Antonymy and Conceptual Vectors

Didier Schwab, Mathieu Lafourcade, Violaine PrinceCOLING 2002

TL;DR

This paper introduces a novel method for identifying and representing antonyms within a conceptual vector framework, crucial for improving natural language processing tasks like meaning representation and text generation. By defining antonymy as a form of symmetry within conceptual vectors, the authors propose functions to compute antonym vectors and evaluate the degree of antonymy between lexical items.

What Problem Does This Paper Address?

Representing meaning in Natural Language Processing (NLP) is a fundamental challenge. While thematic aspects and conceptual vectors have shown promise, existing methods often overlook the crucial phenomenon of negation and antonymy. This paper aims to integrate the study of antonymy into thematic analysis by developing a model compatible with conceptual vectors. The goal is to improve NLP systems by better handling negation, enabling more nuanced meaning representation, and facilitating applications like generative text, where understanding opposing concepts is vital for tasks such as paraphrase generation or summarization.

Methods

The core of the paper lies in defining antonymy within the context of conceptual vectors. The authors propose that two lexical items are in an antonymy relation if there's a symmetry between their semantic components relative to an axis. They identify three types of symmetry media: scalar values, true-false relevance, and cultural opposition. For this paper, they focus on "complementary antonymy," exemplified by pairs like "presence/absence."

Key methodological contributions include:

  • Conceptual Vectors: Representing textual segments (words, phrases, documents) as vectors in a high-dimensional space, where components correspond to activated concepts. These vectors are learned from human usage dictionary definitions.
  • Angular Distance: A measure DA(A, B) based on the cosine similarity of two vectors, quantifying their thematic proximity. A smaller distance (e.g., ≤ 45 degrees) indicates thematic closeness.
  • Antonymy Functions:
    • AntiR(Vitem, Vc): Computes an antonym vector for Vitem in the context of Vc, emphasizing the ideas present in Vitem.
  • AntiC(Ci, Vcontext): Returns the complementary antonym vector for a concept Ci given a context Vcontext, drawing from a manually constructed list of antonym vectors (AVC).
  • Antonymy Evaluation Measure (MantiEval): A pseudo-distance measure to assess whether two lexical items can reasonably be antonyms in a given context. It's calculated as the angular distance between the sum of the two item vectors and the sum of their respective antonym vectors.

Datasets

The paper mentions using a thesaurus for French (Lar92) with 873 concepts for bootstrapping their system. They also refer to a system that counts over 71,000 items for French, with more than 288,000 vectors, of which 2,000 items are concerned by antonymy. These items are either defined through negative sentences or have antonyms directly listed in the dictionary. No new datasets are explicitly introduced or benchmarked.

Key Insights and Ideas

  • Antonymy as Symmetry: The fundamental insight is reframing antonymy not just as an opposite, but as a specific type of symmetry within the semantic space defined by conceptual vectors. This allows for a more formal and computational approach.
  • Context is Crucial: Antonymy is highly context-dependent. The paper emphasizes the need to consider the context (represented by a vector) when determining antonyms, as a word can have different antonyms depending on the situation.
  • Antonymy is Not Exact Inverse of Synonymy: The antonymy evaluation measure suggests that antonyms share some related concepts, making them "negative synonyms" rather than completely unrelated words.
  • Handling Polysemy: The authors acknowledge the difficulty of polysemy in selecting the correct meaning for antonym computation. Their approach uses contextualization to favor specific meanings.
  • Discovering Non-Obvious Antonyms: The method can potentially uncover antonyms for seemingly non-opposable items (like material objects) by focusing on their opposable properties or categories.

Key Results

The paper presents qualitative results and evaluations of their proposed measures.

  • Improved Coherence: An evaluation measure (MantiEval) shows an improvement in the coherence of term vectors after applying their antonymy functions. For instance, the measure between 'existence' and 'non-existence' decreased from 0.44 to 0.33, indicating they are stronger antonyms.
  • Angular Distance Improvement: The authors report an improvement in angular distance through MantiEval ranking to 0.1 radian on 10% of the concerned items.
  • Examples of Antonym Vectors: The paper provides examples of computed antonym vectors for concepts like 'death' and 'life', showing that the closest antonym is 'DEATH' for 'LIFE', with a non-null distance, highlighting the approximate nature of the method.

Limitations and Future Work

  • Manual Construction of AVC: The Antonym Vectors of Concepts (AVC) list is manually constructed and requires validation and expansion by linguists.
  • Computational Cost: The antonymy evaluation measure's computation cost is noted as prohibitive (around a minute on a P4 at 1.3 GHz), limiting its practical use for large-scale lexicon exploration.
  • Polysemy Challenges: While addressed, polysemy remains a significant challenge, especially for the absolute antonymy function, where context selection can be problematic.
  • Future Work: The authors suggest several avenues for future research:
  • Revising and expanding the AVC lists with broader linguistic validation.
  • Learning antonymy relationships from corpora.
  • Exploring improvements to the antonymy measure and functions.
  • Applying the developed functions to text generation and other NLP tasks.

Summary updated

BibTeX
@inproceedings{schwab2002antonymy,
  author = {Didier Schwab and Mathieu Lafourcade and Violaine Prince},
  title = {{Antonymy and Conceptual Vectors}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure