Digitization of Document and Information Extraction using OCR

Rasha Sinha; Rekha B S

REVIEW 5 major objections 5 minor 21 references

Digitization of Document and Information Extraction using OCR

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A hybrid OCR-plus-LLM pipeline extracts structured key-value fields from both scanned and digital documents, gaining flexibility over template-based methods.

desk verdict A clear system write-up with no experimental evidence for its central claim; desk-reject as a research submission. read the letter →

arxiv 2506.11156 v1 pith:DLVQAEOQ submitted 2025-06-11 cs.CV cs.IR

Rasha Sinha , Rekha B S This is my paper · ORCID

classification cs.CVcs.IR

keywords TextExtractionOpticalCharacterRecognitionDocumentParsingLargeLanguageModelsScannedDocumentsDigitalStructuredDataKey-ValuePair

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Retrieving accurate details from documents is hard when the inputs are a mix of scanned images and native digital files with different layouts. This paper proposes a hybrid framework that splits the problem: OCR engines handle scanned inputs, layout-aware parsers handle digital files, and a large language model performs the final semantic step of turning raw text into structured key-value pairs with confidence scores. The paper argues that this approach beats rule-based and template-based systems because it generalises across document categories without reconfiguration. It also reports a comparative evaluation of Tesseract, DocTR, and Google Vision, and claims the LLM stage produces JSON outputs with very few false positives. A sympathetic reader would care because the framework promises to lower the cost of digitising heterogeneous document collections in business and research settings.

What carries the argument

The two-branch extraction pipeline is the central mechanism: scanned inputs are preprocessed and passed to an OCR engine, native digital files are read by layout-aware parsers, and both streams converge into raw text that a large language model turns into structured key-value pairs in a specified JSON schema. The LLM prompt is what supplies the semantic flexibility the paper contrasts with rule- and template-based systems, and the confidence indicator attached to each field is what makes the output usable for downstream automation. The OCR comparison table supplies the quantitative ordering of engines, while the digital-branch claim of near-100% extraction rests on the presence of an embedded text layer and preserved structural metadata.

What would settle it

Run the described pipeline on a labeled slice of FUNSD and SROIE, measure word-level OCR accuracy and field-level extraction F1 for Tesseract, DocTR, and Google Vision, and compare the LLM-extracted JSON fields against the ground-truth labels. If the measured OCR accuracies deviate substantially from the reported 85%, 91%, and 94%, or if the LLM stage does not beat a simple template heuristic on the same fields, the central comparative claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a cohesive document-understanding pipeline can be assembled from off-the-shelf components: scanned inputs are preprocessed with binarization and skew correction, then passed to an OCR engine (Tesseract, DocTR, or Google Vision API), while native digital documents are read directly with pdfplumber, PyMuPDF, or python-docx. The raw text from either branch is then sent to a large language model such as GPT-4 or Mistral, prompted to identify key-value pairs (name, date of birth, address, total amount) and output them as JSON with confidence indicators. The paper reports that Google Vision reaches about 94% OCR accuracy on the mixed test set, DocTR about 91%, Tesseract about 85%, that digital parsing stays near 100%, and that the LLM stage produces very few false positives. On the strength of these numbers, the paper concludes that the hybrid approach achieves greater flexibility and semantic precision than rule- and template-based methods across scanned and digital document categories.

Load-bearing premise

The load-bearing premise is that the accuracy figures for Tesseract, DocTR, Google Vision, and the near-100% digital parsing confidence are real measured outcomes from a defined test set, rather than illustrative numbers written to support the conclusion.

Editorial extensions

If this is right

A single pipeline can ingest both scanned images and native PDFs or DOCX files, so organisations do not need separate extraction systems for archival versus born-digital documents.
The use of LLMs for key-value extraction means field formats can vary without rewriting rules, and the JSON-schema output plugs directly into downstream databases and indexing systems.
For digital documents, the parser branch avoids OCR errors entirely, keeping extraction confidence near 100%.
The comparative OCR data gives deployers a practical trade-off: Google Vision for maximum accuracy on complex layouts, DocTR for a free alternative that handles PDFs, and Tesseract as an open-source baseline.
If the reported accuracy holds, the approach would outperform rule-based and template-based extraction on noisy and varied document collections.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the framework attaches confidence scores, one natural extension is a rejection rule that flags low-confidence extractions for human review, which would fit regulated document workflows.
An ablation that runs the same OCR output through a template extractor versus the LLM prompt would isolate how much of the claimed flexibility comes from the semantic stage alone.
The reported ordering of OCR engines (Google Vision above DocTR above Tesseract) could be stress-tested on multilingual and heavily degraded scans, which the paper itself names as future work.
Field-level F1 on the named FUNSD and SROIE subsets, rather than word-level OCR accuracy, would measure the end-to-end extraction quality that the pipeline promises.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Rasha and Rekha have written a clear, well-organized description of a document-extraction pipeline: OCR for scanned documents, layout parsers for digital PDFs, then an LLM prompt step that outputs key-value pairs as JSON. As a project write-up it is readable, and the choices are all reasonable. But as a research paper it does not support its central claim. The abstract says the approach is significantly better than rule-based and template-based methods, yet no rule-based or template-based system is implemented or measured anywhere in the text.

The paper gets some things right. The architecture is straightforward, the OCR comparison covers Tesseract, DocTR, and Google Vision, and the discussion of digital-document parsing (embedded text, structural metadata) is sensible. Those parts could be useful to a practitioner who wants a starting point.

The problems are in the evaluation. Section V lists datasets (FUNSD, SROIE, custom invoices, government docs, papers) but gives no counts, no splits, no annotation rules, and no metric definitions. Table 2 reports ~85%, ~91%, ~94% accuracy, but the unit is undefined – character, word, or field level? – and there are no error bars or sample sizes. The claim that digital parsing confidence 'stays steadily near 100%' is asserted, not measured. The LLM extraction 'great accuracy with very few false positives' has no precision, recall, or F1 numbers. The reference list compounds the problem: several literature-review paragraphs describe papers that don't match the cited references, and [13] and [20] are duplicates, as are [12] and [17]. That sloppiness makes it hard to trust the accuracy numbers as anything other than illustrative.

The load-bearing defect is straightforward: the headline claim about beating templates has no experimental support. Without a measured baseline and a defined test set, the paper cannot be verified. The authors mention F1-score in the conclusion but never report it. A serious referee could spend a full review asking for numbers that simply aren't there.

Who is this for? Possibly a course project or internal technical report. It does not add anything new for the OCR or document-AI community. I would not bring it to reading group, would not cite it, and would not send it to peer review. If the authors add a real evaluation with a fixed corpus, ground truth, and one template baseline, a short workshop paper could be plausible.

Referee Report

5 major / 5 minor

Summary. The paper proposes a document digitization and information extraction pipeline that combines OCR for scanned documents, layout-aware parsers for digital documents, and LLMs for semantic key-value extraction. The stated contributions are a comparative evaluation of Tesseract, DocTR, and Google Vision API; a hybrid pipeline with preprocessing; and LLM-based structured output with confidence scores. The abstract claims significant improvements over rule-based and template-based methods. The experimental sections report accuracy numbers for the OCR engines and assert near-perfect extraction for digital documents and high LLM extraction accuracy with few false positives. The manuscript is a high-level system description without an implementation, dataset protocol, or baseline evaluation.

Significance. If the claimed pipeline were validated, it would be a useful applied contribution: an OCR-plus-LLM architecture for flexible structured extraction from heterogeneous documents, with a practical comparison of OCR engines. The paper also identifies relevant tools and a reasonable pipeline decomposition. However, the result is currently not supported by the evidence in the manuscript. There is no reproducible code, no dataset description with counts or annotation rules, no defined accuracy metric, no error analysis, and no baseline rule-based or template-based system. The paper's own conclusion (Section VII) mentions metrics such as word-level OCR accuracy and F1-score that never appear in the reported experiments. The literature review contains numerous citation mismatches and duplicate references, so the prior-art grounding is unreliable. As written, the paper is an architecture sketch with unsupported quantitative claims rather than a validated evaluation.

major comments (5)

[V.A, V.B, Table 2] The OCR comparison in Table 2 reports accuracy values of ~85%, ~91%, and ~94% for Tesseract, DocTR, and Google Vision API, but the accuracy metric is never defined (character-level, word-level, or field-level), the dataset is not described in terms of document counts, splits, or ground truth, and no error bars or per-category breakdowns are provided. Section V.A lists FUNSD, SROIE, a custom invoice set, government documents, and research papers, but gives no annotation protocol, no train/test split, and no document counts. Without a defined measurement protocol, the rows of Table 2 and the corresponding text in Section VI.A cannot be independently verified, so the comparative OCR claim is unsupported.
[VI.C] The claim that 'The LLM-driven method achieved great accuracy with very few false positives, successfully producing results that met a specified JSON format' is not accompanied by any quantitative evaluation: no precision, recall, F1, confusion counts, or sample outputs with ground truth. The preceding paragraph (Section V.D) mentions evaluating the model's ability to adhere to a JSON schema, but no results of such an evaluation are reported. The conclusion in Section VII references 'F1-score for extracting fields' as an assessment metric, yet this metric is never computed or reported anywhere in the manuscript.
[Abstract and Introduction] The central claim of the paper, stated in the abstract and in Section I, is that the approach demonstrates 'significant improvements over traditional rule-based and template-based methods.' No rule-based or template-based extraction system is implemented, measured, or compared at any point in the manuscript. The evaluation in Section V compares OCR engines and reports LLM extraction assertions, but never compares the overall pipeline against a baseline. This is a load-bearing omission: the headline comparative claim has no quantitative support in the present text.
[V.C.2 and VI.B] The assertion that digital document parsing confidence 'stays steadily near 100%' (Section V.C.2) and the corresponding '~100%' accuracy in Section VI.B are stated without any measurement. No extraction errors, no test documents, no metric, and no comparison are provided. Additionally, the system's LLM 'confidence indicators' (Sections I and V.D) appear to be used as evidence of correct extraction, but the paper's own literature review entry [4] warns that LLMs exaggerate their certainty; furthermore, the cited reference [4] in the bibliography does not match that description, so the warning is not actually backed by the listed source.
[II, References] The literature review contains serious citation integrity problems. The descriptions of [1], [2], [4], and [5] in Section II do not correspond to the works actually listed in the references: [1] is described as a study of digitization procedures with Adobe Acrobat but is cited as a TrOCR receipt-recognition paper; [2] is described as an OCR history paper but is cited as a deep-learning OCR model for Old Turkic; [4] is described as an LLM confidence study but is cited as an OCR post-processing NLP pipeline; [5] is described as handwritten exam grading but is cited as a legal-domain document analysis framework. In addition, references [13] and [20] are exact duplicates, and [12] and [17] are duplicates. These errors prevent the reader from locating the claimed prior art and undermine the literature grounding of the proposed approach.

minor comments (5)

[VI, Figures 2 and 3] Figures 2 and 3 are referenced in the results section but no captions, axes labels, or data sources are provided; the figures appear to be placeholders. Figure 1, the pipeline overview, is also referenced but not reproduced in the text.
[V.D] The sentence 'Evaluates the model’s capability to generate structured outputs that adhere to a specified JSON schema' is a sentence fragment rather than a complete sentence and should be integrated into the surrounding text.
[Table 1] In Table 1, the Google Vision API row lists the same strengths as the EasyOCR row ('Better support for handwritten and cursive text'), which appears to be a copy-paste error; Google Vision's stated advantage for complex layouts, mentioned elsewhere, is missing here.
[III.A] Section III.A begins by saying OCR serves as the foundation for converting handwritten exam sheets, but the stated scope of the paper is general document digitization; this inconsistent terminology should be harmonized throughout.
[V.A and VI.B] The names 'FUNSD' and 'SROIE' are standard benchmarks, but the paper does not state which subset or split was used, nor does it report any comparison with published results on those datasets; this should be clarified to allow apples-to-apples evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the paper's unsupported accuracy claims are evidentiary defects, not circular reasoning.

