Extracting Ontological Selectional Preferences for Non-Pertainym Adjectives from the Google Corpus

John J. Tanner, Fernando GomezAAAI 2010 · DOI

Introduction

Word Sense Disambiguation (WSD) remains a fundamental challenge in Natural Language Processing. While human language processing effortlessly resolves ambiguity—distinguishing "batter" as a culinary ingredient in "crispy batter" versus an athlete in "leadoff batter"—computational models often struggle with such localized context. Traditional WSD systems typically rely on broad context windows, frequently overlooking the precise grammatical relationships between adjectives and the nouns they modify. This paper presents a robust methodology for extracting "selectional preferences"—the statistical tendency for specific adjectives to modify particular noun categories—using the Google Web 1T corpus and a novel Scaled Lemmata Fraction (SSLF) algorithm.

Methodology

The proposed approach focuses on the top 3,000 non-pertainym adjectives, utilizing WordNet to map nouns into a hierarchical ontology.

Type Counting and Ontological Mapping

A critical distinction in this work is the use of "type counting" over "token counting." Rather than aggregating every occurrence of an adjective-noun pair, the system counts unique noun lemmata within a category that co-occur with a given adjective. This prevents high-frequency nouns from disproportionately biasing the preference scores for an entire ontological category.

The Scaled Lemmata Fraction (SSLF)

The core innovation, SSLF, addresses the varying "decay rates" in frequency across different noun categories. By scaling the denominator of the probability fraction to account for low-frequency members, the system mitigates biases toward common categories. Furthermore, scores are normalized against a baseline—the frequency of nouns following the word "the"—to distinguish genuine selectional preference from general linguistic frequency.

Experimental Evaluation

The system was evaluated using the Google Web 1T 5-gram corpus for co-occurrence data and Wikipedia for real-world testing pairs. A pseudo-homonym test was employed, where the system must choose the correct noun sense when paired with a "confounder" that does not typically co-occur with the adjective.

Performance Results

The Scaled Lemmata Fraction consistently outperformed existing state-of-the-art methods, such as the LKM method (Lapata et al.).

Table 1: Performance comparison (F1 scores) across adjective frequency bands for various methods.

As shown in Table 1, SSLF achieved an F1 score of 83.6% in the highest frequency band (< 1K), a significant improvement over the 72.1% baseline. Notably, even the unscaled "Lemmata Fraction" showed superior performance to previous complex models, validating the type-based approach.

Confidence and Precision

The study also analyzed the relationship between the "score gap"—the difference between the preference scores of the correct and incorrect senses—and the resulting precision.

Table 2: Mean precision and distribution of pseudo-homonyms relative to the score difference.

The data indicates that as the score difference increases, precision improves dramatically. For cases where the score difference exceeded 3.0, the system achieved near-perfect precision (99.9%), although these high-confidence cases represent a smaller fraction of the total dataset.

Discussion and Future Work

While highly effective, the system faces challenges with "systemic polysemy" (e.g., "orange" as both fruit and tree) and occasional organizational gaps in WordNet. Additionally, "weak selectors"—adjectives like "large" or "nice" that modify a broad range of nouns—remain difficult to disambiguate. Future research will explore extending this type-based scaling methodology to other grammatical dependencies, such as verb-subject and verb-object relationships, to further enhance ontological preference extraction.

Summary updated

BibTeX
@inproceedings{tanner2010extracting,
  author = {John J. Tanner and Fernando Gomez},
  title = {{Extracting Ontological Selectional Preferences for Non-Pertainym Adjectives from the Google Corpus}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7719},
}
Figure