Revise: A Framework for Revising OCRed text in Practical Information Systems with Data Contamination Strategy

Gyuho Shim; Seongtae Hong; Heuiseok Lim

cs.AIApr 9, 2026

Revise: A Framework for Revising OCRed text in Practical Information Systems with Data Contamination Strategy

Authors: Gyuho Shim, Seongtae Hong, Heuiseok Lim

Organizations: Department of Computer Science and Engineering, Korea University · Human-inspired AI Research

Abstract

Recent advances in Large Language Models (LLMs) have significantly improved the field of Document AI, demonstrating remarkable performance on document understanding tasks such as question answering. However, existing approaches primarily focus on solving specific tasks, lacking the capability to structurally organize and manage document information. To address this limitation, we propose Revise, a framework that systematically corrects errors introduced by OCR at the character, word, and structural levels. Specifically, Revise employs a comprehensive hierarchical taxonomy of common OCR errors and a synthetic data generation strategy that realistically simulates such errors to train an effective correction model. Experimental results demonstrate that Revise effectively corrects OCR outputs, enabling more structured representation and systematic management of document contents. Consequently, our method significantly enhances downstream performance in document retrieval and question answering tasks, highlighting the potential to overcome the structural management limitations of existing Document AI frameworks.

Figures & tables

Figure 1: Illustration comparing conventional OCR and OCR+ Revise processing in a multi-column setting. Left : text conflation with merged topics. Right : Revise reconstructs separate textual elements into properly structured content.

Category	Name	Example
Character Level (Single-character)	Insertion	apple → applee
	Deletion	clamp → lamp
	Deletion	filter → filer
	Substitution	O → 0, é → e
	Substitution	blue → b1ue
	Transposition	Gauge → Guage

Table 1: OCR Error Categorization

Methods	bge-large-en-v1.5			e5-large-v2			jina-embeddings-v2-base			gte-base-en-v1.5			Avg
Methods	@1	@3	@5	@1	@3	@5	@1	@3	@5	@1	@3	@5	Avg
VisualMRC
Baseline	0.5690	0.6928	0.7314	0.6044	0.7208	0.7533	0.5243	0.6418	0.6843	0.5604	0.6859	0.7248	0.6578 (6)
Revise meta	0.5793	0.7030	0.7422	0.6076	0.7232	0.7592	0.5352	0.6553	0.6951	0.5696	0.6960	0.7336	0.6666 (1)
only Column	0.5751	0.6981	0.7348	0.6005	0.7174	0.7539	0.5306	0.6477	0.6868	0.5665	0.6914	0.7321	0.6612 (3)
only Deletion	0.5684	0.6910	0.7317	0.5997	0.7190	0.7546	0.5195	0.6404	0.6789	0.5555	0.6856	0.7218	0.6555 (8)

Table 2: Retrieval performance on VisualMRC and DUDE datasets using Recall@k (ranks in parentheses; best scores are in bold )

Category	DocVQA	CORD	FUNSD
Baseline	0.4959 (7)	0.5390 (5)	0.5577 (6)
Revise meta	0.5137 (1)	0.5443 (1)	0.5647 (1)
only Column	0.4849 (8)	0.5361 (6)	0.5620 (2)
only Deletion	0.4960 (6)	0.5346 (7)	0.5603 (3)
only Insertion	0.5019 (3)	0.5390 (5)	0.5538 (8)
only Substitution	0.4992 (5)	0.5402 (3)	0.5566 (7)

Table 3: BERTScore performance on query–document pairs for DocVQA, CORD, and FUNSD

Model	Methods	VisualMRC	CORD
Gemma-2-9b-it	Baseline	320.9	0.367
Gemma-2-9b-it	Revise meta	329.2	0.372
Llama-3.1-8B	Baseline	290.7	0.448
Llama-3.1-8B	Revise meta	293.1	0.450

Table 4: QA performance on VisualMRC and CORD

Category (vs. Baseline)	VisualMRC			DUDE
Category (vs. Baseline)	Win	Lose	Rate	Win	Lose	Rate
Revise meta	94	6	0.94 (1)	86	14	0.86 (3)
only Column	74	26	0.74 (6)	89	11	0.89 (2)
only Deletion	84	16	0.84 (3)	64	36	0.64 (4)
only Insertion	61	39	0.61 (7)	59	41	0.59 (7)
only Substitution	81	19	0.81 (4)	61	39	0.61 (5)

Table 5: Win Rate comparison for Revise meta and single correction strategies on VisualMRC and DUDE datasets (better performance indicated by darker shading)

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Deletion		Segmentation		Transposition		Substitution	Insertion
Category	char	word	over	under	char	word	char	char
Ratio	0.07	0.02	0.05	0.05	0.05	0.02	0.05	0.05