full rationale

No circularity found. The paper contains no derivation chain in the technical sense: there are no equations, no fitted parameters, and no predicted quantity that is constructed from the same data that defines it. The central comparative claim in the abstract (“significant improvements over traditional rule-based and template-based methods”) is unsupported because no rule-based or template-based baseline is implemented or measured, and the accuracy figures in Table 2 (Tesseract ~85%, DocTR ~91%, Google Vision ~94%) and the “near 100%” digital parsing confidence in Section V.C.2 are not tied to a described dataset, metric, or measurement protocol. These are evidentiary and reproducibility defects, not circular reductions: the claims are asserted rather than derived from their own inputs. The paper cites [4] on LLM confidence exaggeration, but it does not use that citation as load-bearing proof of its own accuracy; if anything, the citation weakens the paper’s use of self-reported confidence, which is a validity concern rather than a circular step. There are no self-citations used as justification and no uniqueness theorems or imported ansatze. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper's central accuracy claims rest on unmeasured, hand-selected percentages and on unstated assumptions about OCR tool behavior and LLM reliability. No free parameter is fit to data in a reproducible way; the numbers function as rhetorical support rather than experimental results.

free parameters (4)

OCR accuracy for Tesseract v4 = ~85%
Reported in Table 2 as an evaluation result, but no test set, sample size, or measurement procedure is given; it appears as an illustrative number rather than a measured value.
OCR accuracy for DocTR = ~91%
Same as above, from Table 2.
OCR accuracy for Google Vision API = ~94%
Same as above, from Table 2.
Digital document parsing accuracy = ~100%
Stated in Section V.C.2 and VI.B as inherent to digital PDFs, without any test on varied or image-based PDFs.

assumptions (3)

domain assumption The OCR accuracy values in Table 2 are comparable across engines under identical conditions.
Section V.B describes a comparative analysis but never specifies the dataset, preprocessing, or metric; the comparison is asserted.
domain assumption LLMs can accurately convert OCR raw text into key-value pairs with few false positives.
Section VI.C claims this outcome without reporting precision, recall, or F1, and without giving the prompt schema.
domain assumption Digitally created PDFs always retain selectable text layers enabling near-perfect extraction.
Section V.C.2 claims this as inherent, but scanned PDFs and many born-digital PDFs with embedded images violate it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Digitization of Document and Information Extraction using OCR." pith.science (2026). https://pith.science/paper/DLVQAEOQ

@misc{pith2026250611156,
  author       = {Pith},
  title        = {Pith review of: Digitization of Document and Information Extraction using OCR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLVQAEOQ}},
  note         = {Machine review of arXiv:2506.11156}
}

read the original abstract

Retrieving accurate details from documents is a crucial task, especially when handling a combination of scanned images and native digital formats. This document presents a combined framework for text extraction that merges Optical Character Recognition (OCR) techniques with Large Language Models (LLMs) to deliver structured outputs enriched by contextual understanding and confidence indicators. Scanned files are processed using OCR engines, while digital files are interpreted through layout-aware libraries. The extracted raw text is subsequently analyzed by an LLM to identify key-value pairs and resolve ambiguities. A comparative analysis of different OCR tools is presented to evaluate their effectiveness concerning accuracy, layout recognition, and processing speed. The approach demonstrates significant improvements over traditional rule-based and template-based methods, offering enhanced flexibility and semantic precision across different document categories

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

[13]

KuroNet: Pre-modern Japanese Kuzushiji Character Recognition with Deep Learning,

T. Clanuwat, A. Lamb, and A. Kitamoto, "KuroNet: Pre-modern Japanese Kuzushiji Character Recognition with Deep Learning," in Proc. Int. Conf. Doc. Anal. Recognit. (ICDAR) , 2019

work page 2019
[20]

DeepNetDevanagari: A Deep Learning Model for Devanagari Ancient Character Recognition,

S. R. Narang, M. Kumar, and M. K. Jindal, "DeepNetDevanagari: A Deep Learning Model for Devanagari Ancient Character Recognition," Multimedia Tools Appl. , vol. 80, pp. 20671–20686, 2021

work page 2021
[17]

Deep Learning for Document Image Enhancement,

Y. Yang, S. Sun, W. Li, and J. Wang, "Deep Learning for Document Image Enhancement," IEEE Trans. Image Process. , vol. 28, no. 5, pp. 2420–2435, 2019

work page 2019
[4]

Extending TrOCR for Text Localization-Free OCR of Full-Page Scanned Receipt Images,

H. Zhang, E. Whittaker, and I. Kitagishi, "Extending TrOCR for Text Localization-Free OCR of Full-Page Scanned Receipt Images," in Proc. IEEE/CVF Int. Conf. Comput. Vis. Workshops (ICCVW) , 2023, pp. 1479–1485

work page 2023
[1]

