Pith. sign in

REVIEW 2 major objections 8 minor 29 references

Unsupervised adaptation turns Latin and Greek language models into sentence encoders that beat every baseline on biblical text-reuse tasks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 12:02 UTC pith:EGBINOPF

load-bearing objection Solid applied win for historical NLP: unsupervised corpus adaptation beats strong baselines on a real multi-author reuse benchmark, with one load-bearing caveat about incomplete gold. the 2 major comments →

arxiv 2607.24542 v1 pith:EGBINOPF submitted 2026-07-27 cs.CL cs.DLcs.IR

From transcription to semantic corpus analysis: unsupervised learning of sentence representations for ancient languages

Th{\'e}otime de la Selle (ISC , HiSoMA , CNRS) This is my paper
classification cs.CL cs.DLcs.IR
keywords sentence embeddingsunsupervised representation learningancient languagesLatinAncient Greektext reuse detectioninformation retrievaldigital humanities
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Automatic transcription now yields large, messy corpora in ancient languages, but the sentence embeddings needed for semantic search and reuse detection transfer poorly from modern tools. This paper shows that two fully unsupervised methods—TSDAE and contrastive sentence embedding—can reshape specialized Latin and Greek token models into corpus-specific sentence encoders using only raw sentences. On 2,935 expert-verified biblical parallels in Augustine, Jerome, and Athanasius, the adapted encoders outperform multilingual, pooled specialized, distilled, and supervised fine-tuned baselines at both detecting reuse and retrieving the matching verse. Contrastive adaptation reaches strong retrieval with only 4–8k in-domain sentences and tens of seconds on a laptop GPU, transfers across works and authors, and improves further when retrained directly on simulated post-transcription noise. The result is a practical, annotation-free path from raw transcription to cross-corpus semantic analysis, released as an online tool.

Core claim

Fully unsupervised TSDAE and contrastive sentence embedding (CSE) adaptation of specialized Latin and Ancient Greek token-level language models into corpus-specific sentence encoders outperforms multilingual, pooled specialized, distilled, and supervised fine-tuned baselines on both binary reuse detection and correspondence retrieval over 2,935 expert-verified biblical parallels. The two strategies are complementary: TSDAE leads detection when a large in-domain corpus is available, while CSE leads retrieval, saturates at 4–8k raw sentences, and transfers across works, authors, and noisy post-ATR text when retrained on it.

What carries the argument

Two unsupervised objectives that convert a pretrained token encoder into a sequence encoder from raw sentences alone: TSDAE (a sequential denoising auto-encoder whose information bottleneck is the sentence vector) and CSE (contrastive learning with dropout-generated positive pairs). They reshape anisotropic token spaces so cosine similarity becomes discriminative for detection and retrieval.

Load-bearing premise

The expert gold indexes of biblical reuses are complete enough that model rankings on precision-style and Hits@k metrics can be trusted, even though the paper finds nearly half of the top-scoring false positives on one major work are real but unannotated reuses.

What would settle it

Have experts complete the De Trinitate gold index by adjudicating the highest-STS false positives, then re-evaluate every model on detection and Hits@k; if the adapted encoders no longer lead or the order among them flips, the central performance claim does not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A newly transcribed Latin or Greek work can receive a corpus-specific sentence encoder from a few thousand raw sentences and no similarity labels.
  • Reuse identification should be scored as two tasks—binary detection and verse retrieval—because the best model systematically differs by task.
  • CSE trained directly on imperfect ATR output is a practical remediation step, not only a clean-text method.
  • UMAP atlases of the adapted spaces double as exploratory maps of thematic clusters linking patristic sentences to biblical books.
  • The full pipeline—segmentation, unsupervised fine-tuning, and cross-corpus search—is usable by non-specialists through the released online tool.

Where Pith is reading between the lines

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

  • Because many top false positives are genuine missing annotations, the same encoders can serve as discovery instruments that enlarge the gold indexes used to evaluate them.
  • CSE’s few-thousand-sentence regime makes per-work or even per-manuscript adaptation realistic for large ATR archives with local orthographic shift.
  • Separating detection from retrieval likely matters for other dense intertextual settings—classical allusions, legal borrowing—where source corpora overlap and labels are incomplete.
  • TSDAE-style reconstruction adaptation for Greek becomes feasible once larger coherent patristic sentence collections exist.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 8 minor

Summary. The manuscript adapts token-level masked LMs for Latin (LaBerta, Latin BERT) and Ancient Greek (Logion, GreBerta) into sentence encoders using two fully unsupervised objectives — TSDAE and SimCSE-style contrastive training (CSE) — and evaluates them on biblical reuse identification in patristic literature, decomposed into binary detection (AP, AUC-ROC, F1max on max-similarity scores) and correspondence retrieval (Hits@k against a verse index). The benchmark comprises 2,935 expert-verified parallels across four works (Augustine's De Trinitate and Contra Faustum, Jerome's Adversus Iovinianum, Athanasius' De decretis), with comparisons against multilingual encoders (LaBSE, mE5), pooled specialized LMs, a distilled encoder (SPhilBerta), supervised fine-tuned variants from Schelb et al., and a grid-searched Passim baseline, plus corpus-size ablations (App. B) and simulated HTR/abbreviation noise regimes (App. A). The adapted encoders lead on both tasks with complementary profiles (TSDAE for detection, CSE for retrieval), CSE saturates at 4–8k in-domain sentences, transfers across works and authors, and noise-aware retraining recovers much of the degradation. The detection/retrieval dissociation is argued to be a methodological point of general relevance. Code, data, a released model (PatriLaSE), and an online tool (Paraphrasis) are provided.

Significance. If the results hold, this is a useful and timely contribution at the NLP/digital-humanities boundary. It ships the largest quantitative evaluation of ancient-language text reuse to date (2,935 expert-verified parallels in two languages, versus ≤550 in prior quantitative work, as the paper documents in §2), an argued and empirically supported task decomposition showing that detection and retrieval rank models differently, and a practically important data-efficiency result: CSE adaptation at 4–8k raw sentences in tens of seconds on a laptop GPU makes per-corpus adaptation realistic for exactly the post-ATR setting the paper targets. The evaluation is genuinely external (expert gold labels never used in training), the comparative claims are falsifiable, and the pipeline is reproducible in principle: public repository, released encoder, and a working tool for non-specialists. The noise-robustness experiments, while based on synthetic proxies, are a step toward honest evaluation of post-HTR conditions that the field generally skips.

major comments (2)
  1. [§8 Limitations; Tables 3–5; Eqs. (3) and (6)] The headline claim is comparative, yet the metrics are computed against a gold index the paper itself shows to be substantially incomplete: §8 reports that nearly 50% of the 100 highest-STS false positives on De Trinitate are genuine but unannotated reuses. This bias is not a uniform offset across models. On detection (AP, Eq. 3), a better retriever places more true-but-unannotated reuses at the top of the ranking, where each counts as a false positive, so AP is depressed most for exactly the strongest systems; the AP gap between LaBerta-TSDAE (0.81) and LaBerta-CSE_trin (0.80) is well inside the plausible size of this differential effect. On retrieval (Hits@k, Eq. 6), unannotated genuine correspondences — including intra-biblical parallels the paper itself flags in §5.1 — can outrank the gold verse and count as misses, again non-uniformly. The sharp contrast between De Trinitate (~50% o
  2. [§6.2–6.3, Tables 4, 6, 9, 10; App. B] All results appear to be single training runs, yet several load-bearing comparisons rest on differences of 1–3 points: the transfer ordering LaBerta-CSE_cfaust (0.60) ≈ CSE_trin (0.60) > CSE_hier (0.59) in Table 4, from which the 'graceful, near-monotonic degradation with domain distance' conclusion is drawn; the retrieval saturation claim across 4k–9k sentences in Table 9 (Hits@1 0.603–0.611) and the CSE_Aug degradation claim (0.596 vs 0.611); and the Greek comparison in Table 6 where N=106 makes one point ≈ one query (the text acknowledges this only in passing). Since a CSE run costs tens of seconds (§7.2 fn. 11), reporting mean ± std over ≥5 seeds for the CSE variants — and at minimum bootstrap confidence intervals on Hits@k and AP for all models — is cheap and would either firm up or appropriately temper these claims. Relatedly, the phrase 'significantly higher/lower' is used through
minor comments (8)
  1. [§3.4, §4] Reproducibility gaps in the configuration: the SimCSE temperature τ is never given a value in §3.4 despite being named in Eq. (2), and the sentence-segmentation min/max word thresholds of §4 are described but never quantified. Both should be stated (or pointed to in the repo config).
  2. [§6.1, Tables 3–5] LaBerta-CSE_trin is trained (unsupervised) on sentences from De Trinitate and evaluated on the same work; similarly CSE_cfaust/CSE_hier. Labels are unused so this is not leakage, but the in-work condition should be explicitly flagged in the table captions or text of §6.1–6.2, since the transfer rows are the clean evidence and the in-work rows otherwise invite misreading.
  3. [Appendix C, Table 11] The Passim grid search (App. C) selects the configuration maximizing Hits@1 on the De Trinitate test set itself, i.e., the baseline is tuned on test data. This inflates the baseline and is therefore conservative for the paper's claim, but it should be acknowledged explicitly as such.
  4. [§4.1, Appendix A, Table 7] The artificial-noise regimes (App. A) rest on the assumption that 5%-CER character confusions and stochastic abbreviation classes are informative proxies for real post-ATR text. Please justify the 5% CER target against measured CERs of actual HTR systems on medieval Latin (e.g., CATMuS/eScriptorium benchmarks the authors cite) and state the sensitivity of the Table 7 conclusions to the CER level.
  5. [§7.3, Figures 3–4] The UMAP discussion slides from visualization to causal explanation ('thereby elucidating the retrieval enhancement', 'which accounts for the difficulty'). A 2D projection cannot establish why retrieval improves; please temper this language or support it with a quantitative correlate (e.g., mean cosine distance between reuse pairs across models).
  6. [§2] Novelty claims ('first time TSDAE is applied on ancient languages models, and the first time that SimCSE is applied on Latin models', §2) should be hedged to 'to our knowledge' consistently — the first clause has the hedge, the second does not — and the relation to Krahn et al. 2023 (SimCSE compared for Greek) stated in one sentence.
  7. [Abstract, §6.2, References, Figure 2, Tables 3–10] Copyediting: abstract has run-together text ('analysestext'); 'hieronimian' (§6.2) should be 'Hieronymian'; the Caffagni et al. 2025b reference contains a leftover editorial note ('To verify: LNCS volume/pages...'); footnotes render as '1Corpus', '2Textual', '3https'; 'NaR' in Figure 2 is never expanded; bolding of best scores is inconsistent across tables (e.g., Table 3 F1max vs Table 4); 'finetuned'/'fine-tuned' used interchangeably.
  8. [§4 Table 1, §5.3, §8] Table 1 reports 570 reuses for Adversus Iovinianum while §8 mentions 269 allusions (44 in multi-reuse sentences); a brief note on how quotations vs allusions compose the 570 would help readers interpret the detection/retrieval split. Also state in §5.3 whether sentences with multiple gold verses are counted once or multiple times in Hits@k.

