs.CV

Learning Diachronic Representations of Ancient Greek Letterforms

John Pavlopoulos , Spyros Barbakos , Lavinia Ferretti , Dionysis Voulgarakis , Asimina Paparrigopoulou , Maria Konstantinidou , Giuseppe De Gregorio , Isabelle Marthot-Santaniello
show 2 more authors
Paraskevi Platanou Holger Essler
This is my paper

Pith reviewed 2026-06-26 00:30 UTC · model grok-4.3

classification 💻 cs.LG cs.CLcs.CV
keywords diachronic representation learningancient Greek letterformssupervised contrastive losslacuna-driven augmentationmanuscript degradationstylistic clusteringhistorical handwriting recognition
0
0 comments X

The pith

Similarity-weighted contrastive loss plus lacuna augmentations produces embeddings that separate ancient Greek letter classes across centuries and visualize their evolution.

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

The paper establishes that training neural networks with a contrastive loss biased toward dynamically estimated similarities between letter classes, together with augmentations that insert realistic manuscript gaps, yields embeddings robust to the style shifts and degradations that occur over more than a millennium of Greek handwriting. A sympathetic reader would care because standard image models fail when data are scarce and the visual forms of the same symbol change systematically; respecting the intrinsic relationships among letters and simulating the actual corruptions present in manuscripts supplies the missing inductive bias. The resulting embeddings support downstream tasks such as clustering instances, recovering stylistic subgroups, and generating prototype images that trace how letter shapes transitioned from one period to the next.

Core claim

A similarity-weighted supervised contrastive loss that uses dynamically estimated inter-class similarities, combined with a lacuna-driven augmentation scheme that simulates realistic manuscript corruptions, enables both a lightweight CNN and a pretrained ResNet to achieve strong recognition performance on three new diachronic Greek letter datasets (Hell-Char, PaLit-Char, Med-Char) while producing embeddings that separate character classes more coherently than PCA or generic pretrained models; these embeddings further support clustering, identification of stylistic subgroups, and construction of prototype images that visualize diachronic evolution and transitional letterforms.

What carries the argument

The similarity-weighted supervised contrastive loss, which biases embeddings using dynamically estimated inter-class similarities, together with the lacuna-driven augmentation scheme.

If this is right

  • Embeddings support clustering of letter instances across periods.
  • Embeddings allow identification of stylistic subgroups within a given century range.
  • Prototype images constructed from the embeddings visualize diachronic evolution and transitional letterforms.
  • The same strategies produce robust representations under scarce, temporally evolving, and noisy conditions.

Where Pith is reading between the lines

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

  • The same loss and augmentation design could be tested on other historical scripts that exhibit gradual visual change, such as Latin or Arabic paleography.
  • Prototype images might serve as quantitative references for dating undated manuscripts by measuring distance to period-specific centroids.
  • Improved separation of classes could directly raise accuracy in downstream optical character recognition pipelines for medieval Greek texts.

Load-bearing premise

That dynamically estimated inter-class similarities supply a reliable and stable bias for the contrastive loss and that lacuna-driven augmentations sufficiently capture the distribution of real manuscript degradations in the evaluation sets.

What would settle it

If embeddings trained with the proposed loss and augmentations fail to separate character classes more coherently than PCA on the PaLit-Char or Med-Char sets, or if the generated prototypes do not align with known historical transitions, the central claim is falsified.

Figures

Figures reproduced from arXiv: 2606.24984 by Asimina Paparrigopoulou, Dionysis Voulgarakis, Giuseppe De Gregorio, Holger Essler, Isabelle Marthot-Santaniello, John Pavlopoulos, Lavinia Ferretti, Maria Konstantinidou, Paraskevi Platanou, Spyros Barbakos.

Figure 1
Figure 1. Figure 1: Alpha from Hell-Char (leftmost) with rectangular (RE, 2nd) and lacuna [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Letter frequency in our Hell-Char subset. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative forms of letter Alpha using cluster medoids (§3.4). [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Boxplot of years per letter for misclassified out-of-distribution images of [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE plot of ResNet18+lf+dscl embeddings on Med-Char. Points are coloured by century (blue for older, red for more recent). One prototype image per letter-century group is shown, selected as the sample closest to the group centroid, to visualize graphemic clustering and local diachronic variation [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
read the original abstract

