Structure-Aware Text Recognition for Ancient Greek Critical Editions

Nicolas Angleraud; Antonia Karamolegkou; Benoît Sagot; Thibault Clérice

cs.CVMar 3, 2026

Structure-Aware Text Recognition for Ancient Greek Critical Editions

Authors: Nicolas Angleraud, Antonia Karamolegkou, Benoît Sagot, Thibault Clérice

Organizations: Inria, Paris, France

Abstract

Recent advances in visual language models (VLMs) have transformed end-to-end document understanding. However, their ability to interpret the complex layout semantics of historical scholarly texts remains limited. This paper investigates structure-aware text recognition for Ancient Greek critical editions, which have dense reference hierarchies and extensive marginal annotations. We introduce two novel resources: (i) a large-scale synthetic corpus of 185,000 page images generated from TEI/XML sources with controlled typographic and layout variation, and (ii) a curated benchmark of real scanned editions spanning more than a century of editorial and typographic practices. Using these datasets, we evaluate three state-of-the-art VLMs under both zero-shot and fine-tuning regimes. Our experiments reveal substantial limitations in current VLM architectures when confronted with highly structured historical documents. In zero-shot settings, most models significantly underperform compared to established off-the-shelf software. Nevertheless, the Qwen3VL-8B model achieves state-of-the-art performance, reaching a median Character Error Rate of 1.0% on real scans. These results highlight both the current shortcomings and the future potential of VLMs for structure-aware recognition of complex scholarly documents.

Figures & tables

\fontspec_if_language:nTF ENG \addfontfeature Language=English Figure 1 : Example of the richness of the layout in three different pages.

Corpus	Pages	Words ${}_{\text{med}}$	%Ref	%Note	%Head.
Synthetic	185k	242	78.6%	6.1%	55.2%
Real	460	215	83.0%	32.3%	30.7%

\fontspec_if_language:nTF ENG \addfontfeature Language=English Table 1 : Comparison of structural characteristics across synthetic and real data. The total count of pages, the median number of words per page, and the proportion of pages containing at least one instance of the respective structural element ( \fontspec_if_language:nTF ENG \addfontfeature Language=English<ref> , \fontspec_if_language:nTF ENG \addfontfeature Language=English<note> , or \fontspec_if_language:nTF ENG \addfontfeature Language=Englishmarkdown header for titles and subtitles.

\fontspec_if_language:nTF ENG \addfontfeature Language=English Figure 2 : Overview of the synthetic data generation pipeline. TEI/XML corpora are converted into structured Markdown annotations and rendered into PDF page images via LaTeX, producing aligned text–image pairs.

Model	Phase	LR	Rank	$\alpha$	Eff. batch	Epochs
DeepSeek-OCR-2	Synthetic	2e-4	16	16	32	1
	Real	2e-4	32	32	16	20
	Synth $\to$ Real	2e-4	32	32	16	20
LightOnOCR-2	Synthetic	2e-4	16	32	32	1
	Real	6e-5	16	32	16	20
	Synth $\to$ Real	6e-5	16	32	16	20

\fontspec_if_language:nTF ENG \addfontfeature Language=EnglishTable 2: LoRA fine-tuning hyperparameters. Synth

\to

Real warm-starts from the best synthetic checkpoint.

Model	Train set	Test set	CER ${}_{\text{med}}$	CER ${}_{\text{mean}}$	WER ${}_{\text{med}}$	WER ${}_{\text{mean}}$
Tesseract	—	real	$7.2$	$9.9$	$24.3$	$28.9$
Kraken (End to end)	real	real	$5.3$	$6.8$	$16.9$	$9.3$
DeepSeek-OCR-2	—	synth	$3.8$	$4.6$	$18.6$	$20.5$
	—	real	$10.2$	$13.8$	$31.6$	$35.0$
	synth	synth	$1.6$	$98.3$	$7.9$	$111.4$
	synth	real	$5.5$	$52.7$	$15.5$	$71.1$

\fontspec_if_language:nTF ENG \addfontfeature Language=English Table 3 : Pure OCR results with markup stripped before scoring. Scores are percentages (%). Train set indicates fine-tuning data (— for zero-shot).

Model	Train	Hdr F1	Ref F1	Note F1	Tab ${}_{1\text{st}}$ Spec	Tab ${}_{1\text{st}}$ Rec
DeepSeek-OCR-2	\fontspec_if_language:nTF ENG \addfontfeature Language=Englishsynth	$18.9$	$47.6$	$0.0$	$100.0$	$0.0$
	\fontspec_if_language:nTF ENG \addfontfeature Language=Englishreal	$27.1$	$46.0$	$13.9$	$100.0$	$0.0$
	\fontspec_if_language:nTF ENG \addfontfeature Language=Englishsynth $\to$ real	$39.6$	$56.7$	$27.2$	$100.0$	$0.0$
LightOnOCR-1B	\fontspec_if_language:nTF ENG \addfontfeature Language=Englishsynth	$39.0$	$43.9$	$7.4$	$100.0$	$0.0$
	\fontspec_if_language:nTF ENG \addfontfeature Language=Englishreal	$27.1$	$72.8$	$47.3$	$92.3$	$38.5$
	\fontspec_if_language:nTF ENG \addfontfeature Language=Englishsynth $\to$ real	$21.9$	$51.5$	$19.9$	$100.0$	$0.0$

\fontspec_if_language:nTF ENG \addfontfeature Language=English Table 4 : Structure recognition (%) on the \fontspec_if_language:nTF ENG \addfontfeature Language=Englishreal test set. Headers: count-based F1; references and notes: longest common subsequence F1. Tab

{}_{1\text{st}}

: specificity (Spec) and recall (Rec) for the first paragraph with indentation.

