Hybrid OCR-LLM Framework for Enterprise-Scale Document Information Extraction Under Copy-heavy Task面向企业级文档信息提取的混合OCR-LLM框架：针对高重复性任务的优化

Our extraction framework comprises two primary components: text extraction and LLM-based information extraction. Figure 2 presents an overview of the complete pipeline. Initially, various tools are employed to extract textual content from heterogeneous file formats, including Markdown, Word, Excel, PDF, and images. The extracted text is then passed to a large language model (LLM) to perform task-specific information extraction.我们的提取框架包含两个主要组件：文本提取和基于LLM的信息提取。图2展示了完整流水线的概览。最初，使用各种工具从异构文件格式（包括Markdown、Word、Excel、PDF和图像）中提取文本内容。提取的文本随后被传递给大语言模型（LLM）以执行特定任务的信息提取。

To support different document structures and information needs, we implement multiple extraction strategies. Specifically, our methods fall into three categories, as shown in Figure 3: direct extraction, replacement-based extraction, and table-based extraction, which will be described in detail later. Additionally, we incorporate a multimodal model to directly extract information from images, serving as a baseline for comparison with text-based pipelines.为了支持不同的文档结构和信息需求，我们实现了多种提取策略。具体而言，我们的方法分为三类，如图3所示：直接提取、基于替换的提取和基于表格的提取，稍后将详细描述。此外，我们还结合了一个多模态模型来直接从图像中提取信息，作为与基于文本的流水线进行比较的基准。

The quality of information extraction heavily depends on the accuracy and completeness of upstream text extraction. To support diverse input formats, we integrate a suite of specialized extraction tools, each selected based on its ability to handle specific file types and structural characteristics.信息提取的质量在很大程度上取决于上游文本提取的准确性和完整性。为了支持多样的输入格式，我们集成了一套专门的提取工具，每种工具的选择都基于其处理特定文件类型和结构特征的能力。

MarkItDown serves as a general-purpose parser for structured text formats, supporting Markdown, Word, Excel, and standard PDF documents. It preserves key structural features such as headings, tables, and semantic markers, enabling clean and semantically meaningful text extraction across multiple formats.MarkItDown作为一种通用的结构化文本格式解析器，支持Markdown、Word、Excel和标准PDF文档。它保留了标题、表格和语义标记等关键结构特征，能够在多种格式中实现清晰且具有语义意义的文本提取。

Docling is applied to both Word and PDF files, offering enhanced layout analysis and document hierarchy preservation. It is particularly effective in maintaining reading order and spatial layout, which are essential for downstream structure-aware tasks.Docling应用于Word和PDF文件，提供增强的布局分析和文档层次结构保留。它在保持阅读顺序和空间布局方面特别有效，这对于下游结构感知任务至关重要。

MinerU complements Docling by targeting complex PDF layouts, including multi-column formatting, dense tables, and embedded formulas. It provides fine-grained spatial structure recovery and is optimized for documents where layout fidelity is critical.MinerU通过针对复杂PDF布局（包括多栏格式、密集表格和嵌入式公式）来补充Docling。它提供细粒度的空间结构恢复，并针对布局保真度至关重要的文档进行了优化。

PaddleOCR and EasyOCR are employed for image-based documents. PaddleOCR is used as the primary engine due to its high accuracy in multilingual printed text, while EasyOCR acts as a fallback for cases involving handwritten content or degraded image quality.PaddleOCR和EasyOCR用于基于图像的文档。PaddleOCR因其在多语言印刷文本方面的高准确性而被用作主要引擎，而EasyOCR则作为涉及手写内容或图像质量下降情况下的后备方案。

The use of multiple tools for overlapping file types is intentional. Our goal is to systematically evaluate different extraction solutions on the same document type in order to identify the most effective extraction strategy. Evaluation focuses on two key dimensions: (1) the accuracy of extracted textual content, and (2) the fidelity of spatial and structural information preservation, which is crucial for position-aware downstream tasks such as table extraction or form understanding.为了重叠的文件类型使用多种工具是有意为之的。我们的目标是在同一文档类型上系统地评估不同的提取解决方案，以确定最有效的提取策略。评估侧重于两个关键维度：(1) 提取文本内容的准确性，以及 (2) 空间和结构信息保留的保真度，这对于表格提取或表单理解等位置感知下游任务至关重要。

Once the raw text is extracted, we apply LLM-based methods to retrieve the target information. To accommodate different document layouts and content patterns, we design three complementary extraction strategies, each tailored to specific structural characteristics.一旦提取出原始文本，我们就会应用基于LLM的方法来检索目标信息。为了适应不同的文档布局和内容模式，我们设计了三种互补的提取策略，每种策略都针对特定的结构特征进行了量身定制。

Direct extraction. This method applies LLMs or VLMs to perform end-to-end information extraction. In the text-based variant, raw text obtained from upstream extraction tools is passed to an LLM along with task-specific prompts. The vision-based variant bypasses OCR and directly uses document images as input to multimodal models, enabling better spatial understanding at the cost of higher inference overhead.直接提取。此方法应用LLM或VLM执行端到端的信息提取。在基于文本的变体中，从上游提取工具获得的原始文本与特定任务的提示一起传递给LLM。基于视觉的变体则绕过OCR，直接将文档图像作为输入提供给多模态模型，以更高的推理开销为代价实现了更好的空间理解。

Replace extraction. To address the challenges of repetitive pattern documents, we adopt a two-step approach. First, structured elements, such as identifiers, are replaced with unique placeholders. The LLM is then prompted to retrieve associated fields based on these placeholders. This design improves consistency, reduces ambiguity, and allows efficient batch processing through prompt reuse.替换提取。为了应对重复模式文档的挑战，我们采用了两步法。首先，将标识符等结构化元素替换为唯一占位符。然后，提示LLM根据这些占位符检索相关字段。这种设计提高了一致性，减少了歧义，并允许通过提示复用进行高效的批处理。

Table extraction. For documents with tabular layouts, we combine LLM-based structure recognition with deterministic parsing. The LLM identifies table regions and target cell coordinates, while a rule-based parser extracts the content. This approach minimizes hallucination and reduces generation cost by limiting model output to positional metadata.表格提取。对于具有表格布局的文档，我们将基于LLM的结构识别与确定性解析相结合。LLM识别表格区域和目标单元格坐标，而基于规则的解析器提取内容。这种方法通过将模型输出限制为位置元数据，最大限度地减少了幻觉并降低了生成成本。

To determine the optimal strategy for different document types, we systematically compare the performance of these methods when paired with various text extraction tools. This enables us to identify the most effective combination in terms of both content accuracy and time consumption, ensuring adaptability and robustness in real-world document scenarios.为了确定不同文档类型的最佳策略，我们系统地比较了这些方法与各种文本提取工具配对时的性能。这使我们能够识别在内容准确性和时间消耗方面最有效的组合，从而确保现实文档场景中的适应性和稳健性。

We adopt Qwen2.5-7B for text-based extraction and Qwen2.5-VL-7B for vision-language tasks, based on their strong balance between performance and efficiency. The 7B scale offers competitive extraction accuracy with significantly lower inference cost compared to larger models, enabling practical deployment on a single GPU.我们采用Qwen2.5-7B进行基于文本的提取，采用Qwen2.5-VL-7B进行视觉语言任务，这是基于它们在性能和效率之间取得的良好平衡。7B规模提供了具有竞争力的提取准确性，且推理成本远低于大型模型，从而能够在单块GPU上进行实际部署。

Qwen2.5 models also provide robust multilingual support—particularly for Chinese, Japanese, and Korean—and share a unified architecture across text and vision variants, simplifying prompt design and enabling consistent evaluation across modalities. Their strong instruction-following ability ensures reliable structured outputs (e.g., JSON, placeholders), reducing post-processing overhead. These capabilities make them well-suited for our production-oriented information extraction pipeline.Qwen2.5模型还提供了强大的多语言支持——特别是针对中文、日文和韩文——并在文本和视觉变体之间共享统一的架构，简化了提示设计并实现了跨模态的一致评估。它们强大的指令遵循能力确保了可靠的结构化输出（例如JSON、占位符），减少了后处理开销。这些能力使它们非常适合我们面向生产的信息提取流水线。

