DocOCR-Eval: A Correction-Based Framework for OCR Tool Selection Without Ground Truth

Lawrence Chun Man Lau; Puzhen Wu; Sirui Li; Wei Liu; Yifan Peng; Yihao Ding; Zihan Xu

REVIEW 4 major objections 5 minor 251 references

The paper claims that OCR tool selection can be done without ground-truth labels by measuring how much a multimodal language model corrects each engine's output, and shows that aggregating several models yields rankings that closely match h

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 · deepseek-v4-flash

2026-08-02 14:51 UTC pith:UGJP7HMP

load-bearing objection DocOCR-Eval's idea is good and the ground-truth validation is honest, but the evidence base (50 images, four engines, no error bars, no code) is too thin to trust the high NDCG numbers as a general result yet. the 4 major comments →

arxiv 2607.16203 v1 pith:UGJP7HMP submitted 2026-05-05 cs.LG cs.AIcs.CL

DocOCR-Eval: A Correction-Based Framework for OCR Tool Selection Without Ground Truth

Zihan Xu , Puzhen Wu , Lawrence Chun Man Lau , Wei Liu , Sirui Li , Yifan Peng , Yihao Ding This is my paper

classification cs.LG cs.AIcs.CL

keywords OCR tool selectionannotation-free evaluationMLLM correctionAverage Normalized Levenshtein Similarityranking consistencyerror diagnosisdocument parsingweak supervision

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.

The paper tries to establish that you can choose the best OCR engine for a given document collection without any manually labeled ground truth. The trick is to have a multimodal language model (MLLM) try to correct each engine's output and then rank the engines by how little correction their outputs need: fewer, smaller corrections mean a better engine. The authors show that averaging the correction scores across multiple MLLMs makes these automatic rankings increasingly close to rankings computed from human annotations, with high NDCG scores on several benchmarks. If correct, this removes the main bottleneck — annotation cost — that currently prevents adaptive, per-collection OCR tool selection.

Core claim

The central claim is that the discrepancy between an OCR engine's raw output and an MLLM's corrected version is a valid proxy for recognition quality, and that ranking engines by post-correction ANLS reproduces annotation-based ordering. The paper demonstrates this through a three-stage pipeline: error diagnosing (character noise, tokenization, semantic consistency), conditional correction (text-based for character/token errors, vision-based re-OCR for semantic inconsistencies), and ranking by ANLS. Across five public benchmarks, the average NDCG between correction-based and ground-truth rankings reached 0.9752 on one English dataset, 0.9679 on a Chinese dataset, and 0.9699 on a French datas

What carries the argument

The core mechanism is the correction-discrepancy signal: each OCR output block is passed through three binary detectors (character noise, tokenization, semantic consistency), which trigger either a text-only corrector or a vision-based re-OCR corrector. The engine is then scored by the Average Normalized Levenshtein Similarity (ANLS) between the raw and corrected outputs, and engines are ranked by this score, optionally averaged over multiple MLLMs to reduce individual model bias.

Load-bearing premise

The entire framework rests on the premise that the amount (and kind) of correction an MLLM applies to an OCR output is a valid and consistent measure of that output's quality — better engines need smaller or more neutral corrections, and the MLLM is equally reliable across all engines compared.

What would settle it

Construct a held-out set with known ground truth, run the correction-based ranking, and search for counterexamples: a low-quality engine whose output is left nearly unchanged because the corrector mistakes plausible-sounding errors for correct text, or a high-quality engine whose accurate output is needlessly rewritten. If such cases are common, the ANLS scores will not track true error rates and the ranking will invert.

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

If this is right

OCR tool selection can be automated for a new document collection with zero labeled samples, enabling deployment-time adaptation.
As new OCR engines appear, the best tool can be re-evaluated and swapped without spending money or time on annotations.
The optimal correction strategy is dataset-dependent — some collections benefit from text-only normalization, others need vision-based re-OCR — so the pipeline should be tuned per collection.
The correction step adds latency and cost, so this approach is best suited to high-accuracy settings, not lightweight retrieval or coarse indexing.
Correction-based rankings could serve as a weak supervision signal for training downstream document-understanding models without ground truth.

Where Pith is reading between the lines

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

The same discrepancy-ranking logic could be applied to other 'pick the best tool without labels' problems, such as speech-to-text or translation engines, whenever a strong model can act as a corrector.
A corrector that is biased toward a particular textual style (for example, normalizing dialect, script, or formatting differences) could silently favor engines that produce the corrector's preferred form rather than the most accurate form — a risk that is not explicitly tested in the paper.
The ensemble could be made more robust by combining corrector decisions (e.g., agreement-based scoring) rather than merely averaging ANLS scores, which is a testable extension of the paper's approach.
The reported NDCG values come from benchmarks where ground truth exists for validation; a true test would require deploying the framework on a label-free collection and checking downstream task performance.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

DocOCR-Eval's idea is good and the ground-truth validation is honest, but the evidence base (50 images, four engines, no error bars, no code) is too thin to trust the high NDCG numbers as a general result yet.

read the letter

DocOCR-Eval is a sensible new idea: instead of paying for ground truth to compare OCR engines, run the outputs through an MLLM corrector and rank by how much the corrector changes them. The paper validates that proxy against real annotation-based rankings on three datasets, which is the right thing to do — it doesn't just assert that the proxy works. The staged diagnosis (character, tokenization, semantic) before choosing text- vs vision-based correction is a reasonable way to make the corrections conditional, and the observation that the optimal correction strategy differs by dataset (RXPAD wants text-only normalization) is a useful practical detail. No prior work does this, so the combination is new.

The problems are mostly about how much evidence exists. Only 50 images per dataset, no error bars, no multiple seeds. The 'progressively improves with aggregation' claim is shown as one jump from a single MLLM to an average of four, not a real curve. The ranking-consistency table covers only three of the five datasets, though the abstract claims all five. And there is no released code or prompts, so reproducing the numbers is hard.

The deeper conceptual issue is the monotonicity assumption: that the discrepancy between OCR output and MLLM correction tracks true OCR quality. The paper never probes this directly. If the MLLM has a stylistic preference — say, it normalizes punctuation or expands abbreviations — it will 'correct' accurate but non-preferred text, and that bias becomes a ranking error. The stress-test note worries about hallucination in weak languages; that's a legitimate worry, but it has to be weighed against the fact that the NDCG numbers on EPHOIE, FUNSD, and RXPAD are high (0.97-0.98) and come from a comparison with external labels, so the proxy is at least useful on these datasets, not obviously broken.

Who should read this? Practitioners assembling document-AI pipelines who want a cheap way to pick an OCR engine, and researchers working on unsupervised evaluation. I'd send it to peer review but ask for a major revision: add error bars and more engines/MLLMs, test the monotonicity assumption directly (e.g., by corrupting clean text in known ways and checking whether the correction score changes monotonically), and release the prompts and code. The framework is worth the referee's time, but the current evidence does not justify 'extensive experiments'.

Referee Report

4 major / 5 minor

Summary. The paper proposes DocOCR-Eval, a ground-truth-free framework for ranking OCR engines on a given document collection. For each OCR output block, MLLM-based diagnosing modules flag character-level, tokenization, and semantic errors; a text-based or vision-based corrector then edits flagged blocks; ANLS between original and corrected outputs is averaged over documents and engines are ranked by minimal discrepancy. The paper evaluates four OCR engines and two MLLMs against annotated benchmarks (FUNSD, SROIE, EPHOIE, RXPAD, and XFUND subsets) and reports NDCG values around 0.95–0.98 for correction-based vs ground-truth rankings on three datasets, concluding that aggregating multiple MLLMs progressively approximates annotation-based ordering without labels.

Significance. If the central claim holds, the paper makes a practically useful contribution: label-free OCR tool selection across languages and domains, with a simple multi-stage correction protocol and broad benchmark coverage. The framework does not feed validation ground truth back into the ranking formula, so circularity is not a serious concern. Strengths include the systematic comparison of traditional OCR engines and MLLMs, the multi-MLLM aggregation idea, and candid discussion of latency/cost trade-offs in Sec. 3.4. However, the result is only as strong as the untested monotonicity assumption that post-correction discrepancy faithfully proxies OCR quality, and the current evidence is statistically thin.

major comments (4)

[Sec. 2.2] The central premise is that ANLS between an OCR block and its MLLM-corrected version is a monotone proxy for OCR quality. This is asserted in Sec. 2.2 but never justified theoretically or tested directly. If the MLLM has a stylistic bias (punctuation normalization, abbreviation expansion, script conventions) or hallucinates, it will 'correct' high-quality text into a preferred form and penalize the better engine. The paper's own finding that the optimal correction strategy is dataset-dependent (Sec. 3.2.2, RXPAD) confirms that the proxy is not a fixed quality measure. I ask for a controlled monotonicity test — e.g., injecting known character/word errors at graduated rates and checking that post-correction ANLS decreases monotonically — or a correlation analysis between correction-based and GT-based ANLS per engine/dataset.
[Sec. 2.3.1 / Tables 4–5] All experiments are run on one random 50-image subsample per benchmark, with no seeds, repeated subsampling, or confidence intervals. Consequently the reported NDCG values (e.g., 0.9752, 0.9679, 0.9699) are point estimates whose noise level is unknown. The claim that multi-model aggregation 'progressively improves' alignment (Abstract; Sec. 3.2.1) is supported only by comparing k=1 and k=4 with no variance. Report mean±std over at least 5 subsample seeds and, for the aggregation claim, plot NDCG as a function of the number of MLLMs.
[Sec. 3.2.2 / Table 5] The ranking-consistency analysis is presented only for EPHOIE, FUNSD, and RXPAD. SROIE and XFUND — which are included in Tables 3–4 and are central to the cross-domain/cross-language claim — are absent from the headline NDCG results. Either add the correction-based ranking tables for those datasets or restrict the 'across five benchmarks' claim to the ANLS table.
[Sec. 3.2.2 / Table 5] Table 5 reports NDCG only for Qwen3-VL, not for the multi-MLLM aggregate that is the paper's proposed remedy for single-model bias. Table 4 gives average ANLS after aggregation but no ranking consistency. Thus the abstract's claim that aggregation 'progressively improves alignment with annotation-based rankings' is not directly evidenced in ranking space. Report NDCG for each individual MLLM and for the k-model average.

minor comments (5)

[Table 2] The caption says 'H: Handwritten' but FUNSD, SROIE, and XFUND are not handwritten; only RXPAD is. Please correct the notation.
[References] The GPT-4.1-mini citation points to 'Hello gpt-4o' (OpenAI, 2024); this appears to be the wrong model/version. Please verify.
[Sec. 2.3.3] ROUGE-L, CER, and word-level F1 are listed as evaluation metrics, but Figure 2 only reports a subset; state explicitly which metrics appear in each figure/table.
[Table 5] The Avg.d column does not equal the arithmetic mean of the five per-strategy d values (e.g., FUNSD: mean is 0.9770 not 0.9752; RXPAD: 0.9717 not 0.9699). Clarify the aggregation formula or round consistently.
[Sec. 3.3] The case study text says Step 1 detects 'ACCEPRANCE-ANTRO' while Figure 1 shows 'ACCEPTANCE-ANTRO'; align the example. Also, in Figure 1 the flow might be clearer if the three diagnostic signals are shown as vectors.

Circularity Check

0 steps flagged

No significant circularity: the correction-based ranking is validated externally against ground-truth annotations.

full rationale

The paper's derivation chain is: (i) define an annotation-free score as ANLS between original OCR output and MLLM-corrected output; (ii) rank engines by this score; (iii) evaluate the resulting ranking against ground-truth-based rankings using NDCG on five benchmarks. Nothing in this chain fits parameters to the labels being predicted. The correction-based score is not derived from, nor defined in terms of, the annotation-based ranking; the ground-truth rankings in Tables 3-5 are used only as external validation targets. The core assumption that correction discrepancy is monotone in OCR quality is a substantive empirical hypothesis (a validity risk, especially under MLLM stylistic bias or hallucination), not a tautology: the paper could in principle fail, and indeed reports imperfect NDCG values below 1 on several datasets. Self-citations (Ding et al. 2023, 2025a, 2025b) appear only in the motivation and related context, not as load-bearing evidence for the ranking claim. No uniqueness theorem, fitted parameter, or ansatz is imported from the authors' prior work. Therefore no definitional circularity, fitted-input-as-prediction, or self-citation chain is present. The appropriate verdict is no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The framework introduces no new physical or mathematical entities, but it relies on domain assumptions about MLLM correction behavior and diagnostic reliability. The only hand-chosen experimental quantities are the 50-image subsample and the number of aggregated MLLMs; neither is fitted to the data, but both shape the reported results.

free parameters (2)

Evaluation subsample size per dataset = 50 images
All experiments use a randomly sampled 50-image subset of each test split (Section 2.3.1). Results and rankings depend on this choice, and no multiple-seed analysis is provided.
Number of aggregated MLLMs = 4 (Qwen3-VL-4B, Qwen3-VL-8B, GPT-4.1-mini, Gemini-2.5-flash-lite)
The aggregation claim is demonstrated only with all four models averaged together; no curve over k=1,2,3,4 is shown, so 'progressively improves' is not directly evidenced.

axioms (4)

domain assumption Discrepancy between original OCR output and MLLM-corrected output is a monotone proxy for OCR error rate.
Section 2.2 defines OCR quality this way. If MLLM corrections hallucinate, normalize style, or reflect model preference, a perfect OCR could show large discrepancy and rankings would be biased.
domain assumption MLLM binary diagnostics (Dc, Dt, Ds) detect true errors with sufficient precision and recall.
Section 2.2.3 triggers correctors based on these decisions. Systematic false positives would equalize discrepancies across engines; systematic false negatives would suppress needed corrections.
domain assumption The annotated benchmarks used for validation are representative of real-world document collections and their ground-truth annotations are correct.
Section 2.3.1 introduces the datasets; NDCG is computed against these GT labels, so annotation errors or domain gaps transfer directly into the validation results.
standard math ANLS and NDCG are valid metrics for the intended comparisons.
ANLS is a length-normalized edit similarity and NDCG assumes top-ranked positions matter more; these are standard but carry assumptions about what 'good' OCR ranking means.

pith-pipeline@v1.3.0-alltime-deepseek · 8152 in / 9401 out tokens · 93152 ms · 2026-08-02T14:51:28.863488+00:00 · methodology

0 comments

read the original abstract

Document parsing is a foundational step for document understanding tasks such as visual question answering and key information extraction, as it transforms unstructured scanned images into structured representations by extracting textual, visual, and layout information. While numerous Optical Character Recognition (OCR) engines and multimodal large language models (MLLMs) have been developed for this purpose, selecting an appropriate document parsing solution for a given document collection remains challenging, particularly in label-scarce settings. In this work, we conduct a systematic evaluation of text recognition performance across a diverse set of OCR engines and state-of-the-art MLLMs on multiple scanned document benchmarks spanning different domains and languages. Motivated by the limited contextual reasoning capabilities of many OCR engines and the high cost of manual annotations, we propose DocOCR-Eval, an annotation-free evaluation framework for automatic OCR assessment and selection. DocOCR-Eval employs a three-staged correction and ranking strategy to approximate annotation-based tool ordering without ground-truth labels. We show that aggregating across multiple MLLMs progressively improves alignment with annotation-based rankings. Extensive experiments further demonstrate that reliable OCR tool selection can be achieved in realistic, label-limited settings, providing practical guidance for deploying document parsing systems across diverse real-world document collections.