For scanned documents, techniques like binarization and skew correction are utilized to enhance text clarity and improve OCR accuracy

Preprocessing: Input documents receive customized preprocessing according to their source. For scanned documents, techniques like binarization and skew correction are utilized to enhance text clarity and improve OCR accuracy. To prepare and clean the data before extraction, parsing libraries and noise removal algorithms are employed for digital documents....

work page
[2]

● python-docx: Utilized for retrieving text and formatting details from DOCX documents

Instruments Employed: To analyze and retrieve data from digitally generated documents, the subsequent libraries and tools are utilized: ● pdfplumber: Aids in retrieving text from PDFs, including table data and formatting specifics ● PyMuPDF (fitz): Provides detailed access to page content, such as metadata, embedded fonts, and images, with excellent fidel...

work page
[5]

A Deep Learning Based Optical Character Recognition Model for Old Turkic,

S. H. Taheri, H. Kosarirad, I. A. Gallego, and N. Taheri, "A Deep Learning Based Optical Character Recognition Model for Old Turkic," EAI Endorsed Trans. AI Robot. , vol. 4, Apr. 2025

work page 2025
[3]

● Maintained Structural Metadata: Internal components such as paragraphs, headings, and tables improve segmentation and contextual understanding

Fidelity of Extraction: Digital documents facilitate accurate data extraction because of their structural features: ● Text Layer Accessibility: In contrast to raster images, digital PDFs retain text as character objects, which greatly improves extraction precision and removes errors caused by OCR. ● Maintained Structural Metadata: Internal components such...

work page

Show all 21 references

[6]

Image Text Extraction and Natural Language Processing of Unstructured Data from Medical Reports,

I. Malashin et al., "Image Text Extraction and Natural Language Processing of Unstructured Data from Medical Reports," Mach. Learn. Knowl. Extr. , vol. 6, pp. 1361–1377, 2024

2024
[7]

A Novel Pipeline for Improving Optical Character Recognition through Post-processing Using Natural Language Processing,

A. Rakshit, S. Mehta, and A. Dasgupta, "A Novel Pipeline for Improving Optical Character Recognition through Post-processing Using Natural Language Processing," arXiv preprint arXiv:2307.04245 , 2023

2023 arXiv
[8]

TransDocAnalyser: A Framework for Offline Semi-structured Handwritten Document Analysis in the Legal Domain,

S. Chakraborty, G. Harit, and S. Ghosh, "TransDocAnalyser: A Framework for Offline Semi-structured Handwritten Document Analysis in the Legal Domain," arXiv preprint arXiv:2306.02142 , 2023

2023 arXiv
[9]

Handwritten Text Recognition Using Convolutional Neural Network,

A. Mishra, A. S. Ram, and K. C, "Handwritten Text Recognition Using Convolutional Neural Network," arXiv preprint arXiv:2307.05396 , 2023

2023 arXiv
[10]

OCR-free Document Understanding Transformer,

G. Kim et al., "OCR-free Document Understanding Transformer," arXiv preprint arXiv:2111.15664 , 2021

2021 arXiv
[11]

TrOCR: Transformer-based Optical Character Recognition with Pretrained Models,

M. Li et al., "TrOCR: Transformer-based Optical Character Recognition with Pretrained Models," in Proc. AAAI Conf. Artif. Intell. , vol. 37, no. 11, 2023

2023
[14]

Documentation of Old Turkic Runic Inscriptions of the Altai Mountains Using Photogrammetric Technology,

M. V. Vavulin, "Documentation of Old Turkic Runic Inscriptions of the Altai Mountains Using Photogrammetric Technology," Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. , vol. 42, pp. 257–261, 2017

2017
[16]

Historical Document Image Restoration Using Multispectral Imaging System,

R. Hedjam and M. Cheriet, "Historical Document Image Restoration Using Multispectral Imaging System," Pattern Recognit. , vol. 47, no. 6, pp. 2022–2030, 2014

2022
[18]

Recognition of Historical Scripts Using Shallow Convolutional Neural Networks,

A. Jindal and C. Arora, "Recognition of Historical Scripts Using Shallow Convolutional Neural Networks," Int. J. Doc. Anal. Recognit. , vol. 24, no. 2, pp. 89–98, 2021

2021
[19]

