[Paper Review] Automatic Wrappers for Large Scale Web Extraction

Nilesh Dalvi, Ravi Kumar+1 more|arXiv (Cornell University)|Mar 12, 2011

Web Data Mining and Analysis17 references11 citations

TL;DR

This paper presents a novel framework for automatic wrapper induction that achieves high-accuracy web information extraction at web scale by making wrapper algorithms robust to noisy, unsupervised training data. By leveraging cheaply generated noisy labels from dictionaries and regular expressions, the system eliminates the need for costly site-level supervision, enabling fully unsupervised training and achieving state-of-the-art accuracy in large-scale web extraction, as deployed in production at Yahoo!.

ABSTRACT

We present a generic framework to make wrapper induction algorithms tolerant to noise in the training data. This enables us to learn wrappers in a completely unsupervised manner from automatically and cheaply obtained noisy training data, e.g., using dictionaries and regular expressions. By removing the site-level supervision that wrapper-based techniques require, we are able to perform information extraction at web-scale, with accuracy unattained with existing unsupervised extraction techniques. Our system is used in production at Yahoo! and powers live applications.

Motivation & Objective

• Address the challenge of high-cost site-level supervision in wrapper induction for large-scale web data extraction.
• Enable accurate information extraction from web pages without requiring manually labeled training data.
• Develop a framework that tolerates noisy, automatically generated training data from heuristics like dictionaries and regular expressions.
• Achieve production-grade accuracy in web-scale information extraction using unsupervised learning techniques.
• Bridge the gap between unsupervised extraction methods and the accuracy required for real-world deployment.

Proposed method

• Design a wrapper induction framework that is resilient to noise in training data, allowing the use of automatically generated, low-quality labels.
• Use weak supervision signals such as dictionaries and regular expressions to generate large-scale, noisy training examples.
• Apply a noise-tolerant learning algorithm that identifies and corrects inconsistencies in the noisy training data during wrapper training.
• Leverage a probabilistic or discriminative model to learn extraction rules from the noisy data while minimizing error propagation.
• Integrate the trained wrapper into a scalable pipeline suitable for processing millions of web pages.
• Validate the framework’s robustness and accuracy through evaluation on real-world web data and comparison with existing unsupervised methods.

Experimental results

Research questions

• RQ1 Can wrapper induction be made robust enough to handle large volumes of noisy, automatically generated training data?
• RQ2 To what extent can site-level supervision be eliminated in web extraction while maintaining high accuracy?
• RQ3 How effective are heuristic-based labeling methods (e.g., dictionaries, regex) when combined with noise-tolerant learning?
• RQ4 Can a fully unsupervised wrapper induction system achieve production-level accuracy on real-world web data?
• RQ5 What is the performance gain of the proposed framework over existing unsupervised or weakly supervised web extraction techniques?

Key findings

• The proposed framework achieves high extraction accuracy without requiring any site-level supervision, enabling fully unsupervised training.
• The system successfully learns from noisy training data generated via dictionaries and regular expressions, significantly reducing labeling costs.
• The framework demonstrates superior accuracy compared to existing unsupervised extraction techniques, as validated in real-world deployment at Yahoo!.
• The method scales effectively to web-scale data, supporting large-scale information extraction in production environments.
• The framework is deployed in production at Yahoo! and powers live applications, confirming its robustness and practical viability.
• The approach outperforms prior unsupervised methods by effectively mitigating noise in training data while preserving extraction precision.