To evaluate our multi-method extraction framework, we construct a large-scale synthetic dataset of Chinese identity documents using GPT-4. The dataset simulates real-world diversity in format and content while preserving perfect ground truth.为了评估我们的多方法提取框架，我们使用GPT-4构建了一个大规模的中文身份证明文档合成数据集。该数据集模拟了现实世界中格式和内容的多样性，同时保留了完美的真实标签（Ground Truth）。

Data Generation Pipeline. We generate 10–30 identity entries per document using the Faker library (zh_CN locale), producing realistic Chinese names and 18-digit ID numbers. These entries are embedded into semantically plausible documents (e.g., insurance forms, travel records, registration sheets) via GPT-4 prompts. Each document contains contextual information such as dates, headers, and auxiliary fields.数据生成流水线。我们使用Faker库（zh_CN区域设置）为每份文档生成10–30个身份条目，产生逼真的中文姓名和18位身份证号。这些条目通过GPT-4提示嵌入到语义合理的文档（如保险单、旅行记录、登记表）中。每份文档都包含日期、标题和辅助字段等上下文信息。

Document Formats. To evaluate cross-modality robustness, each generated Markdown document is converted into four formats:文档格式。为了评估跨模态稳健性，每份生成的Markdown文档都被转换为四种格式：

• •

  PNG (100): Rendered HTML to image (via imgkit), simulating scanned documentsPNG (100)：通过imgkit将HTML渲染为图像，模拟扫描文档

  Report issue for preceding element
• •

  DOCX (100): Word documents with tables and formatted sectionsDOCX (100)：带有表格和格式化部分的Word文档

  Report issue for preceding element
• •

  PDF (100): Generated using WeasyPrint, preserving layout and structurePDF (100)：使用WeasyPrint生成，保留布局和结构

  Report issue for preceding element
• •

  XLSX (100): Spreadsheets with tabular identity dataXLSX (100)：带有表格化身份数据的电子表格

  Report issue for preceding element

The final dataset comprises 400 documents with over 10,000 name-ID pairs.最终数据集包含400份文档，包含超过10,000对姓名-ID组合。

We evaluate extraction performance at the (name, ID number) pair level.我们在（姓名，身份证号）对级别评估提取性能。

Accuracy Metrics.准确性指标。

• •

  Precision: Ratio of correctly extracted pairs to all extracted pairs $exactmatchrequired$精确率（Precision）：正确提取的对数与所有提取的对数之比。

  Report issue for preceding element
• •

  Recall: Ratio of correctly extracted pairs to total ground truth pairs召回率（Recall）：正确提取的对数与真实标签总对数之比。

  Report issue for preceding element
• •

  F1 Score: Harmonic mean of precision and recallF1分数：精确率和召回率的调和平均值。

  Report issue for preceding element

Efficiency Metrics.效率指标。

• •

  Text Extraction Time: Time for text extraction using each tool (MarkItDown, PaddleOCR, EasyOCR, etc.)文本提取时间：使用每种工具（MarkItDown、PaddleOCR、EasyOCR等）进行文本提取的时间。

  Report issue for preceding element
• •

  LLM Time: Inference time for the extraction model (excluding text extraction)LLM时间：提取模型的推理时间（不包括文本提取）。

  Report issue for preceding element
• •

  Total Time: End-to-end latency from input to structured output总时间：从输入到结构化输出的端到端延迟。

  Report issue for preceding element

Robustness Metrics.稳健性指标。

• •

  Success Rate: Percentage of documents processed without fatal errors成功率：无致命错误处理的文档百分比。

  Report issue for preceding element
• •

  Per-Format Accuracy: Extraction metrics broken down by document type (PNG, DOCX, etc.)分格式准确率：按文档类型（PNG、DOCX等）细分的提取指标。

  Report issue for preceding element

We evaluate our multi-method extraction framework on a corpus of 400 synthetic Chinese identity documents spanning four formats (PNG, DOCX, XLSX, PDF), utilizing three extraction paradigms across 16 OCR-LLM configurations. Our findings underscore substantial variation in performance across both extraction strategies and document formats, with method efficacy closely aligned with the structural characteristics of each format.我们在包含四种格式（PNG、DOCX、XLSX、PDF）的400份合成中文身份文档语料库上评估了我们的多方法提取框架，利用了16种OCR-LLM配置下的三种提取范式。我们的研究结果强调了不同提取策略和文档格式之间的性能差异巨大，且方法的有效性与每种格式的结构特征密切相关。

Experimental Overview. The evaluation encompasses 16 extraction methods categorized into three distinct paradigms: (1) Direct extraction, which leverages document parsers or multimodal models; (2) Replace-based extraction, which employs rule-based template matching; and (3) Table-based extraction, which exploits spatial structure for field localization. Each method was evaluated on up to 100 documents per supported format, resulting in 2,500 test instances. We report precision, recall, F1 score, and processing latency, further decomposed into OCR and LLM inference time.实验概览。评估涵盖了分为三个不同范式的16种提取方法：(1) 直接提取，利用文档解析器或多模态模型；(2) 基于替换的提取，采用基于规则的模板匹配；(3) 基于表格的提取，利用空间结构进行字段定位。每种方法都在每种支持格式的最多100份文档上进行了评估，产生了2,500个测试实例。我们报告了精确率、召回率、F1分数和处理延迟，并进一步分解为OCR和LLM推理时间。

Table 1 summarizes the performance of the best and worst-performing methods for each document type. For structured formats (DOCX/XLSX), table-based methods achieve perfect $F_{1}$ scores (1.0), with docling_table and markitdown_table delivering 100% success rates while maintaining minimal latency (0.3–0.5s). In stark contrast, replace-based methods perform poorly on these formats, with markitdown_replace achieving only $F_{1}$=0.969 for both DOCX and XLSX, representing the worst performance with perfect rates dropping to 59% and 54% respectively. For image-based formats (PNG), the multimodal approach outperforms all OCR-based methods with an $F_{1}$ score of 0.999, while easyocr_table catastrophically fails with $F_{1}$=0.000 and 0% success rate despite minimal processing time (1.5s). PDF extraction exhibits the most extreme performance variance: docling_table maintains perfect accuracy ($F_{1}$=1.0) with low latency (1.6s), whereas mineru_replace completely fails ($F_{1}$=0.000, 0% success rate), highlighting severe method-format incompatibilities. Direct extraction methods consistently incur high LLM inference latency (13.4–13.6s), accounting for over 90% of total processing time, while table-based approaches achieve 40–50$\times$ speedup through minimal LLM usage.表1总结了每种文档类型表现最好和最差的方法的性能。对于结构化格式（DOCX/XLSX），基于表格的方法实现了完美的F1分数（1.0），其中docling_table和markitdown_table实现了100%的成功率，同时保持了极低的延迟（0.3–0.5秒）。相比之下，基于替换的方法在这些格式上表现不佳，markitdown_replace在DOCX和XLSX上仅达到F1=0.969，表现最差，完美率分别降至59%和54%。对于基于图像的格式（PNG），多模态方法以F1=0.999的得分优于所有基于OCR的方法，而easyocr_table尽管处理时间极短（1.5秒），却以F1=0.000和0%的成功率彻底失败。PDF提取表现出最极端的性能差异：docling_table保持了完美的准确性（F1=1.0）且延迟较低（1.6秒），而mineru_replace则完全失败（F1=0.000，成功率为0%），突显了严重的方法-格式不兼容性。直接提取方法始终会导致较高的LLM推理延迟（13.4–13.6秒），占总处理时间的90%以上，而基于表格的方法通过最少的LLM使用实现了40–50倍的加速。

Figure 4 provides a holistic view of performance across all 16 methods and formats via dual heatmaps. The $F_{1}$ score heatmap reveals distinct performance patterns: table-based methods exhibit binary characteristics—either achieving near-perfect extraction ($F_{1}\approx 1.0$) or complete failure ($F_{1}=0.0$), indicating strong format dependency. Direct extraction methods demonstrate more consistent but suboptimal performance across formats ($F_{1}\approx 0.77$–$0.99$), while replace-based methods show the highest variability, ranging from moderate success to total failure. The processing time heatmap complements these findings, showing that methods with perfect $F_{1}$ scores often achieve the fastest processing times (0.3–1.6s for table-based approaches), while direct methods consistently require 13–15s due to LLM overhead. The multimodal method presents an outlier with 33.9s processing time for PNG files, trading computational efficiency for extraction accuracy. Empty cells in both heatmaps denote unsupported format-method combinations, particularly evident for specialized parsers like mineru (PDF-only) and OCR-based methods (image formats only). The visualization conclusively demonstrates that no single method achieves universal optimality across all formats, necessitating format-specific extraction strategies for optimal performance.图4通过双热力图提供了所有16种方法和格式的整体性能概览。F1分数热力图揭示了明显的性能模式：基于表格的方法表现出二元特征——要么实现近乎完美的提取（F1≈1.0），要么完全失败（F1=0.0），表明了强烈的格式依赖性。直接提取方法在各种格式上表现出更一致但非最优的性能（F1≈0.77–0.99），而基于替换的方法表现出最高的变异性，从适度成功到完全失败不等。处理时间热力图补充了这些发现，显示出具有完美F1分数的方法往往实现最快的处理时间（基于表格的方法为0.3–1.6秒），而由于LLM开销，直接方法始终需要13–15秒。多模态方法是一个异常值，PNG文件的处理时间为33.9秒，以计算效率换取了提取准确性。热力图中的空白单元格表示不支持的格式-方法组合，这在像mineru（仅限PDF）和基于OCR的方法（仅限图像格式）这样的专用解析器中尤为明显。该可视化结论性地证明，没有任何单一方法能在所有格式上实现通用最优，因此需要针对特定格式的提取策略以获得最佳性能。

Image-based Documents (PNG). For scanned documents with identity cards, we observe stark contrasts between multimodal and OCR-based pipelines, as illustrated in Figure 5. The multimodal vision-language model demonstrates exceptional accuracy ($F_{1}=0.999\pm 0.007$) through direct visual feature processing, effectively bypassing the cascading errors inherent in sequential OCR pipelines. However, this end-to-end approach incurs prohibitive computational costs, requiring $33.91\pm 9.49$ seconds per document due to high-dimensional visual encoding and transformer-based attention mechanisms. Among OCR-based methods, we observe significant performance variations based on both the OCR engine and extraction strategy. PaddleOCR consistently outperforms EasyOCR across all extraction paradigms, achieving $F_{1}$ scores of 0.997, 0.961, and 0.997 for direct, replacement, and table-based extraction respectively, while EasyOCR exhibits substantially degraded performance ($F_{1}=0.760$, $0.688$, and $0.000$). This performance disparity stems fundamentally from EasyOCR’s inability to preserve document spatial structure during text extraction, which disrupts the positional relationships critical for accurate field identification in structured documents like identity cards.基于图像的文档（PNG）。对于带有身份证的扫描文档，我们观察到多模态流水线与基于OCR的流水线之间的强烈对比，如图5所示。多模态视觉语言模型通过直接处理视觉特征，展示了卓越的准确性（F1=0.999±0.007），有效地绕过了顺序OCR流水线中固有的级联错误。然而，这种端到端方法产生了昂贵的计算成本，由于高维视觉编码和基于Transformer的注意力机制，每份文档需要33.91±9.49秒。在基于OCR的方法中，我们观察到基于OCR引擎和提取策略的显著性能差异。PaddleOCR在所有提取范式中始终优于EasyOCR，直接、替换和基于表格的提取的F1分数分别为0.997、0.961和0.997，而EasyOCR表现出显著下降（F1分别为0.760、0.688和0.000）。这种性能差异从根本上源于EasyOCR在文本提取过程中无法保留文档空间结构，这破坏了在身份证等结构化文档中进行准确字段识别至关重要的位置关系。

