PaperlayerView exact revision on arXiv

Source-linked AI summary

PACE: Publisher-Adaptive Content Extraction via Agentic Automation

Zhanlin Liu, Munirathnam Srikanth

arXiv:2608.27466v1cs.CLcs.AI

TL;DR

Web extraction must balance accuracy, scalability, and adaptability across publisher-specific layouts and increasingly rich LLM pipeline outputs. PACE learns publisher-specific configurations with LLM-guided agentic analysis, then uses a fixed deterministic template for inference. Across article-body, metadata, and multimodal benchmarks, it outperforms scalable non-manual baselines while approaching manually engineered parsers, though some richer-task references are not human-annotated.

  • Problem

    Existing web extractors struggle to jointly provide accuracy, scalability, and adaptability across publisher-specific layouts and richer outputs such as metadata, images, and tables.

  • Method

    PACE uses agentic LLM analysis of representative pages and user requirements to learn reusable publisher-specific configurations, then instantiates a fixed deterministic extractor template.

  • Results

    Across article-body, metadata, and multimodal benchmarks, PACE outperforms scalable non-manual baselines while approaching manually engineered publisher-specific parsers.

  • Takeaways & Limitations

    Agentic configuration learning supports scalable, flexible LLM-ready page representations beyond article text, including metadata, images, and tables.

  • Takeaways & Limitations

    Metadata and table scores rely on GPT-5.4-generated references, while image evaluation uses a deterministic HTML-derived heuristic rather than human-labeled annotations.

Abstract

from arXiv · show

Web content extraction is essential for reliable LLM data pipelines, yet existing methods often struggle to jointly satisfy accuracy, scalability, and adaptability. General-purpose extractors can be applied broadly, but they are often brittle on publisher-specific layouts and richer extraction targets such as metadata, images, and tables. Direct LLM-based extraction offers greater flexibility, but incurs substantial cost and latency at scale, while manually engineered publisher-specific parsers can achieve high accuracy but require substantial human effort to build and maintain. We introduce PACE, an agentic framework for learning publisher-specific extraction configurations from representative pages and user requirements. During training, PACE uses LLMs to analyze page structure and aggregate reusable extraction patterns. At inference time, the learned configurations instantiate a fixed deterministic extractor template, enabling scalable extraction without additional LLM calls. Experiments spanning article-body, metadata, and multimodal extraction show that PACE outperforms scalable non-manual baselines while approaching the quality of manually engineered publisher-specific parsers. PACE achieves stronger extraction of article text, metadata, images, and tables, demonstrating that agentic configuration learning can automate publisher-specific extraction for LLM-ready page representations beyond article text.

1 Introduction

Web extraction must handle publisher-specific layouts and diverse outputs for LLM applications, but existing approaches trade off accuracy, scalability, adaptability, and engineering effort. PACE addresses this by learning reusable publisher-specific configurations with LLMs and executing them through a deterministic extractor template.

  • Motivation: LLM pipelines increasingly require requirement-adaptive representations containing article text, metadata, tables, figures, images, captions, links, or structured outputs.Extraction targets vary across retrieval, multimodal grounding, information extraction, and tool-use systems.
  • Limitations of Existing Methods: General-purpose extractors rely on fixed structural assumptions and may struggle with publisher-specific templates, interactive pages, and targets outside predefined schemas.These methods commonly use text density, link density, readability signals, or fixed training distributions.
  • Limitations of Existing Methods: Direct LLM extraction is flexible, whereas manually engineered publisher-specific parsers can be reliable but require substantial engineering effort to build and maintain.The existing tradeoff spans accuracy, scalability, adaptability, and maintenance cost.
  • PACE: PACE learns publisher-specific extraction behavior from representative pages and user-defined requirements through agentic orchestration, page analysis, pattern aggregation, and optional feedback.The learned output is a reusable publisher-level configuration rather than unconstrained extractor code.
  • PACE: PACE separates LLM-guided training from deterministic template-based inference, enabling scalable extraction without additional per-page LLM calls.The template applies publisher-specific selectors, exclusion rules, fallback logic, and formatting instructions.
  • Results: PACE provides empirical evidence across article text, metadata, and multimodal extraction, performing strongest among evaluated scalable non-manual extractors while approaching manually engineered parsers.Its supported outputs extend beyond article text to images, tables, and other LLM-ready page representations.

2 Related Work

Prior extraction systems balance broad applicability, efficiency, and learned structure against limitations in publisher adaptation and richer output targets. PACE connects these research directions by using agentic LLM analysis to learn publisher-specific configurations while retaining deterministic runtime extraction.

  • Heuristic and Machine-Learning Extraction: Heuristic systems such as Boilerpipe, jusText, trafilatura, and news-please provide broadly applicable and computationally efficient boilerplate removal.Their approaches combine structural, textual, density, link, and readability signals.
  • Heuristic and Machine-Learning Extraction: Fixed heuristics may struggle with publisher-specific templates, interactive pages, metadata, in-content images, captions, and tables.Their extraction behavior is largely fixed and typically assumes conventional article text.
  • Heuristic and Machine-Learning Extraction: Machine-learning methods learn richer structural and textual patterns than fixed heuristics but remain constrained by their training distributions and task definitions.BoilerNet formulates boilerplate removal as neural sequence labeling over HTML text blocks.
  • Heuristic and Machine-Learning Extraction: Learned extractors may transfer unreliably to new publishers, evolving layouts, or targets beyond article-body text.PACE instead learns publisher-specific configurations at deployment time from representative pages.
  • LLM-Based Extraction: LLM-based extraction supports natural-language schemas and varied output formats but incurs recurring inference costs, latency, and potential nondeterminism when applied to every page.PACE uses LLMs during training and a deterministic template during inference to avoid per-page calls.
  • Publisher-Specific Extraction: Publisher-specific parsers can achieve high-quality extraction, but manually inspecting DOM structures, writing rules, and maintaining parsers limits scalability.PACE automates customization through agentic configuration learning and template-based deterministic inference.