Learning representations that remain robust across centuries of variation in handwriting is a key challenge in diachronic representation learning. Taking one of the longest continuously used writing systems, ancient Greek, as a case study, we introduce three datasets for diachronic representation learning: Hell-Char, a curated training set spanning the 3rd-1st centuries BCE, and two evaluation sets, PaLit-Char (2nd-5th c. CE) and Med-Char (9th-14th c. CE). To address the challenges of symbolic variation, scarce data, and systematic degradation, we propose: a similarity-weighted supervised contrastive loss that biases embeddings using dynamically estimated inter-class similarities, and a lacuna-driven augmentation scheme that simulates realistic manuscript corruptions. Trained with these strategies, both a lightweight CNN and a pretrained ResNet achieve strong recognition performance and produce embeddings that more coherently separate character classes than PCA or generic pretrained models. These embeddings enable clustering, identification of stylistic subgroups, and construction of prototype images that visualize diachronic evolution and transitional letterforms. Our results demonstrate that respecting intrinsic inter-letter relationships and augmenting with domain-informed corruptions yield robust, interpretable representations, offering a transferable paradigm for representation learning under scarce, temporally evolving, and noisy conditions. Code and data available at: https://github.com/ipavlopoulos/diachronic-greek-letterforms.

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

1 major / 2 minor

Summary. The manuscript introduces three datasets (Hell-Char for training spanning 3rd-1st c. BCE, PaLit-Char and Med-Char for evaluation in later centuries) for diachronic representation learning of ancient Greek letterforms. It proposes a similarity-weighted supervised contrastive loss that biases embeddings via dynamically estimated inter-class similarities, together with a lacuna-driven augmentation scheme simulating manuscript corruptions. The central claim is that lightweight CNN and pretrained ResNet models trained under these strategies achieve strong recognition performance, yield embeddings that separate character classes more coherently than PCA or generic pretrained models, and support downstream tasks including clustering, stylistic subgroup identification, and prototype-image visualization of diachronic evolution.

Significance. If the empirical results hold, the work supplies a concrete, transferable paradigm for representation learning under scarce, temporally evolving, and noisy conditions, with direct relevance to digital humanities and historical document analysis. The public release of code and data is a clear strength that supports reproducibility.

major comments (1)
  1. Abstract: the claims of 'strong recognition performance' and 'more coherently separate character classes than PCA or generic pretrained models' are load-bearing for the central contribution, yet the abstract supplies no quantitative metrics, baseline comparisons, ablation results, or error analysis; without these, the magnitude and robustness of the reported gains cannot be assessed.
minor comments (2)
  1. Methods: the dynamic estimation of inter-class similarities used inside the contrastive loss requires an explicit stability analysis or sensitivity check across random seeds and training epochs.
  2. Experiments: the lacuna-driven augmentation scheme should be accompanied by a quantitative comparison showing how closely the simulated corruptions match the degradation statistics observed in the held-out PaLit-Char and Med-Char sets.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We address the single major comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [—] Abstract: the claims of 'strong recognition performance' and 'more coherently separate character classes than PCA or generic pretrained models' are load-bearing for the central contribution, yet the abstract supplies no quantitative metrics, baseline comparisons, ablation results, or error analysis; without these, the magnitude and robustness of the reported gains cannot be assessed.

    Authors: We agree that the abstract would be strengthened by including concrete quantitative support for the central claims. The full manuscript reports recognition accuracies, embedding separation metrics (e.g., via silhouette scores or nearest-neighbor classification), and comparisons against PCA and generic pretrained models in Sections 4 and 5, along with ablations of the proposed loss and augmentation. In the revised version we will condense the key figures (e.g., top-1 accuracy on Hell-Char, PaLit-Char, and Med-Char, plus a brief statement on embedding coherence) into the abstract while preserving its length constraints. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

This is empirical ML work introducing new datasets (Hell-Char, PaLit-Char, Med-Char) with held-out evaluation sets, a contrastive loss, and domain-specific augmentations. The central claims concern recognition performance and embedding quality on those sets; no derivation, prediction, or uniqueness theorem is asserted that reduces by construction to fitted parameters or self-citations. The abstract and described approach contain no load-bearing self-referential steps of the enumerated kinds. Standard empirical pipeline with external validation sets yields a self-contained result.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only view provides no explicit free parameters, axioms or invented entities; the central claim rests on the unverified effectiveness of the dynamically estimated similarities and the realism of the lacuna augmentations.

pith-pipeline@v0.9.1-grok · 5829 in / 1049 out tokens · 33949 ms · 2026-06-26T00:30:58.370389+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

27 extracted references · 11 canonical work pages · 1 internal anchor

  1. [1]

    In: Bausi, A., Borbone, P.G., Briquel- Chatonnet, F., Buzi, P., Gippert, J., Macé, C., Maniaci, M., Melissakis, Z., Parodi, L.E., Witakowski, W

    Bianconi, D.: Greek Palaeography. In: Bausi, A., Borbone, P.G., Briquel- Chatonnet, F., Buzi, P., Gippert, J., Macé, C., Maniaci, M., Melissakis, Z., Parodi, L.E., Witakowski, W. (eds.) Comparative Oriental Manuscript Studies. An Introduction, pp. 297–305. Trendition, Hamburg (2015)

    2015

  2. [2]

    Digital Signal Processing 149, 104477 (2024)

    Boudraa, M., Bennour, A., Al-Sarem, M., Ghabban, F., Bakhsh, O.A.: Con- tribution to historical manuscript dating: A hybrid approach employing hand-crafted features with vision transformers. Digital Signal Processing 149, 104477 (2024). https://doi.org/10.1016/j.dsp.2024.104477

    work page doi:10.1016/j.dsp.2024.104477 2024

  3. [3]

    In: ICCV

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 4778–4788 (2021)

    2021

  4. [4]

    In: Bagnall, R.S

    Cavallo, G.: Greek and Latin Writing in the Papyri. In: Bagnall, R.S. (ed.) TheOxfordHandbookofPapyrology,pp.101–148.OxfordUniversityPress, Oxford, New York (2009)

    2009

  5. [5]

    A Simple Framework for Contrastive Learning of Visual Representations

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: Proceedings of ICML (2020). https://doi.org/10.48550/arXiv.2002.05709

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2002.05709 2020

  6. [6]

    Una introduzione

    Crisci, E., Degni, P.: La scrittura greca dall’antichità all’epoca della stampa. Una introduzione. Carocci, Roma (2011)

    2011

  7. [7]

    ACM Journal on Computing and Cultural Heritage18(3), 1–19 (2025)

    D’Alessandro, T., Cilia, N., De Stefano, C., Fontanella, F., Molinara, M., Scotto di Freca, A., Marthot-Santaniello, I.: A deep transfer learning ap- proach for writer identification in greek papyri. ACM Journal on Computing and Cultural Heritage18(3), 1–19 (2025)

    2025

  8. [8]

    EGRAPSA Hellenistic Dated Papyri Dataset (2025)

    Ferretti, L., Serbaeva Saraogi, O., De Gregorio, G., Marthot-Santaniello, I.: Hell-Date. EGRAPSA Hellenistic Dated Papyri Dataset (2025). https: //doi.org/10.5281/zenodo.15083590, https://zenodo.org/records/15083590

    work page doi:10.5281/zenodo.15083590 2025

  9. [9]

    In: CVPR

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recog- nition. In: CVPR. pp. 770–778 (2016)

    2016

  10. [10]

    https://doi.org/10.5281/zenodo.1194357 (2016)

    He, S., Schomaker, L.R., Samara, P., Burgers, J.: Mps data set with im- ages of medieval charters for handwriting-style based dating of manuscripts. https://doi.org/10.5281/zenodo.1194357 (2016)

    work page doi:10.5281/zenodo.1194357 2016

  11. [11]

    E., & Salakhutdinov, R

    Hinton, G.E., Salakhutdinov, R.R.: Reducing the dimensionality of data with neural networks. Science313(5786), 504–507 (2006). https://doi.org/ 10.1126/science.1127647

    work page doi:10.1126/science.1127647 2006

  12. [12]

    In: Sirat, C., Irigoin, J., Poulle, E

    Irigoin, J.: L’alphabet grec et son geste des origines au IXe siècle après J.-C. In: Sirat, C., Irigoin, J., Poulle, E. (eds.) L’écriture: le cerveau, l’œil et la main, pp. 299–305. Brepols, Turnhout (1990)

    1990

  13. [13]

    In: NeurIPS

    Khosla,P.,Teterwak,P.,Wang,C.,Sarna,A.,Tian,Y.,Isola,P.,Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: NeurIPS. pp. 18661–18673 (2020), proceedings.neurips.cc Diachronic Greek Letterforms 17

    2020

  14. [14]

    Lecun, L

    LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of IEEE86(11), 2278–2324 (1998). https://doi.org/10.1109/5.726791

    work page doi:10.1109/5.726791 1998

  15. [15]