The spatial information loss particularly impacts table-based extraction, where EasyOCR completely fails ($F_{1}=0.000$) as the LLM cannot generate valid coordinates without reliable spatial encoding. In contrast, PaddleOCR maintains precise spatial mappings, enabling the table-based extraction strategy to achieve optimal performance at $0.63\pm 0.24$ seconds—a $54\times$ speedup over the multimodal baseline—while preserving near-perfect accuracy ($F_{1}=0.997\pm 0.011$). This method minimizes LLM inference to coordinate generation only, leveraging the preserved document structure to reduce computational overhead from $\sim$34 seconds to sub-second latency. The replacement method with PaddleOCR, employing regex patterns for standardized fields and batch processing for names, offers a middle ground at $0.72\pm 0.28$ seconds ($F_{1}=0.961\pm 0.044$), while direct extraction requires $13.75\pm 3.87$ seconds despite high accuracy ($F_{1}=0.997\pm 0.010$) due to processing entire OCR outputs. The PaddleOCR table-based approach thus emerges as the optimal solution for production deployment, combining superior spatial preservation with minimal LLM computation. The marginal 0.2% accuracy trade-off compared to multimodal methods represents an acceptable compromise given the 54-fold efficiency gain crucial for large-scale document processing systems.空间信息丢失对基于表格的提取影响尤为严重，EasyOCR完全失败（F1=0.000），因为没有可靠的空间编码，LLM无法生成有效的坐标。相比之下，PaddleOCR保持了精确的空间映射，使基于表格的提取策略能够在0.63±0.24秒内实现最优性能——比多模态基准快54倍——同时保持近乎完美的准确性（F1=0.997±0.011）。该方法将LLM推理最小化为仅生成坐标，利用保留的文档结构将计算开销从约34秒降低到亚秒级延迟。结合PaddleOCR的替换方法，利用正则表达式处理标准化字段，并对姓名进行批处理，在0.72±0.28秒内提供了中间方案（F1=0.961±0.044），而直接提取尽管准确率很高（F1=0.997±0.010），但由于处理整个OCR输出，仍需要13.75±3.87秒。因此，基于PaddleOCR的表格方法成为生产部署的最佳解决方案，结合了卓越的空间保留与最小的LLM计算。与多模态方法相比，0.2%的准确率权衡在54倍的效率增益面前是可以接受的折衷，这对大规模文档处理系统至关重要。

Structured Office Documents (DOCX/XLSX). Our evaluation of structured office document extraction reveals fundamentally different performance characteristics compared to image-based formats, as illustrated in Figure 6. Native documents exhibit near-perfect accuracy across all extraction methodologies due to their inherent machine-readable structure and preserved spatial information. For DOCX files, both MarkItDown and Docling frameworks successfully extract the document’s spatial structure, enabling high-fidelity extraction across all paradigms. Docling’s table-based approach achieves optimal performance with perfect accuracy ($F_{1}=1.000$) and exceptional efficiency at $0.34\pm 0.03$ seconds—a $41\times$ speedup compared to direct extraction ($13.68\pm 3.85$ seconds)—by leveraging structured table representations that minimize LLM processing to coordinate generation only. MarkItDown’s table method follows closely with identical accuracy ($F_{1}=1.000$) at $0.54\pm 0.13$ seconds, representing a $26\times$ improvement over its direct counterpart.结构化办公文档（DOCX/XLSX）。我们对结构化办公文档提取的评估揭示了与基于图像的格式完全不同的性能特征，如图6所示。原生文档由于其固有的机器可读结构和保留的空间信息，在所有提取方法中表现出近乎完美的准确性。对于DOCX文件，MarkItDown和Docling框架都成功提取了文档的空间结构，实现了所有范式下的高保真提取。Docling的基于表格的方法实现了最优性能，准确率完美（F1=1.000），效率极高，为0.34±0.03秒——比直接提取（13.68±3.85秒）快41倍——通过利用结构化表格表示，将LLM处理最小化为仅生成坐标。MarkItDown的表格方法紧随其后，准确率相同（F1=1.000），时间为0.54±0.13秒，比其直接对应方法提高了26倍。