Figures

Figures reproduced from arXiv: 2607.16203 by Lawrence Chun Man Lau, Puzhen Wu, Sirui Li, Wei Liu, Yifan Peng, Yihao Ding, Zihan Xu.

**Figure 1.** Figure 1: Workflow of the DOCOCR-EVAL with a case study. Text-based Corrector (Ct) revises OCR outputs using linguistic and contextual cues alone, without introducing content beyond the recognized text. The Vision-based Corrector (Cv) additionally incorporates the OCR bounding box cropped from the source image and instructs the MLLM to rerecognize only the visible text within that region. 2.2.3 Error Diagnosis, C… view at source ↗

**Figure 2.** Figure 2: Metrics among diverse datasets by Qwen3- [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

251 extracted references · 4 canonical work pages

[1]

A comparison study on optical character recognition models in mathematical equations and in any language

Francis, Sofi A and Sangeetha, M. A comparison study on optical character recognition models in mathematical equations and in any language. Results in Control and Optimization
[2]

Seeing justice clearly: Handwritten legal document translation with OCR and vision-language models

Nigam, Shubham Kumar and Shukla, Parjanya Aditya and Shallum, Noel and Bhattacharya, Arnab. Seeing justice clearly: Handwritten legal document translation with OCR and vision-language models. arXiv:2512.18004

arXiv
[3]

2025 , howpublished =

Vikas Paruchuri and Datalab Team , title =. 2025 , howpublished =

2025
[4]

2025 , month = aug, version =

The Kraken OCR system , author =. 2025 , month = aug, version =

2025
[5]

2502.18443 , archivePrefix=

Jake Poznanski and Jon Borchardt and Jason Dunkelberger and Regan Huff and Daniel Lin and Aman Rangapur and Christopher Wilhelm and Kyle Lo and Luca Soldaini , year=. 2502.18443 , archivePrefix=

arXiv
[6]

A feasible framework for arbitrary-shaped Scene Text Recognition

Zhang, Jinjin and Wang, Wei and Huang, Di and Liu, Qingjie and Wang, Yunhong. A feasible framework for arbitrary-shaped Scene Text Recognition. arXiv:1912.04561

Pith/arXiv arXiv 1912
[7]

Digital Humanities Quarterly , author =

Calamari -. Digital Humanities Quarterly , author =
[8]

2023 , url =

EasyOCR: Ready-to-use OCR with 80+ Supported Languages , author =. 2023 , url =

2023
[9]

2023 , version =

Tesseract OCR , author =. 2023 , version =

2023
[10]

2025 , eprint=

PaddleOCR 3.0 Technical Report , author=. 2025 , eprint=

2025
[11]

2021 , publisher =

docTR: Document Text Recognition , author=. 2021 , publisher =

2021
[12]

2025 , url =

Ocular: Optical Character Understanding and Layout , author =. 2025 , url =

2025
[13]

2025 , url =

Adobe PDF Extract API , author =. 2025 , url =

2025
[14]

2025 , url =

Mistral OCR , author =. 2025 , url =

2025
[15]

2025 , url =

Google Cloud Vision OCR , author =. 2025 , url =

2025
[16]

2024 , eprint=

MinerU: An Open-Source Solution for Precise Document Content Extraction , author=. 2024 , eprint=

2024
[17]

2025 , url =

pdfminer.six: Python PDF Parser and Analyzer , author =. 2025 , url =

2025
[18]

2025 , howpublished =

LightOnOCR-1B: End-to-End and Efficient Domain-Specific Vision-Language Models for OCR , author =. 2025 , howpublished =

2025
[19]

arXiv preprint arXiv:2511.21631 , year=

Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year=

Pith/arXiv arXiv
[20]

2025 , url =

OpenAI Images & Vision API , author =. 2025 , url =

2025
[21]

arXiv preprint arXiv:2510.18234 , year=

DeepSeek-OCR: Contexts Optical Compression , author=. arXiv preprint arXiv:2510.18234 , year=

Pith/arXiv arXiv
[22]

2025 , journal=

HunyuanOCR Technical Report , author=. 2025 , journal=

2025
[23]

arXiv preprint arXiv:2505.07062 , year=

Seed1.5-VL Technical Report , author=. arXiv preprint arXiv:2505.07062 , year=

Pith/arXiv arXiv
[24]

Aho and Jeffrey D

Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

1972
[25]

arXiv preprint arXiv:2507.09861 , year=

A Survey on MLLM-based Visually Rich Document Understanding: Methods, Challenges, and Emerging Trends , author=. arXiv preprint arXiv:2507.09861 , year=

Pith/arXiv arXiv
[26]

Publications Manual , year = "1983", publisher =

1983
[27]

Information Fusion , pages=

SynJAC: Synthetic-data-driven Joint-granular Adaptation and Calibration for Domain Specific Scanned Document Key Information Extraction , author=. Information Fusion , pages=. 2025 , publisher=

2025
[28]

