Pith. sign in

REVIEW 4 major objections 5 minor 29 references

DocAnnot auto-annotates key-information documents with F1 up to 0.846, and models trained only on its outputs reach F1 0.686 on CORD.

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 ↗

DocAnnot combines an LVLM, OCR, and a spatial matching heuristic to auto-annotate KIE documents at F1 0.68–0.85, and models trained on that data reach roughly 0.68 F1 on CORD.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A credible, incremental auto-annotation pipeline for KIE with a real specification gap in the SICM spatial step; worth a serious referee but needs fixes. the 4 major comments →

arxiv 2607.24745 v1 pith:V3HO4VUB submitted 2026-05-08 cs.IR cs.AIcs.CLcs.CV

DocAnnot -- Accelerating the Creation of Key Information Extraction Datasets with GenAI-Powered Auto-annotation

Siddartha Reddy , Harikrishnan P M , Goutham Vignesh , Varun V , Vishal Vaddina This is my paper
classification cs.IR cs.AIcs.CLcs.CV
keywords Key Information ExtractionAuto-annotationLarge Vision Language ModelOCRSpatial MatchingSICMCORDSROIE
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 reading

The paper claims that KIE training data can be produced automatically, without human labeling, by combining a large vision-language model's reading of a document with OCR bounding boxes and a spatial matching step. The result is a labeling pipeline that reaches 0.679 F1 on CORD receipts and 0.846 on SROIE, and models fine-tuned exclusively on its output (LayoutLMv3, UDOP) land near 0.68 F1 on CORD. A sympathetic reading: the gap to human-annotated training (0.91–0.92 F1) is real, but the method cuts annotation cost enough that a reviewer can refine outputs rather than label from scratch. The claim matters because document-understanding models are bottlenecked by expensive annotation, and the paper shows a spatial-matching step doing measurable work over pure string matching.

Core claim

The central claim is that a three-stage pipeline — LVLM extraction of label-value pairs, OCR of text and bounding boxes, and a Spatially Informed Contextual Matching (SICM) step — can replace manual annotation for KIE datasets. SICM first filters candidate OCR spans by textual similarity to the extracted value, then disambiguates repeated values by Euclidean distance from the key text's bounding box. Evaluated at token level on CORD and SROIE, the pipeline achieves F1 0.679 and 0.846 with one LVLM, and the ablation shows SICM contributes +0.108 F1 on CORD over string matching alone. As a direct corollary, models trained only on these auto-annotations reach F1 0.6765 (LayoutLMv3) and 0.6861 (

What carries the argument

The Spatially Informed Contextual Matching (SICM) algorithm is the load-bearing component. It takes the LVLM's output — label: {key_text: value} — and the OCR's list of (text, bounding box) pairs. Step 1 scores candidate spans by textual similarity (Levenshtein) to the value and keeps a filtered set. Step 2, when the same value appears multiple times, computes Euclidean distance between each candidate's bounding-box center and the bounding box of the key text (e.g., 'SUBTOTAL') as reported by the LVLM, and picks the closest. The work it does is spatial disambiguation: it ties semantics to geometry, which pure string matching cannot do.

Load-bearing premise

The pipeline assumes the LVLM returns the key text exactly as it appears in the document, so that the OCR has a bounding box for that key; if the LVLM paraphrases or the OCR text differs, the spatial reference is undefined and SICM's gain collapses toward string matching.

What would settle it

Take a sample of receipts, prompt the LVLM to paraphrase keys ('Total Amount' instead of 'TOTAL'), run DocAnnot, and compare F1 to the unmodified version; if F1 falls to the no-SICM baseline (0.571 on CORD), that confirms the spatial step depends on exact key-text grounding.

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

If this is right

  • If the framework's F1 numbers hold, KIE dataset creation can shift from labeling every token to an LVLM pass plus reviewer refinement, which is cheaper per document.
  • Downstream models (LayoutLMv3, UDOP) trained exclusively on auto-annotations score within about 0.24 F1 of their human-supervised versions on CORD, suggesting noisy auto-labels are a viable starting point for prototyping.
  • SICM's measured gain over string matching (F1 0.679 vs 0.571 on CORD) implies that spatial proximity resolves a substantial share of label-value ambiguity in receipt-like layouts.
  • Hybrid training with 10–30% human labels lifts UDOP from 0.686 to 0.699–0.724, indicating that small amounts of human supervision can close part of the gap.

Where Pith is reading between the lines

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

  • Extension: SICM's reliance on the LVLM returning exact key text means the method's ceiling is set by the LVLM's faithfulness to the printed string. If keys are paraphrased or normalized, the spatial anchor disappears and the gain over string matching should vanish; a testable prediction is that F1 on CORD drops toward 0.571 when keys are paraphrased.
  • Extension: The same spatial-disambiguation idea could be applied to other geometry-rich annotation tasks, such as form understanding or tables, wherever repeated values need to be tied to nearby keys via bounding boxes.
  • Extension: The paper measures annotation quality at token level; a downstream-aware evaluation (e.g., end-task accuracy on a held-out document set) would show whether the 0.68-F1 auto-labels are good enough for production workflows.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper presents DocAnnot, a framework for automatically creating KIE training datasets. An LVLM extracts label-value pairs as {label: {key text: value}}; OCR supplies text spans with bounding boxes; the proposed SICM algorithm combines textual similarity with spatial proximity to assign labels to OCR segments. The authors report auto-annotation F1 scores of 0.679 on CORD and 0.846 on SROIE with Claude Sonnet 3.5, and show that LayoutLMv3 and UDOP fine-tuned on the auto-annotated data reach 0.6765 and 0.6861 F1 on CORD, respectively. An ablation attributes substantial gains to SICM, and a hybrid experiment shows modest improvements when 10-30% human-annotated samples are added.

Significance. If the method is fully specified and reproducible, the contribution is useful: it offers a practical pipeline for reducing manual annotation effort, evaluated on public benchmarks against human ground truth. The headline annotation evaluation is not circular — it is measured against external human labels — and the downstream fine-tuning experiments are a genuine student-teacher test. The paper also provides a direct ablation isolating SICM. The main significance, however, rests on a method whose spatial-matching component is not completely described, on a small number of unreplicated runs, and on one internally inconsistent numerical claim. The work does not ship code or data, but the proposed framework is sufficiently interesting to merit further scrutiny.

major comments (4)
  1. [Sections 3.2.1 and 3.2.3] The provenance of the key-text bounding box b_k is undefined. Section 3.2.1 defines the LVLM output as {label: {'actual key text in document': value}} with text only, while Section 3.2.3 states that the extracted key text t_k is 'accompanied by its bounding box b_k and obtained during the LVLM stage.' No mechanism in the pipeline produces a bounding box from the LVLM. The only consistent reading is that b_k is obtained by looking up t_k in the OCR output T_B. That lookup fails under paraphrase, case/punctuation normalization, or OCR errors (e.g., 'TOTAL' vs 'Total Amount'), and the paper does not state the matching policy or the failure rate. Since Table 3 attributes a CORD F1 gain of 0.571 to 0.679 to SICM, this gap is load-bearing. Please specify exactly how b_k is computed, what normalization is applied, and how ambiguous or failed lookups are handled.
  2. [Section 3.2.3, Step 1] The textual similarity filter is under-specified: the paper says candidates are filtered 'e.g., by selecting those exceeding a predefined similarity threshold or by choosing the top-N candidates,' but no threshold or N is reported. This is a free parameter that controls the candidate set entering the spatial disambiguation step, and therefore can directly affect the reported SICM gain. Please report the exact filter configuration used for the CORD and SROIE experiments and, ideally, a sensitivity analysis over the threshold/top-N.
  3. [Section 4.2.2, Table 2 and text] There is a direct numerical inconsistency. Table 2 reports UDOP Auto-Annotated F1 as 0.6861, but the surrounding text states that 'multiple fine-tuning runs (ranging from 0.669 to 0.681) confirm that this behavior is consistent.' The stated range excludes the reported 0.6861. If 0.6861 is the best of several runs, that should be stated; if it is the average, the range is wrong or the sample size is needed. This matters because the paper uses this score to claim that a student trained on noisy auto-annotations can slightly exceed the annotation source F1. Please clarify the run count, the selection rule, and report mean ± standard deviation.
  4. [Sections 4.1.2 and 5.1, Tables 1 and 3] All annotation and ablation F1 scores are based on single runs with no error bars or repeated prompts. Given the stochasticity of LVLMs, the difference attributed to SICM (e.g., CORD F1 0.571 to 0.679) could fall within run-to-run variability. Please report the number of repeated evaluations, the variance, or confidence intervals for the auto-annotation and ablation experiments. This is not a claim of circularity; it is a reproducibility requirement for the central quantitative claims.
minor comments (5)
  1. [Section 2.2 and Section 3.2 header] The algorithm is called SICM in most of the paper but 'SCIM' appears in Section 2.2 and in the Section 3.2 header. Please fix the typo.
  2. [Section 4.1.1, Evaluation Metrics] The token-level F1 metric is not fully defined for labels whose values span multiple tokens or occur multiple times in one document. Please specify how partial matches and repeated values are aligned with the ground truth.
  3. [Section 5.1, Table 3] The 'without SICM' baseline is described as using 'fuzzy string matching,' while Section 3.2.3 mentions Levenshtein distance. Please state the exact similarity measure and threshold used in the baseline so the ablation is reproducible.
  4. [Section 5.2 and Conclusion] The conclusion states that hybrid annotation achieves 'F1 scores comparable to human-supervised training,' but Table 2 shows UDOP on human-annotated data reaches 0.921, while the best hybrid result in Table 4 is 0.7241. This wording is an overstatement and should be revised.
  5. [References] Reference [1] is formatted as a bare URL inside a bibliography entry rather than as a standard citation. Please provide a complete bibliographic entry for the Claude model family.

Circularity Check

0 steps flagged

No circular derivation: headline metrics are benchmarked against external human ground truth; the central pipeline is independent of its inputs, though one self-citation is present and one claimed bbox provenance is underspecified.

full rationale

The core claims — auto-annotation F1 (0.679 CORD, 0.846 SROIE) and downstream fine-tuning performance of LayoutLMv3/UDOP on auto-annotated data — are evaluated against external human ground-truth labels and a held-out test set, so they do not reduce by construction to the framework's own outputs. The downstream experiment is a genuine student–teacher transfer test: models are trained on DocAnnot's labels and evaluated on human-annotated CORD test data, not on the same auto-annotated labels. The SICM ablation also compares two variants side by side, so its claimed gain is an empirical result rather than an identity. No self-definitional equation, fitted parameter renamed as prediction, or imported uniqueness theorem is present. The only self-citation is Ref. [9] ('Reddy, S.'), used as general related-work support for LLM-based annotation acceleration; it is not load-bearing for DocAnnot's main derivation. Two non-circular risks are worth noting: (1) Section 3.2.3 asserts that the key-text bounding box bk is "obtained during the LVLM stage," while Section 3.2.1's LVLM output mapping contains only text ({label:{actual key text:value}}) with no coordinates; this is an under-specification or potential internal inconsistency, but it is not a circular reduction. (2) The future-work note acknowledges possible LVLM pretraining overlap with benchmarks, which is a data-contamination risk rather than circularity. These concerns affect completeness and external validity, not the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

No new physical or mathematical entities are introduced. SICM introduces a bbox-center nearest-neighbor heuristic plus an undisclosed similarity threshold, which are free parameters/ad hoc assumptions rather than independent evidence.

free parameters (1)
  • SICM textual similarity threshold / top-N = not reported
    Section 3.2.3 filters OCR candidates using a 'predefined similarity threshold' or top-N; no value or selection method is given, and it directly affects assignment accuracy.
axioms (4)
  • domain assumption The LVLM reliably outputs the literal key text as it appears in the document, allowing a bounding box to be located in OCR output.
    SICM's spatial step needs the key text's bbox bk, but Section 3.2.1 only has LVLM return key_text:value; the paper never specifies how bk is obtained or what happens if key text is paraphrased.
  • domain assumption The benchmark-provided OCR text and bounding boxes are accurate enough for matching.
    CORD/SROIE OCR data is used directly as positional ground truth; OCR errors would propagate to SICM matching.
  • domain assumption Proprietary LVLM outputs (Claude Sonnet 3.5, Gemini 1.5 Pro) are stable and reproducible.
    All results depend on black-box API models; no versions or seeds are pinned, and outputs can change over time.
  • ad hoc to paper Nearest Euclidean distance between bounding-box centers correctly disambiguates repeated values.
    This is the paper's core heuristic; no validation shows bbox-center distance is the best spatial signal, and layouts with columns or multi-column receipts may break it.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DocAnnot -- Accelerating the Creation of Key Information Extraction Datasets with GenAI-Powered Auto-annotation." pith.science (2026). https://pith.science/paper/V3HO4VUB

@misc{pith2026260724745,
  author       = {Pith},
  title        = {Pith review of: DocAnnot -- Accelerating the Creation of Key Information Extraction Datasets with GenAI-Powered Auto-annotation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V3HO4VUB}},
  note         = {Machine review of arXiv:2607.24745}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Key Information Extraction (KIE) is vital for many document applications, but creating training datasets is traditionally a time-consuming manual process. We introduce DocAnnot, a framework that significantly accelerates KIE dataset generation. DocAnnot leverages a Large Vision Language Model (LVLM) for label value extraction, OCR for text/bounding box detection, and a novel Spatially Informed Contextual Matching (SICM) algorithm. SICM improves label-value association by combining spatial relationships and proximity analysis with textual matching. We evaluate our framework on the CORD and SROIE benchmarks, demonstrating its ability to auto-generate annotations with F1-scores of 0.679 and 0.846, respectively. Furthermore, we investigate the effectiveness of using auto-annotated data for fine-tuning downstream KIE models. While human-annotated data remains superior, models trained exclusively on DocAnnot's outputs attain respectable performance (e.g., LayoutLMv3 achieving an F1-score of 0.6765 on CORD). These results show that while our framework significantly reduces reliance on manual effort, it does not yet fully eliminate the need for human intervention. However, by automating the process to a point where reviewers can efficiently refine outputs, our system enables near-perfect annotations with much greater efficiency than manual annotation from scratch. This approach offers substantial time and cost savings, making it valuable for resource-constrained settings and rapid model prototyping.

Figures

Figures reproduced from arXiv: 2607.24745 by Goutham Vignesh, Harikrishnan P M, Siddartha Reddy, Varun V, Vishal Vaddina.

Figure 1
Figure 1. Figure 1: The DocAnnot framework for automated KIE dataset generation. In [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Fine-tuning a KIE model using the output of the DocAnnot framework. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Disambiguation using bounding box proximity. The central element ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] 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 · 1 canonical work pages

  1. [1]

    org/CorpusID:268232499

    The claude 3 model family: Opus, sonnet, haiku.https://api.semanticscholar. org/CorpusID:268232499

  2. [2]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  3. [3]

    Appalaraju, S., Tang, P., Dong, Q., Sankaran, N., Zhou, Y., Manmatha, R.: Doc- formerv2: Local features for document understanding (2023),https://arxiv.org/ abs/2306.01733

    Pith/arXiv arXiv 2023

  4. [4]

    In: Proceedings of the AAAI conference on artificial intelligence

    Appalaraju, S., Tang, P., Dong, Q., Sankaran, N., Zhou, Y., Manmatha, R.: Doc- formerv2: Local features for document understanding. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 709–718 (2024)

    2024

  5. [5]

    5-vl technical report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., et al.: Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  6. [6]

    Bensch, O., Popa, M., Spille, C.: Key information extraction from documents: Evaluation and generator (2021),https://arxiv.org/abs/2106.14624

    Pith/arXiv arXiv 2021

  7. [7]

    Cao, P., Wang, Y., Zhang, Q., Meng, Z.: Genkie: Robust generative multimodal documentkeyinformationextraction.In:ConferenceonEmpiricalMethodsinNat- ural Language Processing (2023),https://api.semanticscholar.org/CorpusID: 264487294

    2023

  8. [8]

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding (2019),https://arxiv.org/ abs/1810.04805

    Pith/arXiv arXiv 2019

  9. [9]

    In: Machine Learning for Health (ML4H)

    Goel, A., Gueta, A., Gilon, O., Liu, C., Erell, S., Nguyen, L.H., Hao, X., Jaber, B., Reddy, S., Kartha, R., et al.: Llms accelerate annotation for medical information extraction. In: Machine Learning for Health (ML4H). pp. 82–100. PMLR (2023)

    2023

  10. [10]

    Neural Comput.9(8), 1735–1780 (Nov 1997).https://doi.org/10.1162/neco.1997.9.8.1735,https: //doi.org/10.1162/neco.1997.9.8.1735

    Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput.9(8), 1735–1780 (Nov 1997).https://doi.org/10.1162/neco.1997.9.8.1735,https: //doi.org/10.1162/neco.1997.9.8.1735

    doi:10.1162/neco.1997.9.8.1735 1997

  11. [11]

    In: Proceedings of the 30th ACM International Conference on Multimedia

    Huang, Y., Lv, T., Cui, L., Lu, Y., Wei, F.: Layoutlmv3: Pre-training for docu- ment ai with unified text and image masking. In: Proceedings of the 30th ACM International Conference on Multimedia. p. 4083–4091. MM ’22, Association for Computing Machinery, New York, NY, USA (2022).https://doi.org/10.1145/ 3503161.3548112,https://doi.org/10.1145/3503161.354...

    arXiv 2022

  12. [12]

    In: 2019 International Conference on Document Analysis and Recognition (ICDAR)

    Huang, Z., Chen, K., He, J., Bai, X., Karatzas, D., Lu, S., Jawahar, C.V.: Ic- dar2019 competition on scanned receipt ocr and information extraction. In: 2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE (Sep 2019).https://doi.org/10.1109/icdar.2019.00244,http://dx.doi.org/ 10.1109/ICDAR.2019.00244

    arXiv 2019

  13. [13]

    Jaume, G., Ekenel, H.K., Thiran, J.P.: Funsd: A dataset for form understanding in noisy scanned documents (2019),https://arxiv.org/abs/1905.13538

    Pith/arXiv arXiv 2019

  14. [14]

    In: Al-Onaizan, Y., Bansal, M., Chen, Y.N

    Kim, S., Liao, H., Appalaraju, S., Tang, P., Tu, Z., Satzoda, R.K., Manmatha, R., Mahadevan, V., Soatto, S.: DocKD: Knowledge distillation from LLMs for open- world document understanding models. In: Al-Onaizan, Y., Bansal, M., Chen, Y.N. (eds.) Proceedings of the 2024 Conference on Empirical Methods in Nat- ural Language Processing. pp. 3167–3193. Associ...

    doi:10.18653/v1/2024 2024

  15. [15]

    Advances in neural information processing systems36, 34892–34916 (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023)

    2023

  16. [16]

    Mathew, M., Karatzas, D., Jawahar, C.V.: Docvqa: A dataset for vqa on document images (2021),https://arxiv.org/abs/2007.00398

    Pith/arXiv arXiv 2021

  17. [17]

    Müller,R.,Kornblith,S.,Hinton,G.E.:Whendoeslabelsmoothinghelp?Advances in neural information processing systems32(2019)