3 Methodology

PACE learns reusable publisher-level extraction configurations from representative pages and user requirements, then instantiates a fixed deterministic extractor for new pages. Its pipeline combines LLM-based analysis, pattern aggregation, configurable output handling, optional feedback, and scalable inference without additional LLM calls.

  • 3.1 Problem Formulation: PACE takes representative publisher pages and user requirements defining targets, exclusions, formats, and optional structured or multimodal fields.Each page contains target content interleaved with navigation, advertisements, recommendations, comments, and layout-specific boilerplate.
  • 3.1 Problem Formulation: The objective is to learn a reusable publisher-level configuration that generalizes from representative pages to unseen pages from the same publisher.The configuration is used by a publisher-specific extractor to produce the requested representation.
  • 3.2 Agentic Configuration Learning: An agentic orchestrator coordinates a page-level analyzer and publisher-level aggregator to infer page structure and consolidate recurring patterns.The analyzer operates on individual pages, while the aggregator produces the reusable publisher-level configuration.
  • 3.2 Agentic Configuration Learning: PACE preprocesses reference pages by removing scripts, styles, hidden elements, and common structural noise while preserving visible content structure and text-bearing elements.The simplified DOM focuses subsequent analysis on extraction-relevant structure.
  • 3.2 Agentic Configuration Learning: The analyzer proposes generalized selectors and exclusions, while the aggregator removes one-off rules, generalizes unstable class names, and resolves conflicts.Hash-like or CSS-module suffixes can be converted into robust partial-match selectors.
  • 3.3 Template-Based Extractor Generation and Inference: A reusable fixed template executes DOM traversal, selection, exclusion, fallback extraction, deduplication, normalization, and output formatting using the learned configuration.The same framework supports article text, metadata, images, tables, captions, markdown, JSON, and other requirement-conditioned outputs.
  • 3.3 Template-Based Extractor Generation and Inference: Inference applies learned selectors and rules to new or held-out pages without additional LLM calls, avoiding the cost, latency, and nondeterminism of direct LLM extraction.Optional human review and evaluation-driven refinement can update prompts, requirements, and configurations during training.

4 Experiments

PACE is evaluated on article-body, metadata, image, and table extraction across held-out publisher pages, using reusable configurations learned from representative pages and deterministic inference. It outperforms scalable non-manual baselines across the reported extraction targets while remaining close to the manually engineered Fundus reference for article-body extraction.

  • Evaluation settings: PACE uses GPT-5.4 during training to learn publisher-level configurations, then applies instantiated extractors to held-out pages without additional LLM calls.The fixed runtime supports reusable, scalable extraction after configuration learning.
  • Evaluation settings: PACE is evaluated on article-body extraction using 71 held-out pages from 16 publishers and on metadata and multimodal extraction across 21 publishers or domains.The broader benchmark covers metadata fields, in-content images, and tables, including five table-rich domains.
  • Metadata and multimodal extraction: PACE achieves higher table F1 than trafilatura on all five table-rich domains, with scores ranging from 0.885 to 1.000.The results cover pages with rowspan and colspan structures, multi-level headers, and row-header stubs.
  • Article-body extraction: PACE achieves the strongest ROUGE-L F1 and WER among scalable non-manual article-body methods, remains close to Fundus, and ranks in the top two on 11 of 16 publishers.Fundus is the manually engineered publisher-specific reference baseline.
  • Metadata and multimodal extraction: Image extraction favors PACE overall, but results are less uniform: it wins on most publishers, ties The Telegraph, and trails on LATimes and The Nation.The reported overall image F1 is 0.516 for PACE versus 0.318 for trafilatura.

5 Conclusion and Future Work

PACE learns publisher-specific extraction configurations and uses them with a fixed extractor template for scalable inference without per-page LLM calls. Across article-body, metadata, multimodal, and table extraction, it outperforms scalable non-manual baselines while approaching manually engineered parsers; future work extends its scope.

  • Conclusion: PACE learns publisher-specific extraction configurations from representative pages and user requirements, then instantiates a fixed extractor template for scalable inference.The learned configurations enable inference without per-page LLM calls.
  • Conclusion: PACE achieves the strongest performance among scalable non-manual methods while approaching manually engineered publisher-specific parsers on article-body extraction.The article-body benchmark covers 16 news publishers and 71 held-out pages.
  • Conclusion: PACE outperforms trafilatura on metadata, images, and tables in a 21-publisher benchmark.These results cover metadata and multimodal extraction beyond article text.
  • Future Work: Future work will extend PACE beyond independent publisher-specific extractors and beyond news to product pages, documentation, forums, chats, and enterprise knowledge bases.The paper also proposes exploring shared template families or more general extractors that reuse common layout patterns.
Loading 2608.27466v1…