Circularity Check

0 steps flagged

No circularity: unsupervised adaptation trained only on raw sentences; claims are empirical benchmark results against external expert gold, not derivations that reduce to inputs.

full rationale

This is an empirical NLP/digital-humanities paper. The load-bearing claims are comparative performance numbers (AP, AUC-ROC, F1 max, Hits@k) of TSDAE- and CSE-adapted encoders versus multilingual, pooled specialized, distilled, and supervised baselines on expert-verified biblical-reuse gold (2,935 parallels). Training objectives (TSDAE reconstruction Eq. 1; CSE InfoNCE Eq. 2) consume only unlabeled raw sentences; gold reuse labels are used solely at evaluation and are not inputs to the adapters. In-work CSE variants (e.g. LaBerta-CSE_trin on De Trinitate) adapt to domain text from the evaluated work but do not fit the detection/retrieval targets; cross-work and cross-author transfers (Tables 4–5) and Greek replication (Table 6) further separate training data from the measured ordering. UMAP atlases are post-hoc geometric illustrations, not predictive derivations. Self-citations (e.g. de la Selle & Mellerin 2026) motivate the task and do not underwrite uniqueness or force the results. Gold incompleteness (§8) is an evaluation-validity concern, not circularity of the derivation chain. No self-definitional step, fitted-input-as-prediction, uniqueness import, or ansatz smuggling is present.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

Load-bearing commitments are standard ML/IR practice plus domain choices about segmentation, similarity, and noise simulation. No new physical entities. Free parameters are ordinary training hyper-parameters set to literature defaults or small grids, plus the artificial CER and abbreviation rates used only in robustness tables—not fitted to maximize the main clean-text claim.

free parameters (4)
  • CSE learning rate, batch size, temperature τ, weight decay, warmup = LR=1e-5, batch=32, 1 epoch, mean pooling
    Set to Gao et al. defaults / paper choices (LR 1e-5, batch 32, etc.); not swept exhaustively (§3.4, §8).
  • TSDAE deletion ratio, LR, batch size, CLS pooling = deletion=0.6, LR=2e-5, batch=16, 1 epoch
    Deletion ratio 0.6 is Wang et al. recommended default; other knobs fixed without full sweep.
  • Artificial HTR target CER and abbreviation class probabilities = CER≈5%; abbr. token rate ~25–35%
    Hand-set noise model (CER≈5%; class probabilities yielding ~25–35% abbreviated tokens) for robustness tables only (App. A).
  • Sentence segmentation min/max word thresholds
    Chosen to avoid unstable short embeddings and long-sentence dilution (§4); exact thresholds affect unit of evaluation.
axioms (6)
  • domain assumption Cosine similarity (inner product of L2-normalized embeddings) is an adequate STS measure for ranking reuse candidates.
    Used throughout §5–6 via FAISS IndexFlatIP; standard in dense retrieval but not proved optimal for allusive patristic reuse.
  • domain assumption Max similarity of a query sentence to any biblical verse is a sufficient score for both binary detection and correspondence retrieval.
    Task definition in §5.1; one-to-many and multi-reuse sentences are acknowledged but collapsed to max.
  • domain assumption Expert BiblIndex/TEI annotations are a valid (if incomplete) gold standard for biblical quotation and allusion.
    Evaluation backbone §4–6; incompleteness admitted in §8.
  • domain assumption InfoNCE with dropout positives (SimCSE) and TSDAE reconstruction improve alignment/uniformity of anisotropic LM spaces.
    Imported from Gao et al. 2021, Wang et al. 2021, Wang & Isola 2020; §2–3.
  • ad hoc to paper Artificial character-level HTR confusions and stochastic scribal abbreviations are informative proxies for post-ATR robustness.
    §4.1 and App. A; not calibrated on a real paired HTR gold corpus of the same works.
  • standard math Standard Transformer encoder/decoder training math (cross-entropy reconstruction; contrastive InfoNCE).
    Eqs. (1)–(2) in §3.
invented entities (1)
  • Paraphrasis online tool / PatriLaSE encoder release independent evidence
    purpose: Package segmentation, fine-tuning, and cross-corpus search for non-specialists; release the large-corpus Latin TSDAE model.
    Engineering artifacts, not theoretical entities; independent_evidence true via public URLs claimed in the paper.

pith-pipeline@v1.2.0-grok45-kimik3 · 27328 in / 3926 out tokens · 88893 ms · 2026-07-31T12:02:55.360371+00:00 · methodology

