pdf parsing

This study addresses the limitation of existing keyword extraction methods, which predominantly focus on plain text while neglecting visual and audio modalities, and the absence of dedicated multimodal datasets. To bridge this gap, the authors introduce the first multimodal dataset comprising 1,000 academic papers, each annotated with human-labeled keywords and accompanied by original text, images (with OCR-extracted text), and audio transcripts (via automatic speech recognition). Leveraging this resource, the paper systematically evaluates the impact of individual modalities and their fusion—under both unsupervised and supervised settings—on keyword extraction performance. Experimental results demonstrate that integrating multimodal textual information significantly enhances extraction accuracy, with distinct modalities offering complementary cues, thereby confirming the effectiveness and necessity of multimodal modeling for academic keyword extraction.

Existing PDF parsing tools are primarily designed for academic papers and struggle to accurately process pedagogical documents—such as legal textbooks—that exhibit complex, implicitly structured hierarchies. To address this, we propose a hierarchical text segmentation framework integrating structure-aware preprocessing with large language models (LLMs). Our method jointly leverages OCR-based heading detection, XML structural feature extraction, and contextual semantic modeling to infer implicit heading hierarchies without requiring explicit table-of-contents input. Compared to pure LLM–based or traditional rule-based approaches, our framework significantly reduces false positives and improves segmentation accuracy. When high-quality metadata is available, a supplementary table-of-contents–driven strategy further enhances performance. The source code and benchmark dataset are publicly released to support reproducible research.

This study addresses the challenge of extracting key metadata (e.g., title, authors, affiliations) from German social science PDFs, which exhibit high template heterogeneity and widespread metadata absence. We systematically benchmark state-of-the-art NLP (BERT), computer vision (YOLO/CNN-based layout analysis), and multimodal approaches (LayoutLMv3, text-layout-image joint modeling) on metadata extraction, establishing the first unified evaluation framework tailored to real-world, high-diversity German academic PDFs—balancing accuracy and efficiency. Experimental results show that multimodal methods achieve up to a 12.7% F1-score improvement over unimodal baselines, with layout-aware models significantly outperforming pure text-based ones. To foster reproducibility and FAIR metadata infrastructure, we publicly release the first benchmark dataset dedicated to German social science PDFs, along with fully documented, end-to-end training and inference code.

This work addresses high-accuracy conversion of unstructured/semi-structured documents (e.g., contracts, academic papers, invoices) into structured, machine-readable data. Method: We systematically survey and empirically compare modular pipeline approaches against end-to-end multimodal large models, proposing a unified framework integrating OCR, layout analysis (LayoutParser), graph neural networks, vision-language models (VLMs), and specialized formula/table recognition. We identify and characterize core bottlenecks—layout understanding, dense text recognition, and cross-modal alignment—for the first time. Contribution/Results: We establish a comprehensive analytical framework covering methodology, challenges, and benchmarks, revealing >32% performance gaps of current SOTA on complex layouts (e.g., multi-column, nested tables). We propose a “dual-driven” evolution path emphasizing both data diversity and scale, and open-source a larger annotated dataset to significantly advance knowledge base construction and training-data generation for large models.

Scientific PDF information extraction tools suffer from inconsistent input formats, opaque “black-box” behavior, poor fault tolerance, and limited format support—hindering literature analysis efficiency for non-NLP researchers. To address these challenges, we propose the first modular information extraction experimental framework specifically designed for scientific PDFs, enabling model-level decoupling, fine-grained intermediate-state visualization, and unified cross-model evaluation. The framework integrates Hugging Face token classifiers, diverse large language models (LLMs), and domain-specific models within a PDF processing pipeline comprising layout-aware parsing, text reconstruction, and semantic alignment. Evaluated on materials science literature review tasks, it significantly reduces model trial-and-error overhead, improves error attribution accuracy, and enhances system interpretability. The platform delivers a debuggable, reusable, plug-and-play rapid prototyping capability for scientific information extraction, empowering domain experts without NLP expertise.

Existing benchmarks struggle to systematically evaluate the impact of multimodal inputs on structured information extraction. This work proposes the first multimodal extraction benchmark tailored for government forms, generating diverse documents with deterministic ground truth through procedural PDF templates and a reverse annotation pipeline. Each document is provided in four standardized input formats—plain text, layout-preserving text, image, and multimodal—to enable rigorous ablation studies. A three-stage quality control mechanism and compliance-based evaluation of structured outputs ensure high data fidelity and reliable assessment. Experiments reveal that small models (<4B parameters) are primarily limited by structural adherence; fine-tuning a 2B model yields an 81-percentage-point improvement; layout-preserving text consistently outperforms image inputs by 3–18 percentage points; and the benchmark exhibits maximal discriminative power within the 60–95% accuracy range.

Existing PDF parsers often fail to capture critical visual content, erroneously extract irrelevant images, and struggle to accurately associate figures with their captions—limitations that significantly hinder the performance of multimodal retrieval-augmented generation (RAG) systems. To address these challenges, this work proposes a lightweight, production-oriented visual PDF parsing framework that, for the first time in a production setting, integrates spatial heuristics, document layout analysis, and semantic similarity to achieve robust, low-latency detection of visual elements and precise figure–caption alignment. Evaluated on both public and internal datasets, the method attains detection accuracy of at least 96% and caption association accuracy of 93%. When deployed as a RAG preprocessing module, it substantially outperforms current state-of-the-art approaches while reducing inference latency by more than twofold.

This work addresses the challenge that current machine translation systems struggle to preserve alignment between textual content and visual layout—such as tables and mathematical formulas—when processing PDF documents, often resulting in structural distortion. To tackle this issue, the authors construct a multimodal parallel corpus comprising 3,956 PDFs across 15 language pairs, meticulously retaining original layout metadata. They introduce, for the first time, a 45-dimensional geometric feature space to guide K-Medoids sampling, prioritizing visual structural diversity. This dataset enables layout-aware translation that jointly leverages textual and visual context, exposing critical limitations of existing systems in spatial localization and geometric synchronization. It thus establishes a high-fidelity, evaluable benchmark for document translation and reconstruction that faithfully preserves complex layouts.

To address three core challenges in scientific literature information extraction—modeling long documents, understanding multimodal content, and standardizing fine-grained cross-paper information (especially under dynamically evolving data schemas)—this paper proposes SciEx, a modular, decoupled framework. SciEx explicitly separates PDF parsing, multimodal retrieval, LLM-driven extraction, and cross-document aggregation, enabling plug-and-play integration of diverse prompting strategies, foundation models, and inference mechanisms for rapid adaptation. Evaluated across three domain-specific datasets, SciEx achieves high accuracy and consistency in fine-grained information extraction. The study systematically identifies key strengths and bottlenecks of current LLM-based pipelines, offering an extensible and maintainable technical pathway for constructing scientific knowledge graphs that evolve with shifting data patterns and scholarly conventions.

This study investigates whether document information extraction in the era of multimodal large language models (MLLMs) still necessitates traditional OCR preprocessing. Through large-scale benchmarking, the authors evaluate the end-to-end performance of off-the-shelf MLLMs on real-world business documents and propose a direct image input pipeline that bypasses OCR entirely. They develop an LLM-based automated hierarchical error analysis framework to systematically diagnose failure modes and introduce structured prompting, in-context learning, and output schema constraints to significantly enhance extraction accuracy. Experimental results demonstrate that the OCR-free approach achieves comparable accuracy to OCR-augmented pipelines across most scenarios, confirming the feasibility of streamlining the extraction workflow and offering practical guidance for real-world deployment.