Turning Web Text and Search Queries into Factual Knowledge: Hierarchical Class Attribute Extraction

Marius PascaAAAI 2008

Abstract

This paper presents a scalable, weakly supervised framework for constructing a hierarchical knowledge base of class attributes. By leveraging both unstructured Web documents and search query logs, the system extracts thousands of open-domain classes and their associated properties. These classes are anchored into a conceptual hierarchy (WordNet), allowing for the mathematical propagation of attributes across levels. The approach demonstrates high precision and significant scale, requiring minimal human supervision—only five seed examples for a single reference class are needed to prime the entire extraction engine.

Introduction

Traditional Information Extraction (IE) systems often operate in a "closed-loop" fashion, effectively identifying instances of predefined categories like People or Organizations but failing to scale to the "long-tail" diversity of real-world concepts. Manually defining every possible class and its relevant attributes is an intractable task. Previous attempts to automate this process have often relied on Web documents or search logs in isolation, leading to either noisy or sparse results. This research addresses the challenge of scalability and structure by automatically discovering thousands of classes, linking them into a logical hierarchy, and identifying their most relevant properties with minimal human intervention.

Methodology

The proposed framework operates through a four-stage pipeline designed to maximize both coverage and precision by combining the breadth of the Web with the focused intent of search queries.

Instance and Class Extraction

The system first scans approximately 100 million Web documents using "Is-A" patterns (e.g., "...zoonotic diseases such as Venezuelan equine encephalitis..."). To mitigate the noise inherent in Web text, search query logs serve as a validator: candidate instances are discarded if they do not appear as standalone search queries. This ensures that the extracted entities are meaningful to human users.

Examples of open-domain classes and their instances extracted from Web documents and search logs.

Linking to Hierarchies and Attribute Extraction

Extracted classes are mapped to the WordNet hierarchy using a "first-sense" heuristic. Simultaneously, the system mines search logs for attributes by identifying patterns where users search for an instance and a property together (e.g., "cast of [Movie Name]"). This dual-source approach combines the breadth of the Web with the cleaner signal of user intent found in search logs.

Hierarchical Propagation

A key innovation of this work is the mathematical propagation of attributes up the hierarchy. Attributes identified for specific sub-classes (e.g., "side effects" for "painkillers") are promoted to broader concepts (e.g., "Drug") if they are shared by a significant number of sub-concepts. This smoothing process improves precision by filtering out sporadic errors and identifying properties that are universally relevant to a category.

The mapping of attributes to a conceptual hierarchy, illustrating how properties are promoted to higher-level nodes like 'Organism' or 'Person'.

Results

The system successfully extracted over 9,000 open-domain classes containing roughly 200,000 unique instances. Evaluation on a set of 25 target concepts showed that the system achieved a precision of 0.64 at rank 10 and 0.46 at rank 50. Notably, the use of automatically extracted instances from the Web outperformed the use of sparse, manually curated instances from WordNet alone, highlighting the importance of large-scale data mining for knowledge base construction.

Discussion and Conclusion

While the system demonstrates remarkable scale, challenges remain regarding sense ambiguity and determining the optimal level of abstraction for attribute placement. Future work may explore more sophisticated word-sense disambiguation techniques. Nonetheless, the ability to "prime" an entire extraction engine with just five seed attributes (population, area, president, flag, climate) for a single class (European countries) represents a significant step toward truly scalable, open-domain knowledge acquisition.

Summary updated

BibTeX
@inproceedings{pasca2008turning,
  author = {Marius Pasca},
  title = {{Turning Web Text and Search Queries into Factual Knowledge: Hierarchical Class Attribute Extraction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure