Mapping Qualitative Spatial Language to Quantitative Geometry

309 concepts195 papers1 proposal

When a human says "the keys are on the table" or "the coffee shop is just past the park," they are communicating rich spatial arrangements using qualitative, fuzzy, and context-dependent terms. Computers, however, operate in a world of rigid mathematical formalisms: $(x, y, z)$ coordinates, pixel grids, and bounding boxes.

Bridging this gap is one of the central challenges in spatial natural language processing. As Wu et al. observe in On the Description-based Spatial Positioning and the Associated Uncertainty, human spatial language lacks fixed geometric boundaries, creating a fundamental incompatibility with traditional Geographic Information Systems (GIS) and robotic engines. To build systems that truly understand human intent—whether navigating an autonomous vehicle or generating a 3D scene from text—we must understand how to translate qualitative linguistic expressions into quantitative geometric representations.

The Context-Dependency of Prepositions

The simplest approach to parsing spatial language might seem to be hard-coding geometric rules: defining "on" as a vertical offset or "near" as a fixed radius. However, cognitive and computational research shows that this strategy fails because spatial prepositions are deeply tied to object function and real-world context rather than pure geometry.

In her foundational work On the Spatial Uses of Prepositions, Herskovits demonstrates that choosing a preposition involves calculating object function, observer perspective, and relative mobility. For instance, Chang et al. point out in Learning Spatial Knowledge for Text to 3D Scene Generation that the word "on" implies a vertical surface when placing a poster, but a horizontal surface when placing a cup.

Furthermore, spatial terms carry different cognitive profiles. Viechnicki et al. show in Large-Scale Bitext Corpora Provide New Evidence for Cognitive Representations of Spatial Terms that "functional" spatial terms (which involve concepts like force and containment, such as in) vary significantly more across contexts and languages than purely "geometric" spatial terms (which rely on simple axes and vectors, such as above). Because human language is elastic, rigid computational models inevitably break down.

From Binary Logic to Probabilistic Fields

If a system treats a term like "near" as a binary boundary (e.g., "near means within 50 meters"), it creates an unnatural threshold where an object at 51 meters suddenly becomes "far." Humans evaluate spatial relations continuously and intuitively.

To capture this, researchers shift from binary logic to continuous probabilistic fields. Gapp argued in Basic Meanings of Spatial Relations: Computation and Evaluation in 3D Space that computational models must represent how an object can be "more" or "less" in a certain position, rather than relying on strict classifications.

Rigid Binary Model (Fails context shifts):
[Target] ---- 50m ----> | Boundary | -> Not Near

Probabilistic Field Model (Adapts to context):
[Target] ~~~~~~~~~~~~~~~~~~~~~~~~ (Smooth Gradient of Likelihood) ~~~~~~~~~~~~~~~~~~~~~~~~>
         Urban Context: Drops off quickly
         Rural Context: Drops off slowly

A powerful way to operationalize this intuition is through Gaussian Mixture Models (GMMs) and probabilistic heat maps. Skoumas et al. demonstrate in Location Estimation Using Crowdsourced Geospatial Narratives that qualitative descriptions from text can be modeled as spatial probability distributions.

Crucially, these distributions are elastic. As the authors of On Quantifying Qualitative Geospatial Data: A Probabilistic Approach show, a probabilistic model for "near" automatically stretches or contracts based on geographic context—a "short walk" covers a tiny area in a dense urban center but spans kilometers in a rural environment.

Intermediate Formalisms: Qualitative Spatial Calculi

Sometimes, jumping directly from a vague word to a coordinate map is too large a leap. Instead, systems can translate natural language into an intermediate symbolic language—a qualitative spatial calculus—that captures logical relationships without needing exact measurements.

Dan highlights this strategy in From Spatial Relations to Spatial Configurations as a way to bridge linguistic flexibility with formal spatial logic. By mapping prepositions to standardized topological frameworks like Region Connection Calculus (RCC8) or the Dimensionally Extended 9-Intersection Model (DE-9IM), systems can reason logically about space.

As Gao explains in Foundation models for geospatial reasoning: assessing the capabilities of large language models in understanding geometries and topological spatial relations, this process maps vague vernacular phrases (like "along the river") into structured geometric predicates that a computer can evaluate. However, this mapping remains a challenge; Kordjamshidi et al. note in CLEF 2017: Multimodal Spatial Role Labeling Task Working Notes that because human language is polysemous, a single spatial indicator can map to multiple distinct formal relations depending on the surrounding context.

Disambiguation Through Multi-Modal Context and Progressive Refinement

If someone says "the building is near the park," the search area is massive. But if they add "and it's just north of the fountain," the intersection of those two fuzzy regions narrows down the target dramatically.

To resolve the inherent ambiguity of isolated phrases, modern architectures leverage multi-modal data and progressive refinement. For example, Kordjamshidi et al. use global logical constraints over text and images in Spatial Language Understanding with Multimodal Graphs using Declarative Learning based Programming to solve Spatial Role Labeling (SpRL), identifying which objects serve as "trajectors" (the item being located) and "landmarks" (the reference point).

When visual data is unavailable, systems can use sequential linguistic constraints. Zheng et al. propose a "progressive approximation" method in Natural Language Description of Geographic Location Considering Various Spatial Relations with Different References, combining multiple relative landmarks to shrink an infinite directional wedge into a precise target area.

Similarly, in generative tasks, Mittal et al. utilize scene graphs in Interactive Image Generation Using Scene Graphs to incrementally build and modify spatial layouts, transforming step-by-step text instructions into coherent physical environments. By intersecting multiple vague descriptions, the system gradually eliminates ambiguity, arriving at a precise geometric destination.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_spatial_language_ambiguity_18e120a5,
  title = {spatial language ambiguity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatial-language-ambiguity/rdr_18e120a5cb4341d1b67fa566b3b2eead},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_spatial_language_ambiguity_18e120a5_v1,
  title = {spatial language ambiguity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatial-language-ambiguity/rdr_18e120a5cb4341d1b67fa566b3b2eead/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake