Axiomatic Characterization of Synonymy and Antonymy

H. P. EdmundsonCOLING 1967

TL;DR

This paper proposes an axiomatic framework to formally define synonymy and antonymy using relation algebra. By treating these as ternary relations and establishing a set of axioms, the authors provide a rigorous foundation for understanding these linguistic concepts, paving the way for more robust computational linguistics models.

What Problem Does This Paper Address?

Understanding and computationally modeling the relationships between words, specifically synonymy (words with similar meanings) and antonymy (words with opposite meanings), is a fundamental challenge in linguistics and natural language processing. Previous approaches often relied on explicit definitions that varied significantly, leading to inconsistencies. This paper aims to bridge this gap by developing a formal, axiomatic system that can precisely characterize these relationships, moving beyond intuitive notions to a mathematically grounded approach. The authors highlight that existing dictionary definitions and early computational models struggled with the nuances of these relationships, particularly how context and part-of-speech influence meaning.

Methods

The core of this paper is the development of an axiomatic characterization of synonymy and antonymy using the calculus of relations.

  • Relation Algebra: The authors leverage the power of relation algebra, a mathematical system for manipulating relations. They define binary relations (like identity and null relations) and operations (like complement, union, intersection, converse, product, and power).
  • Axiomatic Definitions: Instead of explicit definitions, synonymy ($S_1$) and antonymy ($A_1$) are defined implicitly through a set of axioms. These axioms are expressed in terms of relation algebra.
  • Synonymy Axioms: Reflexive, Symmetric, Transitive (making it an equivalence relation).
  • Antonymy Axioms: Irreflexive, Symmetric, Antitransitive, and Right-identity.
  • Ternary Relations: Crucially, the paper posits that synonymy and antonymy are fundamentally ternary relations (involving three elements, though the formalization focuses on pairwise relationships within a specific context or "intension"). This acknowledges that the relationship between two words can depend on a third factor (like a specific sense or context).
  • Topological Models: Beyond the algebraic approach, the paper explores two topological models for synonymy:
  • Neighborhood Topology: Defines a topology based on "neighborhoods" of words, where a neighborhood is a subset of a word's synonym class containing the word itself. This model satisfies standard neighborhood axioms.
  • Closure Topology: Defines a topology based on a closure operation, where the closure of a set of words is its synonym class. This model satisfies standard closure axioms.

Datasets

Not discussed in the paper. The paper focuses on theoretical formalization rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Synonymy and Antonymy as Relations: The paper firmly establishes synonymy and antonymy as formal relations that can be manipulated using mathematical tools like relation algebra.
  • The Power of Axiomatization: Defining these linguistic concepts implicitly through axioms, rather than explicitly, avoids the ambiguities and inconsistencies inherent in natural language definitions. This provides a robust and consistent framework.
  • Ternary Nature of Relations: The insight that synonymy and antonymy might be best understood as ternary relations (even if formalized pairwise) is significant. It acknowledges that the meaning of "synonym" or "antonym" can be context-dependent.
  • Algebraic Relationships: The derived theorems reveal elegant algebraic relationships, such as $A^2 = S$ (the square of the antonymy relation is the synonymy relation) and $A^{2k} = S$ for even powers, and $A^{2k+1} = A$ for odd powers. This suggests a deep structural connection between the two concepts.
  • Topological Spaces for Semantics: The exploration of topological models offers a novel way to represent semantic relationships without relying on metric distances, which have historically been difficult to define for language.

Key Results

  • Axiomatic System: A consistent set of axioms for synonymy and antonymy is proposed and expressed in relation algebra.
  • Derived Theorems: Numerous theorems are proven, demonstrating the internal consistency of the axiomatic system and revealing interesting properties. For instance:
  • $S^{-1} = S$ and $A^{-1} = A$ (both relations are symmetric).
  • $S^n = S$ for any $n \ge 1$ (synonymy is idempotent under product).
  • $A^2 = S$ and $A^{2k} = S$, $A^{2k+1} = A$ for $k \ge 1$.
  • $A \neq S$ (synonymy and antonymy are distinct).
  • Synonym Classes: The synonym class of a word is shown to be non-empty and forms an equivalence relation, partitioning the vocabulary.
  • Antonym Classes: The antonym class of a word is also shown to be non-empty.
  • Topological Properties: Both neighborhood and closure topologies are shown to be satisfied by the synonymy relation, providing a geometric interpretation of semantic spaces.

Limitations and Future Work

  • Independence of Axioms: The paper states that the independence of the axioms is not treated.
  • Model Consistency: While the consistency of the axioms is shown by exhibiting a model, the paper does not delve into the construction of specific models or their empirical validation.
  • Focus on Content Words: The formalization is restricted to "content words," and the paper notes that the treatment of "function words" (like "and," "the") might require special consideration or additional axioms.
  • Antonymy Topology: Topological models are only considered for synonymy, not antonymy.
  • Future Work: The paper suggests that the relation algebra can be used to produce shorter proofs of theorems and that the topological characterizations provide a nonmetric representation of "semantic space," avoiding the difficulties of metric spaces. The authors imply that further exploration of these algebraic and topological models could lead to more advanced computational linguistic applications.

Summary updated

BibTeX
@inproceedings{edmundson1967axiomatic,
  author = {H. P. Edmundson},
  title = {{Axiomatic Characterization of Synonymy and Antonymy}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1967},
}

Loading related content...

Figure