Adjectival Modification in Text Meaning Representation

Victor Raskin, Sergei NirenburgCOLING 1996

TL;DR

This paper introduces a novel "microtheory" for representing the meaning of adjectives within the MikroKosmos computational semantics framework. It proposes a principled approach to capturing adjectival meaning by linking it to ontological properties and scales, moving beyond traditional linguistic focus on syntactic distinctions. This work is crucial for building more robust and comprehensive natural language understanding systems.

What Problem Does This Paper Address?

The paper tackles the challenge of accurately representing the meaning of adjectives in computational linguistics, a category often overlooked in favor of more syntactically complex structures. Traditional linguistic analyses of adjectives tend to focus on their attributive vs. predicative usage, but this doesn't fully capture their semantic contribution. The authors argue that a more effective approach is needed to integrate adjectival meaning into a broader knowledge representation system, particularly for machine translation and large-scale NLP applications. The core problem is how to systematically derive and represent the semantic contribution of adjectives in a way that is both computationally tractable and semantically rich.

Methods

The authors propose an "ontological approach" to adjectival meaning representation within the MikroKosmos system. This approach relies on a "text meaning representation" (TMR) language, which is frame-based and grounded in an ontology that models world knowledge.

Key methodological aspects include:

  • Ontology-Driven Representation: Adjectives are linked to ontological properties and their associated scales. For scalar adjectives (like "big"), their meaning is represented as a value on a numerical scale corresponding to an ontological property (e.g., SIZE).
  • Lexical Mapping (LEX-MAP): Adjective entries in the lexicon contain a LEX-MAP that specifies how the adjective's meaning contributes to the TMR of the modified noun. This involves mapping adjective senses to ontological concepts or deriving them by modifying property constraints.
  • Syntactic-Semantic Dependency Mapping: A crucial link is established between the syntactic structure of an adjective-noun phrase and its semantic representation, often using special variables.
  • Distinction between Property-Based and Non-Property-Based Modification: The paper highlights a key distinction:
  • Property-Based: Adjectives directly modify a property of the noun (e.g., "big house" modifies the SIZE property of "house").
  • Non-Property-Based: Adjectives characterize the proposition or introduce attitudes, rather than directly modifying a noun property (e.g., "abusive speech" where "abusive" describes the event, or "good" which introduces an evaluative attitude).
  • Grain Size Principle: The authors advocate for a "principle of practical effability," suggesting that the level of detail in semantic representation should match what can be expressed in the target language, often favoring coarser-grained descriptions.

Datasets

The work is based on a corpus of approximately 6,000 English and 1,500 Spanish adjectives obtained from task-oriented corpora. The paper does not introduce new datasets but utilizes existing lexical resources.

Key Insights and Ideas

  • Scalar vs. Non-Scalar is Key: The most important distinction for semantic treatment of adjectives is not attributive/predicative usage, but whether the adjective represents a scalar property or not.
  • Ontology is Foundational: A robust ontology is essential for grounding adjectival meaning in world knowledge and enabling consistent representation.
  • Adjectives as Property Modifiers: For many adjectives, their core semantic contribution is to specify or constrain a property of the modified noun.
  • Beyond Direct Property Modification: Some adjectives (like "good" or temporal ones) don't directly modify a noun's property but rather express attitudes or characterize the entire proposition, requiring different representation strategies.
  • Practical Effability Matters for MT: The level of semantic detail should be guided by what is expressible in the target language for machine translation, leading to principled choices about grain size.

Key Results

The paper presents a detailed methodology for representing adjectival meaning, illustrated with numerous examples of lexical entries for adjectives like "big," "good," "authentic," "fake," and "medical." The authors demonstrate how their approach can handle both simple scalar modifications and more complex non-property-based cases. While quantitative results are not the primary focus, the paper showcases the practical application of their method within the MikroKosmos semantic analyzer, which has been tested on a substantial lexicon of English and Spanish adjectives.

Limitations and Future Work

  • Complexity of Non-Property-Based Modification: The paper acknowledges that representing non-property-based adjectival modification, especially those involving attitudes or event characterization, is more challenging and requires further development.
  • Salient Property Identification: Identifying the "salient property" for certain adjectives (like "good") is difficult and can be computationally expensive. The authors' approach aims to mitigate this by using broader attitude representations.
  • Extrapolation to Other Modifiers: The authors hypothesize that their property-based vs. non-property-based dichotomy might also apply to adverbial modification of verbs and other types of modification, suggesting this as a direction for future research.
  • Language Independence: While the core approach is language-independent, the specific "adjectival superentries" (combinations of meanings) and availability of adjectival expression for a meaning can vary between languages. Further research is needed to fully explore these cross-linguistic variations.

Summary updated

BibTeX
@inproceedings{raskin1996adjectival,
  author = {Victor Raskin and Sergei Nirenburg},
  title = {{Adjectival Modification in Text Meaning Representation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}

Loading related content...

Figure