跳转至

文章背景与核心概要

网页内容提取是构建可靠的大语言模型（LLM）数据管道的核心环节。然而，传统方法长期面临准确性、可扩展性和自适应性之间的“不可能三角”困境：通用提取器泛化能力强但面对特定发布者布局时极为脆弱；直接基于LLM的提取方式灵活却在大规模应用时成本高昂、延迟明显；人工编写的定制解析器虽然精度高，但维护成本极其沉重。

为了解决这一痛点，本文提出了 PACE（Publisher-Adaptive Content Extraction via Agentic Automation）这一全新框架。它利用智能体技术，根据具有代表性的网页和用户需求自动学习特定发布者的提取配置。在训练阶段，PACE通过LLM分析页面结构并综合出可复用的提取模式；在推理阶段，学习到的配置将被实例化到固定且确定性的提取器模板中，从而在无需额外LLM调用的情况下实现高性能、可扩展的内容提取。

跨文章正文、元数据、图像以及表格的多项实验表明，PACE 的性能显著优于现有的可扩展非人工基线方法，并无限逼近人工定制解析器的质量。该研究证明了智能体配置学习能够有效实现LLM就绪页面表示的自动化发布者特定提取，极大扩展了自动化内容解析的边界。

PACE: Publisher-Adaptive Content Extraction via Agentic Automation

PACE: Publisher-Adaptive Content Extraction via Agentic Automation

Summary

Summary

Web content extraction is a vital component of reliable Large Language Model (LLM) data pipelines. However, traditional approaches face a trilemma of balancing accuracy, scalability, and adaptability. General-purpose extractors are broad but brittle when dealing with specific publisher layouts or complex extraction targets (like metadata, images, and tables). Direct LLM-based extraction is flexible yet costly and slow at scale, while manually engineered parsers achieve high accuracy at the heavy expense of human maintenance.

Web content extraction is a vital component of reliable Large Language Model (LLM) data pipelines. However, traditional approaches face a trilemma of balancing accuracy, scalability, and adaptability. General-purpose extractors are broad but brittle when dealing with specific publisher layouts or complex extraction targets (like metadata, images, and tables). Direct LLM-based extraction is flexible yet costly and slow at scale, while manually engineered parsers achieve high accuracy at the heavy expense of human maintenance.

This paper introduces PACE (Publisher-Adaptive Content Extraction via Agentic Automation), an agentic framework designed to learn publisher-specific extraction configurations from representative pages and user requirements: * Training Phase: Utilizes LLMs to analyze page structures and synthesize reusable extraction patterns. * Inference Phase: Instantiates learned configurations onto a fixed, deterministic extractor template, enabling high-performance, scalable extraction without requiring additional LLM calls.

This paper introduces PACE (Publisher-Adaptive Content Extraction via Agentic Automation), an agentic framework designed to learn publisher-specific extraction configurations from representative pages and user requirements: * Training Phase: Utilizes LLMs to analyze page structures and synthesize reusable extraction patterns. * Inference Phase: Instantiates learned configurations onto a fixed, deterministic extractor template, enabling high-performance, scalable extraction without requiring additional LLM calls.

Experiments across article bodies, metadata, images, and tables reveal that PACE significantly outperforms scalable non-manual baselines, closely approaching the quality of manually engineered publisher-specific parsers.

Experiments across article bodies, metadata, images, and tables reveal that PACE significantly outperforms scalable non-manual baselines, closely approaching the quality of manually engineered publisher-specific parsers.

Document Metadata

Document Metadata

Field Details
arXiv ID arXiv:2608.27466
Subjects Computation and Language (cs.CL), Artificial Intelligence (cs.AI)
Authors Zhanlin Liu, Munirathnam Srikanth
Submission Date July 20, 2026
DOI 10.48550/arXiv.2608.27466
Field Details
arXiv ID arXiv:2608.27466
Subjects Computation and Language (cs.CL), Artificial Intelligence (cs.AI)
Authors Zhanlin Liu, Munirathnam Srikanth
Submission Date July 20, 2026
DOI 10.48550/arXiv.2608.27466

Abstract

Abstract

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.

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.

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.

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.

Access & Resources

Access & Resources