A Fluid Knowledge Representation for Understanding and Generating Creative Metaphors

Tony Veale, Yanfen HaoCOLING 2008

TL;DR

This paper introduces "Talking Points," a flexible knowledge representation designed to capture the fluid nature of creative metaphors. By combining objective definitions from WordNet with stereotypical descriptions from the web, they build a scalable system that can generate and interpret metaphors, demonstrating improved clustering accuracy compared to previous methods.

What Problem Does This Paper Address?

Creative metaphors, like describing a burqa as "armor" or "a wearable cage," stretch semantic conventions to make the unfamiliar familiar or the mundane strange. Computational models need a knowledge representation that is equally flexible and accommodating to handle these fluid connections between concepts. Existing approaches often rely on rigid semantic rules or limited knowledge bases, struggling to capture the nuanced and often unconventional relationships that drive creative metaphor. This paper aims to bridge this gap by developing a scalable and flexible knowledge representation for metaphor processing.

Methods

The core of the paper is the "Talking Points" knowledge representation. These points are essentially descriptive fragments of concepts. The authors propose two complementary ways to acquire these points:

  • Objective Talking Points: Extracted from WordNet glosses, these represent consensus definitions. For example, from the gloss "the person who leads a musical group," they derive the talking point "leads:musical_group."
  • Stereotypical Talking Points: Harvested from web text, these capture how people actually speak about concepts, often through similes. They use patterns like "as ADJ as a C" to find associations (e.g., "precise:surgeon") and then further refine these into "has ADJ:facet" relationships (e.g., "has_delicate:hands" for surgeon).

These talking points are then connected to form a "slipnet," a probabilistic network inspired by Hofstadter's work. This slipnet allows for "slippage," a process of conceptual re-representation where concepts can be linked through non-obvious paths, enabling metaphor generation and understanding.

Datasets

  • WordNet: Used for extracting objective talking points and for its hierarchical structure to identify semantic relationships.
  • Web Text (via Google): Used for harvesting stereotypical talking points, particularly through simile patterns and attribute extraction.
  • Existing Benchmarks: The paper replicates clustering experiments from Almuhareb and Poesio (2004, 2005) using their datasets of 214 and 402 nouns, respectively, to evaluate their talking points approach.

Key Insights and Ideas

  • The Power of Stereotypes: The paper strongly suggests that stereotypical descriptions, often dismissed as informal or inaccurate, are crucial for capturing how humans actually conceptualize and relate ideas, making them highly effective for metaphor processing.
  • Talking Points as Concise Representations: The "Talking Points" approach, especially with stereotypical descriptions, provides a surprisingly concise yet effective way to represent concepts, leading to better clustering results with significantly smaller feature sets compared to prior work.
  • Fluidity is Key: The "slipnet" structure, allowing for flexible connections and "slippage" between talking points, is essential for modeling the dynamic and often indirect nature of metaphorical reasoning.
  • Web Harvesting for Scale: Leveraging large-scale web data is presented as a viable and scalable method for acquiring the vast amount of descriptive knowledge needed for robust metaphor processing, moving beyond manually curated resources.

Key Results

The paper demonstrates the effectiveness of their "Talking Points" approach through clustering experiments, replicating and extending the work of Almuhareb and Poesio:

  • Experiment 1 (214 nouns): Their approach achieved a clustering accuracy of 90.2% when using both values and attributes (talking points), significantly outperforming Almuhareb and Poesio's baseline of 85.51%.
  • Experiment 2 (402 nouns): Their method achieved 69.85% accuracy using both values and attributes, which, while lower than the first experiment, still represents a strong performance, especially considering the scale and lack of filtering.
  • The results consistently show that combining both "property values" (adjectives) and "attribute nouns" into integrated talking points yields the highest clustering accuracy.

Limitations and Future Work

  • Scale of Knowledge Base: While the approach is scalable, the authors acknowledge that the knowledge base needs to grow by at least an order of magnitude to be truly comprehensive.
  • Radical Metaphors: The paper notes that highly creative or "ad-hoc" metaphors (like the "Governor as President" example) require more radical slippage patterns that are not fully captured by the current system.
  • Nested Propositions: The "Burqa" examples highlight the need for a more complex representation that can handle nested propositions within talking points.
  • Future Work: The authors aim to develop a general, corpus-grounded mechanism for explaining all metaphor-related slippage, moving beyond handling individual cases. They also anticipate that stereotypical descriptions will be a key area for future growth of the talking points knowledge base.

Summary updated

BibTeX
@inproceedings{veale2008fluid,
  author = {Tony Veale and Yanfen Hao},
  title = {{A Fluid Knowledge Representation for Understanding and Generating Creative Metaphors}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2008},
}
Figure