    2019

  18. [18]

    In: International Conference on Document Analysis and Recognition

    Naparstek, O., Azulai, O., Shapira, I., Amrani, E., Yaroker, Y., Burshtein, Y., Pony, R., Rubinstein, N., Dahood, F.A., Prince, O., et al.: Kvp10k: A comprehen- sive dataset for key-value pair extraction in business documents. In: International Conference on Document Analysis and Recognition. pp. 97–116. Springer (2024)

    2024

  19. [19]

    O’Shea, K., Nash, R.: An introduction to convolutional neural networks (2015), https://arxiv.org/abs/1511.08458

    Pith/arXiv arXiv 2015

  20. [20]

    International Journal of Innovative Technology and Exploring Engineering8, 1613–1617 (07 2019).https://doi.org/10.35940/ijitee.I8156

    Panda, S., Behera, V., Pradhan, A., Mohanty, A.: A rule-based information ex- traction system. International Journal of Innovative Technology and Exploring Engineering8, 1613–1617 (07 2019).https://doi.org/10.35940/ijitee.I8156. 078919

    work page doi:10.35940/ijitee.i8156 2019

  21. [21]

    In: Workshop on Document Intelli- gence at NeurIPS 2019 (2019),https://openreview.net/forum?id=SJl3z659UH

    Park, S., Shin, S., Lee, B., Lee, J., Surh, J., Seo, M., Lee, H.: {CORD}: A consol- idated receipt dataset for post-{ocr} parsing. In: Workshop on Document Intelli- gence at NeurIPS 2019 (2019),https://openreview.net/forum?id=SJl3z659UH

    2019

  22. [22]

    Rouzegar, H., Makrehchi, M.: Enhancing text classification through llm-driven ac- tive learning and human annotation (2024),https://arxiv.org/abs/2406.12114

    Pith/arXiv arXiv 2024

  23. [23]

    In: International Conference of the Cross- LanguageEvaluationForumforEuropeanLanguages.pp.105–117.Springer(2022)

    Skalick` y, M., Šimsa, Š., Uřičář, M., Šulc, M.: Business document information ex- traction: Towards practical benchmarks. In: International Conference of the Cross- LanguageEvaluationForumforEuropeanLanguages.pp.105–117.Springer(2022)

    2022

  24. [24]

    Tang, Z., Yang, Z., Wang, G., Fang, Y., Liu, Y., Zhu, C., Zeng, M., Zhang, C., Bansal, M.: Unifying vision, text, and layout for universal document processing (2023),https://arxiv.org/abs/2212.02623

    Pith/arXiv arXiv 2023

  25. [25]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., Millican, K., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  26. [26]

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention is all you need (2023),https://arxiv.org/abs/1706. 03762

    2023

  27. [27]

    Xu, Y., Xu, Y., Lv, T., Cui, L., Wei, F., Wang, G., Lu, Y., Florencio, D., Zhang, C., Che, W., Zhang, M., Zhou, L.: Layoutlmv2: Multi-modal pre-training for visually- rich document understanding (2022),https://arxiv.org/abs/2012.14740 DocAnnot 15

    Pith/arXiv arXiv 2022

  28. [28]

    In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

    Xu, Y., Li, M., Cui, L., Huang, S., Wei, F., Zhou, M.: Layoutlm: Pre-training of text and layout for document image understanding. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. p. 1192–1200. KDD ’20, ACM (Aug 2020).https://doi.org/10.1145/ 3394486.3403172,http://dx.doi.org/10.1145/3394486.3403172

    arXiv 2020

  29. [29]

    Zhang, Q., Huang, V.S.J., Wang, B., Zhang, J., Wang, Z., Liang, H., Wang, S., Lin, M., He, C., Zhang, W.: Document parsing unveiled: Techniques, challenges, and prospects for structured information extraction (2024),https://arxiv.org/ abs/2410.21169

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.