Model	cSub	cIns	cDel	wSub	wIns	wDel	Br	Ac	Is
	Character edits			Word edits			Diacritics
DeepSeek-OCR-2	$8.94$	$3.15$	$1.10$	$44.21$	$3.85$	$1.99$	$4.8$	$18.2$	$1.5$
LightOnOCR-1B	$6.10$	$5.73$	$1.11$	$35.22$	$6.87$	$2.94$	$7.2$	$6.6$	$1.5$
Qwen3-VL-2B	$10.12$	$3.22$	$1.86$	$52.22$	$4.69$	$3.73$	$11.3$	$24.4$	$1.6$
Qwen3-VL-8B	$7.02$	$3.58$	$2.12$	$39.61$	$2.83$	$2.94$	$7.3$	$15.7$	$1.3$
DeepSeek-OCR-2 \fontspec_if_language:nTF ENG \addfontfeature Language=Englishreal	3.22	$4.57$	$2.10$	18.47	$5.18$	$3.17$	$9.9$	$17.1$	$2.4$

\fontspec_if_language:nTF ENG \addfontfeature Language=English Table 5 : Edit rates and diacritic breakdown on real data. cSub/cIns/cDel : per 100 chars; wSub/wIns/wDel : per 100 words. Br./Ac./Is. : breathing, accent, iota subscript confusions (% of cSub).

\fontspec_if_language:nTF ENG \addfontfeature Language=English Figure 4 : Example of output of Qwen3.5-VL.8b on the test set.

Explore similar work

Reading or Guessing? Visual Grounding Failures of Vision-Language Models for OCR in Ancient Greek Editions
May 26, 2026Antonia Karamolegkou, Nicolas Angleraud, Benoît Sagot +1VLM Evaluation Document Understanding
Inria, Paris, France
From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction
Oct 8, 2026Uddipan Basu Bir, Vincent Christlein, Andreas Maier +1Document Information Extraction VLM Adaptation
Friedrich-Alexander-Universität Erlangen–Nürnberg, Germany
When Do VLMs Help Arabic Manuscript OCR? A Cross-Dataset Study
Aug 23, 2026Moshiur Farazi, Firoj Alam, Abderrahmane Maaradji +3Optical Character Recognition Document Image Understanding
University of Doha for Science and Technology, Doha, Qatar · Qatar Computing Research Institute, Hamad Bin Khalifa University, Doha, Qatar · Northwestern University in Qatar, Doha, Qatar

May 26, 2026J/K move · Enter open · S save

Reading or Guessing? Visual Grounding Failures of Vision-Language Models for OCR in Ancient Greek Editions

Antonia Karamolegkou, Nicolas Angleraud, Benoît Sagot, Thibault Clérice

Inria, Paris, France

VLM Evaluation Document Understanding Historical Document Analysis VLM Hallucination Vision-Language Grounding Optical Character Recognition

Open paper AskPDF arXiv

Recent work has shown that Vision-Language Models (VLMs) used for optical character recognition (OCR) can generate plausible but visually unsupported text, suggesting reliance on language priors. Comparing open-weight VLMs with traditional OCR baselines on low-resource Ancient Greek critical editions, we show that VLM errors often remain fluent even when wrong, producing plausible Greek substitutions where traditional engines produce local recognition noise. To analyze visual evidence during decoding, we introduce controlled image perturbations and token-level grounding measures based on conditional versus image-free decoding distributions. Under character-level perturbations, VLMs diverge sharply from the perturbed ground truth while traditional OCR remains comparatively faithful; however, token-level analysis shows that prior reliance is model-specific: in an OCR-specialist model, fluent lexical errors are produced with little reliance on the image, whereas general-purpose VLMs remain conditioned on the visual input even when wrong. Decode-time interventions fail to reliably restore grounding, while post-OCR language-model correction improves several systems only by repairing text after generation. Our results extend prior evidence of OCR language-prior reliance to low-resource historical documents and a broader set of models, showing that fluent output is not necessarily visually grounded and motivating interpretability-driven evaluation beyond aggregate accuracy.

Abstract

Explore similar work

Reading or Guessing? Visual Grounding Failures of Vision-Language Models for OCR in Ancient Greek Editions

From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction

When Do VLMs Help Arabic Manuscript OCR? A Cross-Dataset Study