A Method of Calculating the Measure of Salience in Understanding Metaphors

Makoto Iwayama, Takenobu Tokunaga, Hozumi TanakaAAAI 1990

Abstract

This paper presents a formal mathematical bridge between information theory and cognitive linguistics to address the challenge of metaphor understanding in Natural Language Processing (NLP). We propose a computational method for quantifying "salience"—the prominence of a concept's traits—using Shannon entropy and property distributions across concept hierarchies. By calculating how typical and unique a property is, our system, AMUSE, automatically selects relevant traits to transfer during metaphor interpretation, effectively distinguishing between literal, metaphorical, and nonsensical statements.

Introduction

Metaphor understanding is a significant hurdle for NLP systems. When a human encounters a metaphor like "Mary’s cheeks are like apples," they instinctively transfer the property of redness while ignoring irrelevant traits like having seeds. Previous AI research identified "salience" as the key to this selection process but lacked a rigorous mathematical framework to calculate it. Without a formal definition, AI systems suffer from "property explosion," attempting to map every possible trait from a source to a target, and fail to reliably distinguish meaningful metaphors from semantic anomalies.

Methodology

The proposed method is built upon the foundation of Probabilistic Concepts, where a concept is defined not by binary attributes but by a set of attributes with probability distributions. For instance, an "Apple" might be represented as 80% red, 15% green, and 5% brown.

Salience is calculated using two primary factors derived from information theory:

  1. Amount of Information of Property (AIP): This measures "typicality." By calculating the entropy of a property's value distribution, we determine its redundancy. A property with a highly concentrated distribution (low entropy) is considered more typical and thus has a higher AIP.
  2. Difference of Property (DP): This measures "diagnostic significance." It compares a property's AIP against "sibling" concepts in an IS-A hierarchy. If a property like "redness" has high AIP for an apple but low AIP for most other fruits, it becomes a highly diagnostic (salient) feature.

Salience is defined as the product of these two: $Salience = AIP \times DP$.

Figure 1: A probabilistic hierarchy of the 'Fruit' concept, showing the distribution of shape and color attributes used to calculate salience.

The AMUSE system implements this via a four-step pipeline: Extraction (identifying target and source), Selection (filtering for high-salience properties), Mapping (finding corresponding target attributes), and Variance (adjusting the final mental model).

Key Insights and Results

The core contribution of this work is the treatment of salience as a statistical filter. High-salience properties are those that are both highly predictable within a category and highly distinctive between categories. For example, while "having seeds" is typical for an apple, it is common to almost all fruits, resulting in a low DP and low overall salience.

Mathematical Validation

The formula correctly identified that "redness" is more salient for an apple than "roundness," as many fruits share a round shape, but fewer share the specific red color distribution. In testing the "cheeks like apples" metaphor, AMUSE ranked "redness" (salience 0.20) and "shape" (0.12) above the threshold, while correctly discarding "having seeds" (0.06).

The Spectrum of Meaning

The authors argue that literal sentences, metaphors, and anomalies exist on a spectrum defined by Comprehensibility of Viewpoint (CV)—a ratio of information successfully transferred. Literal sentences approach a CV of 1.0, while semantic anomalies approach 0.

Limitations and Future Work

Currently, the model relies on a static hierarchy and does not account for dynamic shifts in salience caused by context (e.g., a poem vs. a grocery list). Furthermore, the system requires pre-defined probabilistic data. Future research will explore using incremental clustering algorithms like COBWEB to learn these probabilities from raw data and investigate the scalability of DP calculations in massive, real-world knowledge bases.

Summary updated

BibTeX
@inproceedings{iwayama1990method,
  author = {Makoto Iwayama and Takenobu Tokunaga and Hozumi Tanaka},
  title = {{A Method of Calculating the Measure of Salience in Understanding Metaphors}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure