Introduction
Extracting structured data from the web at scale is a significant challenge due to the sheer variety of website layouts. Traditionally, this process involves "wrapper induction"—learning rules, such as XPath expressions, to identify specific data fields within HTML. However, standard induction methods require extensive manual human labeling, creating a bottleneck that prevents scaling across thousands of domains. This paper presents a noise-tolerant framework that automates this process by leveraging imperfect, automated annotations.
The Challenge of Noisy Supervision
Websites often use scripts to generate pages from underlying databases, resulting in highly regular structures. While automated tools like dictionaries of business names can be used to label data quickly, these tools are inherently "noisy." They frequently miss items, label incorrect text, or introduce ambiguities. In traditional wrapper induction systems, even a single incorrect label can lead to "over-generalization," where the algorithm produces rules that extract irrelevant data or "junk."
The core problem addressed here is building extraction rules that are robust enough to ignore this noise and identify the true underlying data structure of a page.
Methodology: Probabilistic Wrapper Induction
The authors propose a "generate-and-test" framework that treats wrapper induction as a probabilistic inference problem. Instead of treating labels as ground truth, the system seeks the most likely data mapping ($X$) given a set of noisy labels ($L$). This is modeled by maximizing the product of two components:
- Annotation Model ($P(L|X)$): This component accounts for noise by modeling the probability that an imperfect annotator would produce the observed labels if $X$ were the actual data.
- Web Publication Model ($P(X)$): This evaluates the "goodness" of a candidate wrapper based on structural regularity. It assumes that high-quality wrappers will produce lists with consistent schema sizes and visual alignments.
To handle the exponential number of possible label subsets, the researchers developed two efficient search algorithms: BottomUp, a general-purpose black-box algorithm, and TopDown, a high-performance algorithm optimized for feature-based inductors like XPath.
Experimental Evaluation
The framework was evaluated across three domains: DEALERS (330 store locators), DISC (15 music discography sites), and PRODUCTS (10 cellphone shopping sites). Labels were generated using large, pre-existing databases and dictionaries, simulating a real-world automated extraction scenario.
Key Results
The noise-tolerant framework (NTW) demonstrated significant improvements over "naive" approaches that trust all labels. On the DEALERS dataset, NTW achieved nearly 1.0 precision and recall, whereas the naive approach suffered from a sharp drop in precision due to over-generalization.
Furthermore, the TopDown algorithm proved to be highly efficient, typically finding optimal wrappers in under a second per website. This performance makes the system viable for large-scale production environments where thousands of sites must be processed.
Conclusion
By shifting the focus from perfect training data to structural consistency, this research enables web-scale data extraction without the need for manual supervision. The system remains robust even when automated annotators have low recall or mediocre precision, provided the target pages exhibit the regular structure typical of script-generated content. Future work may explore extending this framework to handle more irregular page layouts and automated schema discovery.
Summary updated