Mining Salient Properties from Similes and Queries

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 221 papers32 conceptsfrozen 2026-10-04

Traditional knowledge bases like WordNet are excellent at telling us what a concept is—a "dog" is a "canine," and "chocolate" is a "food." However, they are notoriously poor at capturing what a concept is like. They miss the "loyal" nature of the dog or the "addictive" quality of chocolate. These non-essential but highly salient properties are the bedrock of human common sense, metaphor, and creative reasoning.

Researchers have found that the best way to capture this knowledge isn't by reading encyclopedias, but by mining the specific linguistic "hooks" humans use to express their assumptions: similes and search queries.

The Problem of Reporting Bias

If you mine a massive corpus of news articles or Wikipedia, you will find thousands of mentions of "brown chocolate" but very few mentions of "sweet chocolate." This is reporting bias: people rarely state the obvious. In standard declarative text, we focus on the incidental or the newsworthy, not the stereotypical.

To find the "obvious" properties that define a concept's cultural identity, researchers like Tony Veale and Yanfen Hao argue we should look at similes. A simile like "as busy as a bee" or "as white as snow" explicitly pairs a concept with its most diagnostic attribute. Unlike a random adjective-noun pair, a simile represents an idealized, culturally agreed-upon association. By mining patterns like as [property] as a [noun] from the web, systems can identify the "low-hanging fruit" of common sense that formal dictionaries omit Comprehending and Generating Apt Metaphors.

Queries as Presupposition Engines

While similes provide property-noun pairs, search queries provide a window into human intent and presupposition. Marius Pasca demonstrated that the questions users type into search engines are a "cleaner" source of knowledge than the documents they eventually click on.

When a user asks "Why do pirates wear eye patches?", they aren't just asking a question; they are asserting a fact: pirates wear eye patches. This is a presupposition. By harvesting these "presupposition-laden" questions, an AI can learn that pirates wear eye patches, even if it never finds a declarative sentence stating it Creative Introspection and Knowledge Acquisition.

This approach scales remarkably well. Systems like Quasimodo, developed by Julien Romero and Simon Razniewski, use search engine auto-completions to extract millions of these salient properties, bypassing reporting bias by capturing the mundane facts humans implicitly assume when they seek information Commonsense Properties from Query Logs and Question Answering Forums.

From Static Hierarchies to Functional Ontologies

The goal of this research is to move from "static" ontologies to "functional" ones. In a static system, "axe" is always a "tool." But in a functional system, if we know an axe is "sharp," "heavy," and "dangerous" (properties mined from similes), we can dynamically re-categorize it as a "weapon" in a specific context Making Lexical Ontologies Functional and Context-Sensitive.

This "lateral thinking" allows for more human-like similarity judgments. For example, Tony Veale and Guofu Li developed Thesaurus Rex, which uses web-mined properties to explain why two concepts are similar—not because they share a parent in a tree, but because they share salient traits like "being cold" or "being fragile" Creating Similarity: Lateral Thinking for Vertical Similarity Judgments.

Applications in Creativity and Ranking

Once a system understands these stereotypical intensities, it can perform complex reasoning tasks:

  1. Intensity Ranking: If a system knows that "snow" is the prototypical anchor for "white," it can order other concepts (milk, paper, clouds) based on how closely they approach that intensity Ordering Concepts Based on Common Attribute Intensity.
  2. Creative Naming: By combining these salient properties with phonetic tools, systems can generate evocative brand names. For instance, knowing that "fragrance" and "rinse" share a functional context allows a system to suggest "Fragrinse" for a cleaning product A Computational Approach to the Automation of Creative Naming.
  3. Sentiment Spin: Salient properties allow a system to understand how a metaphor can "spin" a concept. Describing a person as "as cold as ice" uses a physical property to convey a social sentiment, a feat impossible without a map of stereotypical associations A Context-sensitive, Multi-faceted Model of Lexico-Conceptual Affect.

By focusing on the "how" and "why" of human description rather than just the "is" of formal logic, this research direction provides the connective tissue needed for AI to move from literal processing to figurative understanding.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 14 papers
  2. Version 1 · 2026-10-04 · 21 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_simile_based_property_extraction_13d6a593,
  title = {simile-based property extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/simile-based-property-extraction/rdr_13d6a5938f294de3bad94bbd45eb5431},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_mining_salient_properties_from_similes_a_13d6a593_v1,
  title = {Mining Salient Properties from Similes and Queries},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mining-salient-properties-from-similes-and-queries/rdr_13d6a5938f294de3bad94bbd45eb5431/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}