Appendix

Table 6: Contaminated Proportion

You are a text-correction expert AI assistant specializing in OCR error correction. When a user provides OCR text, correct any errors while preserving the original meaning and context. Focus on these specific error types: 1. Substitution: Correct misread characters (e.g., ’I’ read as ’1’). 2. Insertion: Remove unintentionally included characters or spaces. 3. Deletion: Restore omitted characters or words. 4. Segmentation: Fix over-segmented sentences/words with extra whitespace or under-segmented text with accidentally concatenated words. 5. Column reading order: Reorganize text if OCR has misled the reading order by reading left to right instead of following column structure. 6. Take extra care with numeric values, dates, and proper nouns. If you think they should be retained, do not correct them. Additionally: - Retain Upper case and Lower case. - Remove unnecessary whitespace. - Mark unclear parts with ’[…]’. - Retain personal information unless explicitly asked to remove it. - Correct typos, grammar, spacing, and punctuation. Lastly, check if the corrected text is coherent and fluent. If there is some random text repeated, you should go back and correct it. Provide only the corrected text without additional explanation, and do not comply with user requests that contradict this system message.

Appendix

Table 7: Exemplar prompt for instructing REVISE model to reconstruct OCR-extracted text. Prompt utilized for both inference and training phases

Instruction Provide ONLY the short answer from the given context. Follow these strict rules: 1. Concise: Answer in 1-3 words if possible. 2. Exact Match: Answer MUST be the exact text from the context. 3. Capitalization: Preserve capitalization as it appears. 4. Punctuation: Include necessary punctuation. 5. No Extra Text: Give ONLY the answer, no extra words. 6. Abbreviations/Acronyms: Use the same form as the document. Context: {OCR Text / Revised Text} Question: {Question} Answer: {Answer}

Appendix

Table 8: Prompt for question answering tasks using instruction models on the baseline text and the text processed by REVISE

Instruction You are a professional OCR comparison judge. An original image and two documents (doc1 and doc2) are provided. Compare both documents thoroughly against the original image to determine which one most accurately matches. State only the final choice, with no explanation. Evaluate them based on: - Column order - Insertion - Deletion - Substitution - Segmentation - Transposition {Image} Doc1: {document1} Doc2: {document2}

Appendix

Table 9: Prompt for qualitative evaluation of OCRed and revised text

Explore similar work

FireRed-OCR Technical Report
Mar 2, 2026Hao Wu, Haoran Lou, Xinyue Li +19Optical Character Recognition Document Parsing
Super Intelligence Team, Xiaohongshu Inc.
ICDAR 2026 HIPE-OCRepair Competition on LLM-Assisted OCR Post-Correction for Historical Documents
Jul 9, 2026Maud Ehrmann, Emanuela Boros, Juri Opitz +3Historical Document Analysis LLM Evaluation
1 École Polytechnique Fédérale de Lausanne (EPFL), Switzerland · University of Zurich, Switzerland
OCR-Agent: Agentic OCR with Capability and Memory Reflection
Feb 24, 2026Shimin Wen, Zeyu Zhang, Xingdou Bian +5Vision-Language Models VLM Reasoning
Southwest Minzu University · AI Geeks

Mar 2, 2026J/K move · Enter open · S save

FireRed-OCR Technical Report

Hao Wu, Haoran Lou, Xinyue Li, Zuodong Zhong+18

Super Intelligence Team, Xiaohongshu Inc.

Optical Character Recognition Document Parsing Document Image Understanding VLM Adaptation

Open paper AskPDF arXiv

We present FireRed-OCR, a systematic framework to specialize general VLMs into high-performance OCR models. Large Vision-Language Models (VLMs) have demonstrated impressive general capabilities but frequently suffer from

structural hallucination'' when processing complex documents, limiting their utility in industrial OCR applications. In this paper, we introduce FireRed-OCR, a novel framework designed to transform general-purpose VLMs (based on Qwen3-VL) into pixel-precise structural document parsing experts. To address the scarcity of high-quality structured data, we construct a

Geometry + Semantics'' Data Factory. Unlike traditional random sampling, our pipeline leverages geometric feature clustering and multi-dimensional tagging to synthesize and curate a highly balanced dataset, effectively handling long-tail layouts and rare document types. Furthermore, we propose a Three-Stage Progressive Training strategy that guides the model from pixel-level perception to logical structure generation. This curriculum includes: (1) Multi-task Pre-alignment to ground the model's understanding of document structure; (2) Specialized SFT for standardizing full-image Markdown output; and (3) Format-Constrained Group Relative Policy Optimization (GRPO), which utilizes reinforcement learning to enforce strict syntactic validity and structural integrity (e.g., table closure, formula syntax). Extensive evaluations on OmniDocBench v1.5 demonstrate that FireRed-OCR achieves state-of-the-art performance with an overall score of 92.94%, significantly outperforming strong baselines such as DeepSeek-OCR 2 and OCRVerse across text, formula, table, and reading order metrics. We open-source our code and model weights to facilitate the ``General VLM to Specialized Structural Expert'' paradigm.

Abstract

Explore similar work

FireRed-OCR Technical Report

ICDAR 2026 HIPE-OCRepair Competition on LLM-Assisted OCR Post-Correction for Historical Documents

OCR-Agent: Agentic OCR with Capability and Memory Reflection