Chandra and Dexter C

Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

arXiv 1981
[29]

Scalable training of

Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of
[30]

Dan Gusfield , title =. 1997

1997
[31]

Tetreault , title =

Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

2015
[32]

A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =
[33]

FirstName LastName , title =
[34]

FirstName Alpher , title =
[35]

Journal of Foo , volume = 13, number = 1, pages =

FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =
[36]

Journal of Foo , volume = 14, number = 1, pages =

FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =
[37]

FirstName Alpher and FirstName Gamow , title =
[38]

Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

Layoutlm: Pre-training of text and layout for document image understanding , author=. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=. 2020 , publisher=

2020
[39]

Computer Vision and Image Processing - 5th International Conference,

Yash Srivastava and Vaishnav Murali and Shiv Ram Dubey and Snehasis Mukherjee , title =. Computer Vision and Image Processing - 5th International Conference,. 2020 , url =

2020
[40]

LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=. 2021 , url=

2021
[41]

Proceedings of the 30th ACM International Conference on Multimedia , pages=

Layoutlmv3: Pre-training for document ai with unified text and image masking , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=. 2022 , publisher =

2022
[42]

arXiv preprint arXiv:2410.01609 , year=

DAViD: Domain Adaptive Visually-Rich Document Understanding with Synthetic Insights , author=. arXiv preprint arXiv:2410.01609 , year=

arXiv
[43]

arXiv preprint arXiv:2104.08836 , year=

Layoutxlm: Multimodal pre-training for multilingual visually-rich document understanding , author=. arXiv preprint arXiv:2104.08836 , year=

Pith/arXiv arXiv
[44]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Xylayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2022 , publisher =

2022
[45]

arXiv preprint arXiv:1908.03557 , year=

Visualbert: A simple and performant baseline for vision and language , author=. arXiv preprint arXiv:1908.03557 , year=

Pith/arXiv arXiv 1908
[46]

LXMERT: Learning Cross-Modality Encoder Representations from Transformers , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages=. 2019 , url =

2019
[47]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Docformer: End-to-end transformer for document understanding , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=. 2021 , publisher =

2021
[48]

RotateKV : Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations

Su, Zunhai and Wei, Hanyu and Chen, Zhe and Shen, Wang and Li, Linge and Yu, Huangqi and Yuan, Kehong. RotateKV : Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations. Proceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence
[49]

Thirty-Eighth

DocFormerv2: Local Features for Document Understanding , author =. Thirty-Eighth. 2024 , url =

2024
[50]

Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence,

Fast-StrucTexT: An Efficient Hourglass Transformer with Modality-guided Dynamic Token Merge for Document Understanding , author=. Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence,. 2023 , publisher =

2023
[51]

F orm N et V 2: Multimodal Graph Contrastive Learning for Form Document Information Extraction

Lee, Chen-Yu and Li, Chun-Liang and Zhang, Hao and Dozat, Timothy and Perot, Vincent and Su, Guolong and Zhang, Xiang and Sohn, Kihyuk and Glushnev, Nikolay and Wang, Renshen and Ainslie, Joshua and Long, Shangbang and Qin, Siyang and Fujii, Yasuhisa and Hua, Nan and Pfister, Tomas. F orm N et V 2: Multimodal Graph Contrastive Learning for Form Document I...

2023
[52]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Docparser: Hierarchical document structure parsing from renderings , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2021 , url =

2021
[53]

Findings of the Association for Computational Linguistics: EMNLP 2022 , publisher =

XDoc: Unified Pre-training for Cross-Format Document Understanding , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , publisher =. 2022 , url =

2022
[54]

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =

MarkupLM: Pre-training of Text and Markup Language for Visually Rich Document Understanding , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =. 2022 , url =

2022
[55]

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =

LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =. 2022 , url =

2022
[56]

Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual , pages =

Unidoc: Unified pretraining framework for document understanding , author=. Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual , pages =. 2021 , url =

2021
[57]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , publisher =

Selfdoc: Self-supervised document representation learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , publisher =. 2021 , url =

2021
[58]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I , pages=

Lambert: Layout-aware language modeling for information extraction , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I , pages=. 2021 , organization=

2021
[59]

Proceedings of the 29th ACM International Conference on Multimedia , publisher =

Structext: Structured text understanding with multi-modal transformers , author=. Proceedings of the 29th ACM International Conference on Multimedia , publisher =. 2021 , url =

2021
[60]

Proceedings of the AAAI Conference on Artificial Intelligence , pages=

Bros: A pre-trained language model focusing on text and layout for better key information extraction from documents , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=. 2022 , url =

2022
[61]

Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

InfoVQA: Visual Question Answering on Infographics with A Multi-Modal Entity Graph , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=
[62]

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics (ACL) , year=

Compositional Semantic Parsing on Semi-Structured Tables , author=. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics (ACL) , year=
[63]

International Conference on Learning Representations (ICLR) , year=

TabFact: A Large-scale Dataset for Table-based Fact Verification , author=. International Conference on Learning Representations (ICLR) , year=
[64]

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

2022
[65]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

TextVQA: Visual Question Answering with Reading , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=
[66]