VISTA-OCR: Towards Generative and Interactive End-to-End OCR Models,

L. Hamdi, A. Tamasna, P. Boisson, and T. Paquet, "VISTA-OCR: Towards Generative and Interactive End-to-End OCR Models," arXiv preprint arXiv:2504.03621 , 2025

2025 arXiv
[21]

Post-OCR Parsing: Building Simple and Robust Parser via BIO Tagging,

W. Hwang et al., "Post-OCR Parsing: Building Simple and Robust Parser via BIO Tagging," in Workshop on Document Intelligence at NeurIPS , 2019

2019
[22]

Extract Data Points from Invoices with Multi-layer Graph Attention Network and Named Entity Recognition,

C. Zhang et al., "Extract Data Points from Invoices with Multi-layer Graph Attention Network and Named Entity Recognition," in Proc. IEEE Int. Conf. Artificial Intelligence and Computer Applications (ICAICA) , 2022, pp. 1–6

2022
[23]

A Hybrid CNN-Transformer Model for Historical Document Image Binarization,

V. Rezanezhad, K. Baierer, and C. Neudecker, "A Hybrid CNN-Transformer Model for Historical Document Image Binarization," in Proc. 7th Int. Workshop Hist. Doc. Imaging Process. , 2023

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [13]

KuroNet: Pre-modern Japanese Kuzushiji Character Recognition with Deep Learning,

T. Clanuwat, A. Lamb, and A. Kitamoto, "KuroNet: Pre-modern Japanese Kuzushiji Character Recognition with Deep Learning," in Proc. Int. Conf. Doc. Anal. Recognit. (ICDAR) , 2019

work page 2019

[2] [20]

DeepNetDevanagari: A Deep Learning Model for Devanagari Ancient Character Recognition,

S. R. Narang, M. Kumar, and M. K. Jindal, "DeepNetDevanagari: A Deep Learning Model for Devanagari Ancient Character Recognition," Multimedia Tools Appl. , vol. 80, pp. 20671–20686, 2021

work page 2021

[3] [17]

Deep Learning for Document Image Enhancement,

Y. Yang, S. Sun, W. Li, and J. Wang, "Deep Learning for Document Image Enhancement," IEEE Trans. Image Process. , vol. 28, no. 5, pp. 2420–2435, 2019

work page 2019

[4] [4]

Extending TrOCR for Text Localization-Free OCR of Full-Page Scanned Receipt Images,

H. Zhang, E. Whittaker, and I. Kitagishi, "Extending TrOCR for Text Localization-Free OCR of Full-Page Scanned Receipt Images," in Proc. IEEE/CVF Int. Conf. Comput. Vis. Workshops (ICCVW) , 2023, pp. 1479–1485

work page 2023

[5] [1]

For scanned documents, techniques like binarization and skew correction are utilized to enhance text clarity and improve OCR accuracy

Preprocessing: Input documents receive customized preprocessing according to their source. For scanned documents, techniques like binarization and skew correction are utilized to enhance text clarity and improve OCR accuracy. To prepare and clean the data before extraction, parsing libraries and noise removal algorithms are employed for digital documents....

work page

[6] [2]

● python-docx: Utilized for retrieving text and formatting details from DOCX documents

Instruments Employed: To analyze and retrieve data from digitally generated documents, the subsequent libraries and tools are utilized: ● pdfplumber: Aids in retrieving text from PDFs, including table data and formatting specifics ● PyMuPDF (fitz): Provides detailed access to page content, such as metadata, embedded fonts, and images, with excellent fidel...

work page

[7] [5]

A Deep Learning Based Optical Character Recognition Model for Old Turkic,

S. H. Taheri, H. Kosarirad, I. A. Gallego, and N. Taheri, "A Deep Learning Based Optical Character Recognition Model for Old Turkic," EAI Endorsed Trans. AI Robot. , vol. 4, Apr. 2025

work page 2025

[8] [3]

● Maintained Structural Metadata: Internal components such as paragraphs, headings, and tables improve segmentation and contextual understanding

Fidelity of Extraction: Digital documents facilitate accurate data extraction because of their structural features: ● Text Layer Accessibility: In contrast to raster images, digital PDFs retain text as character objects, which greatly improves extraction precision and removes errors caused by OCR. ● Maintained Structural Metadata: Internal components such...

work page