Where Questions

Benny ShanonACL 1979 · DOI

Abstract

Answering a "where" question requires more than a simple retrieval of geographical facts; it necessitates a pragmatic algorithm that selects an appropriate level of detail based on the speaker's location and context. This paper evaluates and extends "Room Theory," demonstrating that human mental maps are significantly influenced by the fame of landmarks and emotional proximity. These factors force a shift between hierarchical (vertical) and landmark-based (horizontal) descriptions, providing a foundational framework for contextually aware question-answering systems.

Introduction

If a person in London asks, "Where is the Empire State Building?" the answer "On Earth" is semantically accurate but pragmatically useless. Conversely, providing a specific room number within the building would be an overwhelming and unnecessary level of detail. The core challenge in spatial communication is identifying the "pragmatic sweet spot"—the specific level of the geographical hierarchy that is most appropriate for the conversational context.

This research investigates how humans navigate these hierarchies and why standard models of spatial reasoning often fail to predict actual speech patterns. For researchers in Artificial Intelligence and Natural Language Processing (NLP), understanding this mechanism is essential for building systems that provide contextually relevant information rather than mere data retrieval.

Theoretical Framework: Room Theory

The study builds upon Room Theory, originally proposed by Rumelhart (1974). This theory posits that humans mentally organize space into a hierarchy of "rooms" (geographical regions). To answer a "where" question, a speaker identifies the smallest possible room that contains both themselves and the object of interest. The "correct" pragmatic answer is then the next smallest unit that contains the object but excludes the speaker’s current location.

Methodology

To test the limits of Room Theory, Benny Shanon conducted empirical studies using several methodologies:

  • Open Questionnaires: Subjects provided spontaneous answers to various "where" questions.
  • Closed Questionnaires: Subjects selected the "best" answer from a provided list or identified all "true" answers.
  • Diverse Populations: Data was gathered from college students in Israel and the United States, children of various ages, and aphasic patients. This diversity allowed the researcher to observe how spatial reasoning develops in childhood and how it may break down under clinical conditions.

Key Insights: Beyond Logical Hierarchies

While Room Theory provides a solid baseline, Shanon’s research reveals that human mental maps are not strictly logical. Several psychological effects distort the hierarchical process:

The Prominence Effect

Certain geographical entities possess "conceptual autonomy" due to their fame. For example, for many Americans, the Eiffel Tower is located "in Paris," but rarely described as being "in France." In the speaker's mental map, Paris is so prominent that it exists on the same conceptual level as a country, bypassing the standard hierarchy.

The Salience Effect

When an object is "emotionally dear" or physically very close to the speaker, the standard "vertical" (hierarchical) algorithm often fails. In these instances, speakers switch to "horizontal" descriptions, relating the object to another nearby landmark (e.g., describing a location as "North of Israel" rather than "in the Middle East").

Floor and Ceiling Effects

  • Floor Effect: When an object is in the speaker's immediate vicinity (such as a landmark in one's own town), the hierarchy becomes too cumbersome to navigate, leading to a preference for horizontal landmarks.
  • Ceiling Effect: For extremely large-scale objects like continents, the algorithm struggles because there are few "rooms" larger than a continent, often resulting in "vacuous" answers like "in the world."

Results and Discussion

The study concludes that Room Theory is a "by and large" model—it correctly predicts general patterns but cannot account for cultural and psychological distortions. Shanon proposes that question-answering is a two-step process:

  1. Appraisal: The speaker assesses the "configuration" (the spatial relationship between the speaker and the object).
  2. Algorithm Selection: The speaker selects the most appropriate strategy—vertical (hierarchical), horizontal (landmark-based), or featural.

Interestingly, in closed tests, subjects frequently chose "I don't know." This was often not a reflection of a lack of factual knowledge, but rather a pragmatic failure; the subjects could not find an answer that felt contextually appropriate.

Conclusion and Future Work

While this research provides a robust look at the topological aspects of spatial questions, it notes that a complete theory must eventually integrate speaker intention and social factors. Factors such as why the person is asking (e.g., for directions vs. general knowledge) and the perceived knowledge level of the listener play critical roles in how we define "where" something is. Future work in NLP must account for these complex configurations to move toward truly human-like spatial reasoning.

Summary updated

BibTeX
@inproceedings{shanon1979where,
  author = {Benny Shanon},
  title = {{Where Questions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1979},
  doi = {10.3115/982163.982183},
}
Figure