Proceedings of the European Conference on Computer Vision (ECCV) , year=

TextCaps: A Dataset for Image Captioning with Reading , author=. Proceedings of the European Conference on Computer Vision (ECCV) , year=
[67]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=

Going full-tilt boogie on document understanding with text-image-layout transformer , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=. 2021 , url =

2021
[68]

Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXVIII , pages=

Ocr-free document understanding transformer , author=. Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXVIII , pages=. 2022 , organization=

2022
[69]

arXiv preprint arXiv:2209.08569 , year=

ERNIE-mmLayout: Multi-grained MultiModal Transformer for Document Understanding , author=. arXiv preprint arXiv:2209.08569 , year=

Pith/arXiv arXiv
[70]

IEEE Transactions on Multimedia , year=

Multimodal pre-training based on graph attention network for document understanding , author=. IEEE Transactions on Multimedia , year=
[71]

Information Sciences , volume=

Document image layout analysis via explicit edge embedding network , author=. Information Sciences , volume=. 2021 , publisher=

2021
[72]

Proceedings of the 30th ACM International Conference on Multimedia , pages=

Dit: Self-supervised pre-training for document image transformer , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=. 2022 , publisher =

2022
[73]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=

VSR: a unified framework for document layout analysis combining vision, semantics and relations , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=. 2021 , organization=

2021
[74]

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

LayoutReader: Pre-training of Text and Layout for Reading Order Detection , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=. 2021 , publisher =

2021
[75]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=

ViBERTgrid: a jointly trained multi-modal 2D document representation for key information extraction from documents , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=. 2021 , organization=

2021
[76]

StructuralLM: Structural Pre-training for Form Understanding , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=. 2021 , url =

2021
[77]

arXiv preprint arXiv:2009.14457 , year=

Towards a multi-modal, multi-task learning based pre-training framework for document representation learning , author=. arXiv preprint arXiv:2009.14457 , year=

Pith/arXiv arXiv 2009
[78]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=

LayoutParser: A unified toolkit for deep learning based document image analysis , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=. 2021 , organization=

2021
[79]

Document Analysis and Recognition--ICDAR 2021 Workshops: Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=

Mtl-foun: A multi-task learning approach to form understanding , author=. Document Analysis and Recognition--ICDAR 2021 Workshops: Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=. 2021 , organization=

2021
[80]

International Conference on Neural Information Processing , pages=

Improving Document Image Understanding with Reinforcement Finetuning , author=. International Conference on Neural Information Processing , pages=. 2022 , organization=

2022

Showing first 80 references.

[1] [1]

A comparison study on optical character recognition models in mathematical equations and in any language

Francis, Sofi A and Sangeetha, M. A comparison study on optical character recognition models in mathematical equations and in any language. Results in Control and Optimization

[2] [2]

Seeing justice clearly: Handwritten legal document translation with OCR and vision-language models

Nigam, Shubham Kumar and Shukla, Parjanya Aditya and Shallum, Noel and Bhattacharya, Arnab. Seeing justice clearly: Handwritten legal document translation with OCR and vision-language models. arXiv:2512.18004

arXiv

[3] [3]

2025 , howpublished =

Vikas Paruchuri and Datalab Team , title =. 2025 , howpublished =

2025

[4] [4]

2025 , month = aug, version =

The Kraken OCR system , author =. 2025 , month = aug, version =

2025

[5] [5]

2502.18443 , archivePrefix=

Jake Poznanski and Jon Borchardt and Jason Dunkelberger and Regan Huff and Daniel Lin and Aman Rangapur and Christopher Wilhelm and Kyle Lo and Luca Soldaini , year=. 2502.18443 , archivePrefix=

arXiv

[6] [6]

A feasible framework for arbitrary-shaped Scene Text Recognition

Zhang, Jinjin and Wang, Wei and Huang, Di and Liu, Qingjie and Wang, Yunhong. A feasible framework for arbitrary-shaped Scene Text Recognition. arXiv:1912.04561

Pith/arXiv arXiv 1912

[7] [7]

Digital Humanities Quarterly , author =

Calamari -. Digital Humanities Quarterly , author =

[8] [8]

2023 , url =

EasyOCR: Ready-to-use OCR with 80+ Supported Languages , author =. 2023 , url =

2023

[9] [9]

2023 , version =

Tesseract OCR , author =. 2023 , version =

2023

[10] [10]

2025 , eprint=

PaddleOCR 3.0 Technical Report , author=. 2025 , eprint=

2025

[11] [11]

2021 , publisher =

docTR: Document Text Recognition , author=. 2021 , publisher =

2021

[12] [12]

2025 , url =

Ocular: Optical Character Understanding and Layout , author =. 2025 , url =

2025

[13] [13]

2025 , url =

Adobe PDF Extract API , author =. 2025 , url =

2025

[14] [14]

2025 , url =

Mistral OCR , author =. 2025 , url =

2025

[15] [15]

2025 , url =

Google Cloud Vision OCR , author =. 2025 , url =

2025

[16] [16]

2024 , eprint=

MinerU: An Open-Source Solution for Precise Document Content Extraction , author=. 2024 , eprint=

2024

[17] [17]

2025 , url =

pdfminer.six: Python PDF Parser and Analyzer , author =. 2025 , url =

2025

[18] [18]

2025 , howpublished =

LightOnOCR-1B: End-to-End and Efficient Domain-Specific Vision-Language Models for OCR , author =. 2025 , howpublished =

2025

[19] [19]

arXiv preprint arXiv:2511.21631 , year=

Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year=

Pith/arXiv arXiv

[20] [20]

2025 , url =

OpenAI Images & Vision API , author =. 2025 , url =

2025

[21] [21]

arXiv preprint arXiv:2510.18234 , year=

DeepSeek-OCR: Contexts Optical Compression , author=. arXiv preprint arXiv:2510.18234 , year=

Pith/arXiv arXiv

[22] [22]

2025 , journal=

HunyuanOCR Technical Report , author=. 2025 , journal=

2025

[23] [23]

arXiv preprint arXiv:2505.07062 , year=

Seed1.5-VL Technical Report , author=. arXiv preprint arXiv:2505.07062 , year=

Pith/arXiv arXiv

[24] [24]

Aho and Jeffrey D

Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

1972

[25] [25]

arXiv preprint arXiv:2507.09861 , year=

A Survey on MLLM-based Visually Rich Document Understanding: Methods, Challenges, and Emerging Trends , author=. arXiv preprint arXiv:2507.09861 , year=

Pith/arXiv arXiv

[26] [26]

Publications Manual , year = "1983", publisher =

1983

[27] [27]

Information Fusion , pages=

SynJAC: Synthetic-data-driven Joint-granular Adaptation and Calibration for Domain Specific Scanned Document Key Information Extraction , author=. Information Fusion , pages=. 2025 , publisher=

2025

[28] [28]

Chandra and Dexter C

Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

arXiv 1981

[29] [29]

Scalable training of

Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

[30] [30]

Dan Gusfield , title =. 1997

1997

[31] [31]

Tetreault , title =

Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

2015

[32] [32]

A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

[33] [33]

FirstName LastName , title =

[34] [34]

FirstName Alpher , title =

[35] [35]

Journal of Foo , volume = 13, number = 1, pages =

FirstName Alpher and FirstName Fotheringham-Smythe , title =. Journal of Foo , volume = 13, number = 1, pages =

[36] [36]

Journal of Foo , volume = 14, number = 1, pages =

FirstName Alpher and FirstName Fotheringham-Smythe and FirstName Gamow , title =. Journal of Foo , volume = 14, number = 1, pages =

[37] [37]

FirstName Alpher and FirstName Gamow , title =

[38] [38]

Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

Layoutlm: Pre-training of text and layout for document image understanding , author=. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=. 2020 , publisher=

2020

[39] [39]

Computer Vision and Image Processing - 5th International Conference,

Yash Srivastava and Vaishnav Murali and Shiv Ram Dubey and Snehasis Mukherjee , title =. Computer Vision and Image Processing - 5th International Conference,. 2020 , url =

2020

[40] [40]

LayoutLMv2: Multi-modal Pre-training for Visually-rich Document Understanding , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=. 2021 , url=

2021

[41] [41]

Proceedings of the 30th ACM International Conference on Multimedia , pages=

Layoutlmv3: Pre-training for document ai with unified text and image masking , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=. 2022 , publisher =

2022

[42] [42]

arXiv preprint arXiv:2410.01609 , year=

DAViD: Domain Adaptive Visually-Rich Document Understanding with Synthetic Insights , author=. arXiv preprint arXiv:2410.01609 , year=

arXiv

[43] [43]

arXiv preprint arXiv:2104.08836 , year=

Layoutxlm: Multimodal pre-training for multilingual visually-rich document understanding , author=. arXiv preprint arXiv:2104.08836 , year=

Pith/arXiv arXiv

[44] [44]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Xylayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2022 , publisher =

2022

[45] [45]

arXiv preprint arXiv:1908.03557 , year=

Visualbert: A simple and performant baseline for vision and language , author=. arXiv preprint arXiv:1908.03557 , year=

Pith/arXiv arXiv 1908

[46] [46]

LXMERT: Learning Cross-Modality Encoder Representations from Transformers , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages=. 2019 , url =

2019

[47] [47]

Proceedings of the IEEE/CVF international conference on computer vision , pages=

Docformer: End-to-end transformer for document understanding , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=. 2021 , publisher =

2021

[48] [48]

RotateKV : Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations

Su, Zunhai and Wei, Hanyu and Chen, Zhe and Shen, Wang and Li, Linge and Yu, Huangqi and Yuan, Kehong. RotateKV : Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations. Proceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence

[49] [49]

Thirty-Eighth

DocFormerv2: Local Features for Document Understanding , author =. Thirty-Eighth. 2024 , url =

2024

[50] [50]

Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence,

Fast-StrucTexT: An Efficient Hourglass Transformer with Modality-guided Dynamic Token Merge for Document Understanding , author=. Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence,. 2023 , publisher =

2023

[51] [51]

F orm N et V 2: Multimodal Graph Contrastive Learning for Form Document Information Extraction

Lee, Chen-Yu and Li, Chun-Liang and Zhang, Hao and Dozat, Timothy and Perot, Vincent and Su, Guolong and Zhang, Xiang and Sohn, Kihyuk and Glushnev, Nikolay and Wang, Renshen and Ainslie, Joshua and Long, Shangbang and Qin, Siyang and Fujii, Yasuhisa and Hua, Nan and Pfister, Tomas. F orm N et V 2: Multimodal Graph Contrastive Learning for Form Document I...