0 comments
read the original abstract

Automatic Text Recognition (ATR) now supplies digital humanities with large volumes of unstructured, heterogeneous, and often noisy text in ancient languages. Downstream semantic analysestext reuse identification, alignment, and semantic search-rely on sentence embeddings, yet existing methods transfer poorly to ancient languages: generic multilingual encoders underperform, specialized language models yield anisotropic representation spaces, and labeled similarity data is unavailable. We study two fully unsupervised strategies - TSDAE and contrastive sentence embedding (CSE) - that adapt a specialized token-level language model into a corpus-specific sentence encoder using only raw sentences. On the philologically central case of biblical reuse in patristic literature (2,935 expert-verified parallels in Latin and Ancient Greek, from Augustine, Jerome, and Athanasius), we decompose reuse identification into two separately evaluated tasks-binary detection and correspondence retrieval-and benchmark the adapted encoders against multilingual, specialized, distilled, and supervised fine-tuned baselines, as well as on artificially noised data simulating HTR artifacts and scribal abbreviations. The adapted encoders outperform all baselines on both tasks, with complementary profiles: TSDAE leads detection given a large in-domain corpus, while CSE leads retrieval, reaches its optimum with as few as 4-8k raw in-domain sentences-a few tens of seconds of training on a laptop GPU-and transfers across works and authors, including to noisy post-ATR text when retrained directly on it. UMAP atlases relate the geometric effect of each strategy to the measured gains, and the full pipeline-segmentation, fine-tuning, cross-corpus semantic search-is made available to non-specialists through the online tool Paraphrasis.

Figures

Figures reproduced from arXiv: 2607.24542 by CNRS), HiSoMA, Th{\'e}otime de la Selle (ISC.

Figure 1
Figure 1. Figure 1: Architectures of the sequence-encoding models built on top of a pretrained token-level encoder. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic representation of text reuse identification decomposed in two algorithmic tasks: [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: UMAP projection of biblical verses and De Trinitate sentences representations from LaBerta encoder. On left and right figures, each point corresponds to a sentence or a verse representation projected on the 2D UMAP space. The figure on the right is similar to the left one with gold reuse data information superposed: De Trinitate sentences containing biblical reuses are indicated by red dark dots combined w… view at source ↗
Figure 4
Figure 4. Figure 4: UMAP projection of biblical verses and De Trinitate sentences representations, from LaBerta-TSDAE en￾coder (left) and LaBerta-CSE encoder (right). Each point corresponds to a sentence or a verse representation projected on the 2D UMAP space. De Trinitate sentences containing biblical reuses are indicated by red dark dots combined with gray lines that point the respective linked biblical verses. 8 Limitatio… view at source ↗

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

29 extracted references · 4 canonical work pages

  1. [3]

    D. Cer, M. Diab, E. Agirre, I. Lopez-Gazpio, and L. Specia. SemEval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation. InProceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 1–14,

    2017

  2. [7]

    URLhttps://aclanthology.org/2025

    CEUR-WS.org. URLhttps://aclanthology.org/2025. clicit-1.33/. T. de la Selle and L. Mellerin. Detection and typology of psalmic text reuses in the new testament.Religions, 17(1),

    2025

  3. [12]

    doi: 10.48550/arXiv.2203.05794. S. Gururangan, A. Marasovi´c, S. Swayamdipta, K. Lo, I. Beltagy, D. Downey, and N. A. Smith. Don’t Stop Pretrain- ing: Adapt Language Models to Domains and Tasks, May

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.05794

  4. [14]

    org/abs/1702.08734

    URLhttp://arxiv. org/abs/1702.08734. arXiv:1702.08734 [cs.CV]. V . Karpukhin, B. O˘guz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W.-t. Yih. Dense passage retrieval for open- domain question answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769–6781,