    In: The 3rd International Workshop on Historical Document Imaging and Process- ing

    Li, Y., Genzel, D., Fujii, Y., Popat, A.C.: Publication date estimation for printed historical documents using convolutional neural networks. In: The 3rd International Workshop on Historical Document Imaging and Process- ing. pp. 99–106. ACM (2015)

    2015

  16. [16]

    In: Coustaty, M., Fornés, A

    Marthot-Santaniello, I., Vu, M.T., Serbaeva, O., Beurton-Aimar, M.: Stylis- tic similarities in greek papyri based on letter shapes: A deep learning ap- proach. In: Coustaty, M., Fornés, A. (eds.) ICDAR Workshops. pp. 307–323. Springer Nature Switzerland, Cham (2023)

    2023

  17. [17]

    IEEE Transactions on Systems, Man, and Cybernetics9(1), 62–66 (1979), https: //doi.org/10.1109/TSMC.1979.4310076

    Otsu, N.: A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics9(1), 62–66 (1979). https: //doi.org/10.1109/TSMC.1979.4310076

    work page doi:10.1109/tsmc.1979.4310076 1979

  18. [18]

    Machine Learning113(9), 6765–6786 (2024)

    Pavlopoulos, J., Konstantinidou, M., Perdiki, E., Marthot-Santaniello, I., Essler, H., Vardakas, G., Likas, A.: Explainable dating of greek papyri im- ages. Machine Learning113(9), 6765–6786 (2024)

    2024

  19. [19]

    on lines and planes of closest fit to systems of points in space (Nov 1901)

    Pearson, K.: On lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science2(11), 559–572 (1901). https://doi.org/10.1080/14786440109462720

    work page doi:10.1080/14786440109462720 1901

  20. [20]

    Journal of computational and applied mathemat- ics20, 53–65 (1987)

    Rousseeuw, P.J.: Silhouettes: a graphical aid to the interpretation and vali- dation of cluster analysis. Journal of computational and applied mathemat- ics20, 53–65 (1987)

    1987

  21. [21]

    ACM SIGKDD Explorations Newsletter 25(1), 36–42 (2023)

    Schubert, E.: Stop using the elbow criterion for k-means and how to choose the number of clusters instead. ACM SIGKDD Explorations Newsletter 25(1), 36–42 (2023)

    2023

  22. [22]

    Pattern Recognition29(4), 641–662 (1996)

    Trier, Ø.D., Jain, A.K., Taxt, T.: Feature extraction methods for character recognition – a survey. Pattern Recognition29(4), 641–662 (1996). https: //doi.org/10.1016/0031-3203(95)00118-2

    work page doi:10.1016/0031-3203(95)00118-2 1996

  23. [23]

    Academy of Management Perspectives , volume =

    Wahlberg, F., Wilkinson, T., Brun, A.: Historical manuscript production date estimation using deep convolutional neural networks. In: ICFHR. pp. 205–210 (2016). https://doi.org/10.1109/ICFHR.2016.0048

    work page doi:10.1109/icfhr.2016.0048 2016

  24. [24]

    In: Pavlopoulos, John, e.a

    West, G., Swindall, M.I., Brusuelas, J.H., Maltomini, F., Gerhardt, M., D’Angelo, M., Wallin, J.F.: A deep learning pipeline for the palaeographi- cal dating of ancient Greek papyrus fragments. In: Pavlopoulos, John, e.a. (ed.) ML4AL ACL Workshop. pp. 177–185. ACL, Hybrid in Bangkok, Thai- land and online (Aug 2024). https://doi.org/10.18653/v1/2024.ml4al...

    work page doi:10.18653/v1/2024.ml4al-1.18 2024

  25. [25]

    In: Proceedings of CVPR

    Woo, S., Debnath, S., Hu, R., Chen, X., Liu, Z., Kweon, I.S., Xie, S.: Con- vnext v2: Co-designing and scaling convnets with masked autoencoders. In: Proceedings of CVPR. pp. 16133–16142 (2023)

    2023

  26. [26]

    In: ICCV

    Zheng, M., Wang, F., You, S., Qian, C., Zhang, C., Wang, X., Xu, C.: Weakly supervised contrastive learning. In: ICCV. pp. 4009–4015 (2021), arXiv:2110.04770

    arXiv 2021

  27. [27]

    In: AAAI

    Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: Random erasing data augmentation. In: AAAI. vol. 34, pp. 13001–13008 (2020)

    2020