While replacement strategies offer competitive latency (Docling: $0.59\pm 0.13$s, MarkItDown: $0.74\pm 0.18$s), they sacrifice accuracy ($F_{1}=0.972\pm 0.038$ and $0.969\pm 0.044$ respectively) due to the LLM’s occasional misidentification when matching ID numbers to corresponding names—a limitation stemming from the model’s reliance on contextual inference rather than explicit structural cues. This name-matching error could be mitigated through prompt engineering to provide more explicit matching instructions or fine-tuning the general-purpose LLM on document-specific extraction tasks. For XLSX spreadsheets, where only MarkItDown provides support, the inherent tabular structure amplifies these efficiency gains: table-based extraction achieves perfect accuracy ($F_{1}=1.000$) with remarkable $0.30\pm 0.02$ seconds processing time—a $44\times$ acceleration over direct methods ($13.53\pm 3.82$ seconds). This dramatic improvement stems from the natural alignment between spreadsheet cell structures and table-based extraction paradigms, eliminating the need for complex content interpretation. The replacement method maintains reasonable efficiency ($0.51\pm 0.11$s) but exhibits similar accuracy degradation ($F_{1}=0.969\pm 0.039$) as observed in DOCX processing, with errors predominantly occurring in name-ID association tasks. These results establish table-based extraction as the unequivocally superior approach for native office documents, where the preserved document structure enables both perfect accuracy and minimal computational overhead, making it ideal for enterprise-scale document processing pipelines requiring both precision and throughput.虽然替换策略提供了具有竞争力的延迟（Docling：0.59±0.13秒，MarkItDown：0.74±0.18秒），但它们牺牲了准确性（F1分别为0.972±0.038和0.969±0.044），原因是LLM在将身份证号与相应姓名匹配时偶尔会出现识别错误——这一局限性源于模型对上下文推理的依赖，而非显式的结构线索。这种姓名匹配错误可以通过提示工程来提供更明确的匹配指令，或在特定文档提取任务上微调通用LLM来缓解。对于仅有MarkItDown支持的XLSX电子表格，固有的表格结构放大了这些效率增益：基于表格的提取实现了完美的准确性（F1=1.000），处理时间仅为0.30±0.02秒——比直接方法（13.53±3.82秒）加速44倍。这种显著的改进源于电子表格单元格结构与基于表格的提取范式之间的自然对齐，消除了对复杂内容解释的需求。替换方法保持了合理的效率（0.51±0.11秒），但在DOCX处理中观察到了类似的准确性下降（F1=0.969±0.039），错误主要发生在姓名-ID关联任务中。这些结果确立了基于表格的提取是原生办公文档的绝对最优方法，其中保留的文档结构实现了完美的准确性和最小的计算开销，使其成为需要精度和吞吐量的企业级文档处理流水线的理想选择。

Portable Document Format (PDF). Our empirical analysis of PDF document processing reveals a fundamental trade-off between computational efficiency and extraction accuracy across different methodological approaches. MarkItDown prioritizes processing speed, achieving minimal OCR preprocessing overhead (0.056s), while Docling and MinerU adopt more computationally intensive strategies with OCR latencies of approximately 1.3–1.5s. However, this efficiency-accuracy trade-off manifests differently across extraction paradigms. MarkItDown’s lightweight approach, which employs symbolic markers for spatial structure representation, exhibits severely degraded performance across all extraction strategies: direct extraction ($F_{1}=0.772$), replacement extraction ($F_{1}=0.570$), and table-based extraction ($F_{1}=0.070$).便携式文档格式（PDF）。我们对PDF文档处理的实证分析揭示了不同方法论在计算效率和提取准确性之间的根本权衡。MarkItDown优先考虑处理速度，实现了最小的OCR预处理开销（0.056秒），而Docling和MinerU采用了计算密集型策略，OCR延迟约为1.3–1.5秒。然而，这种效率-准确性权衡在不同的提取范式中表现不同。MarkItDown的轻量级方法使用符号标记来表示空间结构，在所有提取策略中表现出严重下降：直接提取（F1=0.772）、替换提取（F1=0.570）和表格提取（F1=0.070）。

The divergence in extraction efficacy between Docling and MinerU, despite their comparable preprocessing costs, underscores the critical importance of spatial representation strategies in document understanding. Docling maintains near-perfect accuracy across all extraction paradigms, with its table-based approach achieving optimal performance ($F_{1}=1.0$). In stark contrast, MinerU’s reliance on HTML-style structural tags (e.g., <td>, <tr>) for encoding spatial relationships proves incompatible with text-based information extraction, resulting in catastrophic failure for replacement and table-based methods ($F_{1}=0.0$). The framework only maintains competitive performance in direct extraction mode ($F_{1}=0.9996$), where spatial structure parsing is bypassed entirely. These empirical findings, illustrated in Figure 7, establish Docling’s table-based methodology as the optimal solution for PDF information extraction, successfully reconciling the competing demands of computational efficiency (1.61s total processing time) and extraction fidelity.尽管预处理成本相当，但Docling和MinerU在提取效能上的差异突显了空间表示策略在文档理解中的关键重要性。Docling在所有提取范式中保持了近乎完美的准确性，其表格方法实现了最优性能（F1=1.0）。相比之下，MinerU依赖HTML风格的结构标签（如<td>、<tr>）来编码空间关系，被证明与基于文本的信息提取不兼容，导致替换和表格方法的灾难性失败（F1=0.0）。该框架仅在直接提取模式下保持了竞争性性能（F1=0.9996），该模式完全绕过了空间结构解析。这些如图7所示的实证结果确立了Docling的基于表格的方法论是PDF信息提取的最优解决方案，成功调和了计算效率（1.61秒总处理时间）和提取保真度之间的竞争需求。