2023

[52] [52]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Docparser: Hierarchical document structure parsing from renderings , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2021 , url =

2021

[53] [53]

Findings of the Association for Computational Linguistics: EMNLP 2022 , publisher =

XDoc: Unified Pre-training for Cross-Format Document Understanding , author=. Findings of the Association for Computational Linguistics: EMNLP 2022 , publisher =. 2022 , url =

2022

[54] [54]

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =

MarkupLM: Pre-training of Text and Markup Language for Visually Rich Document Understanding , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =. 2022 , url =

2022

[55] [55]

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =

LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , publisher =. 2022 , url =

2022

[56] [56]

Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual , pages =

Unidoc: Unified pretraining framework for document understanding , author=. Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual , pages =. 2021 , url =

2021

[57] [57]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , publisher =

Selfdoc: Self-supervised document representation learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , publisher =. 2021 , url =

2021

[58] [58]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I , pages=

Lambert: Layout-aware language modeling for information extraction , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I , pages=. 2021 , organization=

2021

[59] [59]

Proceedings of the 29th ACM International Conference on Multimedia , publisher =

Structext: Structured text understanding with multi-modal transformers , author=. Proceedings of the 29th ACM International Conference on Multimedia , publisher =. 2021 , url =

2021

[60] [60]

Proceedings of the AAAI Conference on Artificial Intelligence , pages=

Bros: A pre-trained language model focusing on text and layout for better key information extraction from documents , author=. Proceedings of the AAAI Conference on Artificial Intelligence , pages=. 2022 , url =

2022

[61] [61]

Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

InfoVQA: Visual Question Answering on Infographics with A Multi-Modal Entity Graph , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , year=

[62] [62]

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics (ACL) , year=

Compositional Semantic Parsing on Semi-Structured Tables , author=. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics (ACL) , year=

[63] [63]

International Conference on Learning Representations (ICLR) , year=

TabFact: A Large-scale Dataset for Table-based Fact Verification , author=. International Conference on Learning Representations (ICLR) , year=

[64] [64]

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

2022

[65] [65]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

TextVQA: Visual Question Answering with Reading , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

[66] [66]

Proceedings of the European Conference on Computer Vision (ECCV) , year=

TextCaps: A Dataset for Image Captioning with Reading , author=. Proceedings of the European Conference on Computer Vision (ECCV) , year=

[67] [67]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=

Going full-tilt boogie on document understanding with text-image-layout transformer , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=. 2021 , url =

2021

[68] [68]

Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXVIII , pages=

Ocr-free document understanding transformer , author=. Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXVIII , pages=. 2022 , organization=

2022

[69] [69]

arXiv preprint arXiv:2209.08569 , year=

ERNIE-mmLayout: Multi-grained MultiModal Transformer for Document Understanding , author=. arXiv preprint arXiv:2209.08569 , year=

Pith/arXiv arXiv

[70] [70]

IEEE Transactions on Multimedia , year=

Multimodal pre-training based on graph attention network for document understanding , author=. IEEE Transactions on Multimedia , year=

[71] [71]

Information Sciences , volume=

Document image layout analysis via explicit edge embedding network , author=. Information Sciences , volume=. 2021 , publisher=

2021

[72] [72]

Proceedings of the 30th ACM International Conference on Multimedia , pages=

Dit: Self-supervised pre-training for document image transformer , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=. 2022 , publisher =

2022

[73] [73]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=

VSR: a unified framework for document layout analysis combining vision, semantics and relations , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=. 2021 , organization=

2021

[74] [74]

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

LayoutReader: Pre-training of Text and Layout for Reading Order Detection , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=. 2021 , publisher =

2021

[75] [75]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=

ViBERTgrid: a jointly trained multi-modal 2D document representation for key information extraction from documents , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=. 2021 , organization=

2021

[76] [76]

StructuralLM: Structural Pre-training for Form Understanding , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=. 2021 , url =

2021

[77] [77]

arXiv preprint arXiv:2009.14457 , year=

Towards a multi-modal, multi-task learning based pre-training framework for document representation learning , author=. arXiv preprint arXiv:2009.14457 , year=

Pith/arXiv arXiv 2009

[78] [78]

Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=

LayoutParser: A unified toolkit for deep learning based document image analysis , author=. Document Analysis and Recognition--ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part I 16 , pages=. 2021 , organization=

2021

[79] [79]

Document Analysis and Recognition--ICDAR 2021 Workshops: Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=

Mtl-foun: A multi-task learning approach to form understanding , author=. Document Analysis and Recognition--ICDAR 2021 Workshops: Lausanne, Switzerland, September 5--10, 2021, Proceedings, Part II 16 , pages=. 2021 , organization=

2021

[80] [80]

International Conference on Neural Information Processing , pages=

Improving Document Image Understanding with Reinforcement Finetuning , author=. International Conference on Neural Information Processing , pages=. 2022 , organization=

2022