    Pith/arXiv arXiv 2020

  5. [15]

    doi: 10.18653/v1/2020.emnlp-main.550. A. R. Kashyap, T.-T. Nguyen, V . Schlegel, S. Winkler, S.-K. Ng, and S. Poria. A Comprehensive Survey of Sentence Representations: From the BERT Epoch to the ChatGPT Era and Beyond, Feb

    doi:10.18653/v1/2020.emnlp-main.550 2020

  6. [16]

    arXiv:2305.12641 [cs]

    URLhttp://arxiv.org/ abs/2305.12641. arXiv:2305.12641 [cs]. B. Kiessling, R. Tissot, P. Stokes, and D. Stökl Ben Ezra. eScriptorium: An open source platform for historical docu- ment analysis. InProceedings of the International Conference on Document Analysis and Recognition Workshops (ICDARW),

    Pith/arXiv arXiv

  7. [17]

    URLhttps://aclanthology.org/2023.alp-1.2/

    INCOMA Ltd., Shoumen, Bulgaria. URLhttps://aclanthology.org/2023.alp-1.2/. P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. InAdvances in Neural Information Processing Systems 33 (NeurIPS), pages 9459–9474,

    2023

  8. [18]

    B. Li, H. Zhou, J. He, M. Wang, Y . Yang, and L. Li. On the sentence embeddings from pre-trained language models. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9119–9130,

    2020

  9. [19]

    doi: 10.1093/llc/fqac089

    ISSN 2055-7671, 2055-768X. doi: 10.1093/llc/fqac089. URLhttps://academic.oup.com/dsh/article/38/2/621/6965034. E. Manjavacas, B. Long, and M. Kestemont. On the feasibility of automated detection of allusive text reuse. InPro- ceedings of the 3rd Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Liter...

    work page doi:10.1093/llc/fqac089 2055

  10. [20]

    doi: 10.1007/s11431-020-1647-3. A. Rahlfs and R. Hanhart, editors.Septuaginta: Id est Vetus Testamentum graece iuxta LXX interpretes. Deutsche Bibelgesellschaft, Stuttgart, editio altera edition,

    doi:10.1007/s11431-020-1647-3

  11. [22]

    doi: 10.16995/dscn.9669

    ISSN 1918-3666. doi: 10.16995/dscn.9669. URL https://www.digitalstudies.org/article/id/9669/. N. Reimers and I. Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,

    work page doi:10.16995/dscn.9669 1918

  12. [23]

    Reimers and I

    N. Reimers and I. Gurevych. Making monolingual sentence embeddings multilingual using knowledge distillation. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4512–4525,

    2020

  13. [24]

    Riemenschneider and A

    F. Riemenschneider and A. Frank. Exploring large language models for classical philology. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 15181–15201, 2023a. Introduces the LaBerta and GreBerta models. 18 From Transcription to Semantic Corpus Analysis: Unsupervised Learning of Sentence Representations...

    2023

  14. [25]

    arXiv:2601.07533 [cs.IR]

    URLhttp://arxiv.org/abs/2601.07533. arXiv:2601.07533 [cs.IR]. P. Singh, G. Rutten, and E. Lefever. A pilot study for bert language modelling and morphological analysis for ancient and medieval greek. InThe 5th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2021),

    Pith/arXiv arXiv 2021

  15. [26]

    van den Oord, Y

    A. van den Oord, Y . Li, and O. Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748,

    Pith/arXiv arXiv

  16. [27]

    K. Wang, N. Reimers, and I. Gurevych. TSDAE: Using transformer-based sequential denoising auto-encoder for unsupervised sentence embedding learning. InFindings of the Association for Computational Linguistics: EMNLP 2021, pages 671–688,

    2021

  17. [28]

    L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, and F. Wei. Multilingual E5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672,

    Pith/arXiv arXiv

  18. [29]

    doi: 10.1109/ACCESS.2025.3556899

    ISSN 2169-3536. doi: 10.1109/ACCESS.2025.3556899. URL https://ieeexplore.ieee.org/document/10946892/. 19

    arXiv 2025

  19. [2006]