Our comprehensive evaluation reveals a fundamental insight: the optimal extraction strategy for copy-heavy documents is intrinsically tied to document modality and structure, challenging the prevailing one-size-fits-all approaches in production systems. The 54× performance differential between table-based methods (0.97s) and multimodal approaches (33.91s) underscores a critical trade-off—while end-to-end models offer superior robustness through direct visual understanding, their computational overhead remains prohibitive for high-throughput scenarios characteristic of copy-heavy tasks. This finding suggests that the field’s pursuit of universal extraction models may be misguided for repetitive document processing, where structural priors can be effectively exploited.我们的全面评估揭示了一个根本性洞察：针对大量重复性文档（copy-heavy documents）的最佳提取策略，本质上取决于文档的模态和结构，这挑战了生产系统中普遍存在的“一刀切”方法。基于表格的方法（0.97秒）与多模态方法（33.91秒）之间54倍的性能差异，凸显了一个关键的权衡——尽管端到端模型通过直接的视觉理解提供了卓越的鲁棒性，但其计算开销对于处理大量重复性文档任务所需的高吞吐量场景而言仍然过高。这一发现表明，在该领域追求通用提取模型，对于可以有效利用结构先验的重复性文档处理任务来说，可能方向有误。

The stark performance dichotomy between OCR engines (PaddleOCR F1=0.997 vs. EasyOCR F1=0.000 for table extraction) reveals that spatial structure preservation, rather than character recognition accuracy alone, determines extraction success. This challenges conventional OCR evaluation metrics and highlights the need for structure-aware benchmarks. Moreover, the consistent superiority of table-based methods across native formats (100% accuracy at <1s) demonstrates that explicit structural modeling outperforms both semantic understanding (direct extraction) and pattern matching (replacement extraction) when document layouts are predictable—a defining characteristic of copy-heavy scenarios.OCR引擎之间巨大的性能鸿沟（在表格提取任务中，PaddleOCR F1=0.997，而EasyOCR F1=0.000）表明，决定提取成功与否的关键在于对空间结构的保留，而非仅仅取决于字符识别的准确率。这挑战了传统的OCR评估指标，并突显了对结构感知基准测试的需求。此外，基于表格的方法在原生格式上始终保持卓越表现（<1秒内实现100%准确率），证明了在文档布局可预测的情况下（这是大量重复性文档场景的典型特征），显式的结构建模优于语义理解（直接提取）和模式匹配（替换提取）。

From a production standpoint, our results advocate for adaptive, format-aware architectures over monolithic solutions. The multimodal approach’s resilience to OCR failures positions it as an ideal fallback mechanism, while table-based extraction serves as the workhorse for structured documents. This hierarchical strategy—rapid format detection followed by method-specific routing—enables systems to process heterogeneous document streams at scale without sacrificing accuracy. The framework thus provides a blueprint for reconciling the competing demands of accuracy, efficiency, and robustness in real-world document processing pipelines.从生产角度来看，我们的研究结果主张采用自适应、感知格式的架构，而非单一的整体式解决方案。多模态方法对OCR故障的抵御能力使其成为理想的备选机制，而基于表格的提取则是处理结构化文档的主力。这种分层策略——先进行快速格式检测，再进行针对性的方法路由——使系统能够在大规模处理异构文档流的同时，无需牺牲准确性。因此，该框架为在实际文档处理流水线中协调准确性、效率和鲁棒性之间的竞争需求提供了一份蓝图。