Class-Driven Attribute Extraction

Benjamin Van Durme, Ting Qian, Lenhart K. SchubertCOLING 2008 · OpenReview

TL;DR

This paper introduces a novel method for automatically extracting attributes of concept classes from text, even when explicit examples of those classes are scarce. By leveraging compositional language processing on the British National Corpus and a modified version of the KNEXT system, they demonstrate that using class labels directly in text can yield attribute extraction quality comparable to methods relying on web query logs, highlighting the value of explicit textual mentions.

What Problem Does This Paper Address?

Acquiring knowledge about concepts (like "US President" or "Animal") is crucial for building intelligent systems. Traditional approaches often rely on pre-compiled lists of representative instances (e.g., "Bill Clinton," "George Bush" for "US President") to infer attributes. However, this method struggles with classes that lack readily available instance lists or when attributes are expressed as unary properties (e.g., "illegal" for "Drug," "warm-blooded" for "Animal"), which are often conveyed through prenominal adjectival modification. This paper tackles the challenge of extracting both binary and, importantly, unary attributes by directly utilizing class labels found within text, aiming to improve attribute acquisition, especially for classes with limited explicit examples.

Methods

The core of the methodology lies in a modified version of the KNEXT system, a framework for extracting structured knowledge into Episodic Logic (EL). The system processes syntactic parses of sentences from the British National Corpus (BNC) to generate logical forms.

Key methodological innovations include:

  • Compositional Language Processing: Applying linguistic analysis to derive logical forms from sentence structures.
  • Leveraging Class Labels: Directly using class names (e.g., "Animal") as anchors for attribute extraction, rather than solely relying on instances.
  • Unary Attribute Extraction: A specific focus on extracting attributes expressed as "X CAN BE Y" (e.g., "AN ANIMAL CAN BE WARM-BLOODED"), often missed by instance-based methods.
  • Filtering Non-Compositional Phrases: Developing sophisticated filtering mechanisms (whitelist, graylist, blacklist) to remove spurious or nonsensical attribute-class pairings that arise from idiomatic expressions or ambiguous phrases. This involves analyzing adjective-noun relationships and using external knowledge sources like Wikipedia.

Datasets

  • British National Corpus (BNC): The primary dataset used for extraction. The authors chose the BNC for its breadth of genre and substantial size (100 million words).
  • Wikipedia: Used to construct a blacklist of non-compositional phrases.
  • WordNet: Utilized for classifying words and identifying adjectival uses.
  • Brown Corpus: Used in the initial creation of the whitelist for calculating part-of-speech tagged bigram counts.
  • Web 1T 5-gram collection: Used for examining trigram frequencies to refine blacklist filtering.

No new benchmarks were introduced; instead, the paper focuses on applying and extending existing extraction frameworks to new types of attributes and leveraging a large, publicly available corpus.

Key Insights and Ideas

  • Class Labels are Powerful: Explicitly mentioning class labels in text (e.g., "the class Animal") is a surprisingly effective source for attribute extraction, yielding results comparable to methods using web query logs. This challenges the notion that only instance-based learning is effective.
  • Unary Attributes are Underserved: Traditional methods often overlook unary attributes (like "warm-blooded" for "Animal") because they are typically expressed through prenominal adjectives, which are not easily captured by instance-based inference. This paper makes significant strides in acquiring these.
  • Filtering is Crucial: The quality of extracted attributes heavily depends on robust filtering of non-compositional phrases and ambiguous word senses. The development of a multi-tiered filtering system (whitelist, graylist, blacklist) is a key contribution to improving precision.
  • Sense Ambiguity is a Major Hurdle: Even with good extraction, word sense ambiguity (e.g., "Mars" as a planet vs. a Roman deity) can lead to incorrect class assignments for attributes, highlighting the need for deeper semantic understanding.

Key Results

  • The proposed class-driven attribute extraction method achieved comparable quality to methods using web query logs.
  • Experimental results suggest that exploiting explicit occurrences of class labels in text is a valuable strategy.
  • The filtering mechanisms significantly improved the quality of extracted unary attributes. For example, the "Planet" class saw a noticeable drop in quality when only using gazetteers, highlighting the importance of the filtering and class-label-based approach.
  • The paper presents qualitative comparisons of extracted attributes for various classes (e.g., BasicFood, Religion, Painter) showing the effectiveness of their approach.

Limitations and Future Work

  • Sense Disambiguation: The authors acknowledge that word sense ambiguity remains a significant challenge, leading to incorrect attribute-class mappings (e.g., "bars" for "Planet" derived from "Mars bars").
  • Non-Compositional Phrase Filtering: While advanced, the filtering lists were not fully automated, suggesting room for further development in this area.
  • Sense-Specific Extraction: The paper points to the need for more advanced techniques beyond simple filtering, such as relative discounting of patterns and improved sense disambiguation, to refine extraction quality.
  • Discovering New Topics: The ability to discover "missing topics" (i.e., identifying potential Wikipedia entries based on extracted phrases) is highlighted as a promising area for future research.

Summary updated

BibTeX
@inproceedings{durme2008class,
  author = {Benjamin Van Durme and Ting Qian and Lenhart K. Schubert},
  title = {{Class-Driven Attribute Extraction}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2008},
  url = {https://openreview.net/forum?id=Iv776Vbu7C},
}
Figure