    Revised edition of Rahlfs’ 1935 Handausgabe, corrected by Robert Hanhart

    ISBN 978-3-438-05119-6. Revised edition of Rahlfs’ 1935 Handausgabe, corrected by Robert Hanhart. M. Reboul. Alignement sémantique et manque de données : l’apport des modèles de langue. Le cas du latin et du grec.Digital Studies / Le champ numérique, 13(3), Apr

    1935

  20. [2014]

    Caffagni, F

    D. Caffagni, F. Cocchi, A. Mambelli, F. Tutrone, M. Zanella, M. Cornia, and R. Cucchiara. Benchmarking BERT- based Models for Latin: A Case Study on Biblical References in Ancient Christian Literature. InProceedings of the 21st Conference on Information and Research Science Connecting to Digital and Library Science (IRCDL 2025), volume 3937 ofCEUR Worksho...

    2025

  21. [2017]

    Clérice, A

    T. Clérice, A. Pinche, M. Vlachou-Efstathiou, A. Chagué, J.-B. Camps, M. Gille Levenson, O. Brisville-Fertin, F. Boschetti, F. Fischer, M. Gervers, A. Boutreux, A. Manton, S. Gabay, P. O’Connor, W. Haverals, M. Kestemont, C. Vandyck, and B. Kiessling. CATMuS medieval: A multilingual large-scale cross-century dataset in Latin script for handwritten text re...

    2024

  22. [2019]

    Ethayarajh

    K. Ethayarajh. How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Pro- cessing (EMNLP-IJCNLP), pages 55–65,

    2019

  23. [2020]

    arXiv:2004.10964 [cs]

    URLhttp://arxiv.org/abs/2004.10964. arXiv:2004.10964 [cs]. 17 From Transcription to Semantic Corpus Analysis: Unsupervised Learning of Sentence Representations for Ancient LanguagesA PREPRINT Jérôme.Contre Jovinien. Livre I. Number 637 in Sources Chrétiennes. Les Éditions du Cerf, Paris,

    Pith/arXiv arXiv 2004

  24. [2021]

    Grootendorst

    M. Grootendorst. Bertopic: Neural topic modeling with a class-based tf-idf procedure.arXiv preprint arXiv:2203.05794,

    Pith/arXiv arXiv

  25. [2022]

    T. Gao, X. Yao, and D. Chen. SimCSE: Simple contrastive learning of sentence embeddings. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6894–6910,

    2021

  26. [2023]

    D’Angelo, A

    C. D’Angelo, A. Taddei, and A. Lenci. Detecting Semantic Reuse in Ancient Greek Literature: A Computational Approach. InProceedings of the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025), CEUR Workshop Proceedings, Cagliari, Italy,

    2025

  27. [2024]

    doi: 10.1007/978-3-031-70543-4_11. T. Clérice, S. Gabay, M. Vlachou-Efstathiou, A. Pinche, and B. Sagot. CoMMA, a Large-scale Corpus of Multilingual Medieval Archives. working paper or preprint, Oct

    work page doi:10.1007/978-3-031-70543-4_11

  28. [2025]

    Morales; texte grec repris, avec modifications, de H

    ISBN 978-2-204-15958-6.De decretis Nicaenae synodi(CPG 2120); introduction, traduction et notes par X. Morales; texte grec repris, avec modifications, de H. G. Opitz,Athanasius WerkeII, 1–2 (Berlin, 1935). Comprend en annexe la lettre d’Eusèbe de Césarée et le dossier documentaire nicéen. D. Bamman and P. J. Burns. Latin BERT: A contextual language model ...

    Pith/arXiv arXiv 1935

  29. [2026]

    doi: 10.3390/rel17010088

    ISSN 2077-1444. doi: 10.3390/rel17010088. URLhttps://www.mdpi.com/2077-1444/17/1/88. J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT), pages 4171–4186,

    work page doi:10.3390/rel17010088 2077