A Generative Lexicon Perspective for Adjectival Modification

Patrick Saint-DizierCOLING 1998

TL;DR

This paper proposes an extension to the Generative Lexicon (GL) framework to better model adjectival modification, particularly for polysemous adjectives like "bon" (good). By incorporating elements of the noun's Qualia structure (specifically the Telic role) into semantic composition rules, the authors aim to explicitly capture how adjective-noun combinations derive their meaning, moving beyond simple type coercion.

What Problem Does This Paper Address?

The paper tackles the challenge of accurately representing the meaning of adjective-noun compounds, especially when the adjective is polysemous and its meaning is heavily influenced by the noun it modifies. Traditional approaches, including earlier versions of the Generative Lexicon, often rely on general type coercion mechanisms. However, these can be too broad and don't fully explain the nuanced semantic shifts observed with adjectives like "good," which can mean "functional," "morally sound," "satisfying," or "accurate" depending on the noun. The authors argue for a more explicit integration of the noun's inherent properties (its Qualia structure) to constrain and define the adjective's meaning in context.

Methods

The core of the paper's methodology lies in extending the Generative Lexicon (GL) framework. Key aspects include:

  • Qualia Structure Integration: The paper emphasizes the Telic role of the noun's Qualia structure. This role describes the purpose or function of an object. The authors propose incorporating predicates from this Telic role directly into semantic composition rules.
  • Lexical Conceptual Structure (LCS): They adopt and slightly enhance Jackendoff's LCS framework for representing underspecified meaning, which they find more appropriate than simple type-based representations for capturing semantic polymorphism.
  • Selective Projection: This concept, building on GL's selective binding, describes how the adjective's meaning is projected from the noun's Telic role. The specific meaning of "bon" is determined by the type of the noun and the predicates within its Telic role.
  • Typed A-calculus: Underspecified structures are represented using a typed A-calculus.
  • Rule-Based Composition: Semantic composition rules are developed that explicitly use the Qualia structure to derive the meaning of adjective-noun compounds, rather than relying solely on general inference rules.

Datasets

Not discussed in the paper. The authors mention using a sample of French technical texts for analysis and a comparison with WordNet, but these are not presented as formal datasets for training or evaluation in the typical machine learning sense.

Key Insights and Ideas

  • The Telic Role is Crucial: The Telic role of a noun's Qualia structure is identified as a primary source for constraining and defining the meaning of modifying adjectives, especially for polysemous ones like "good."
  • Beyond Type Coercion: While type coercion is useful, it's insufficient for capturing the fine-grained semantic shifts. Explicitly linking adjective meaning to the noun's inherent purpose (Telic role) provides a more precise mechanism.
  • "Good" is a Case Study: The analysis of "bon" (good) reveals multiple distinct senses (functional, moral, intensifier, quantifier, accurate) that are systematically explained by its interaction with different noun types and their Telic roles.
  • LCS for Underspecification: Jackendoff's LCS framework is well-suited for representing the underspecified nature of adjective meanings before they are fully resolved by the noun.
  • Metaphors and Metonymies are Constrained: The paper suggests that metaphors and metonymies involving adjectives are not arbitrary but are often constrained by the noun's Qualia structure and type, allowing for a more systematic treatment.

Key Results

The paper presents a conceptual framework and analysis rather than quantitative results. The key "results" are the successful demonstration that:

  • The Qualia structure, particularly the Telic role, can be effectively integrated into semantic composition rules to explain adjectival modification.
  • A more precise semantic representation can be achieved for polysemous adjectives by linking their meaning to the specific properties and functions of the modified noun.
  • The proposed approach offers a more systematic way to handle sense variations, metaphors, and metonymies compared to more general type-shifting mechanisms.

Limitations and Future Work

  • Ontological Knowledge: The framework relies on ontological knowledge, which is acknowledged as an area still under development. The paper suggests that using existing ontologies like WordNet or EuroWordNet could be beneficial.
  • Hierarchy of Preferences: For nouns with multiple predicates in their Telic role, a hierarchy of preferences would be a useful extension to better reflect prototypicality.
  • Quantification: The paper notes that the "quantifier" sense of "bon" (sense 4) is distinct and not studied further in this work.
  • Generalization: While the analysis of "bon" suggests a common typology for adjectival modification across different adjectives and potentially languages, the extent of this generality and the rules' susceptibility to linguistic variation require further investigation.
  • Automatic Acquisition: The authors mention the potential for using on-line resources and dictionaries for automatic acquisition of conceptual descriptions, but this is presented as an ongoing task.

Summary updated

BibTeX
@inproceedings{saintdizier1998generative,
  author = {Patrick Saint-Dizier},
  title = {{A Generative Lexicon Perspective for Adjectival Modification}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1998},
}
Figure