A Computational View of the Cognitive Semantics of Spatial Prepositions

Patrick Olivier, Jun-ichi Tsujii

TL;DR

This paper proposes a computational model for understanding spatial prepositions (like "in front of," "behind," "left of") by drawing on cognitive linguistics. It translates abstract conceptualizations of space into a formal representation that can be used to automatically generate visual depictions of described scenes.

What Problem Does This Paper Address?

Understanding spatial language is a long-standing challenge in natural language processing. While many computational approaches focus on formal logic and truth conditions, this paper argues that a more fruitful path lies in cognitive linguistics, specifically Ronald Langacker's theory of Cognitive Grammar. This theory emphasizes how meaning is constructed through conceptualization and "conventional imagery" – the different ways we can mentally represent the same information. The authors aim to bridge the gap between these descriptive linguistic theories and computational implementation, tackling the problem of generating adequate meanings for spatial descriptions in specific situations.

Methods

The core of the paper's approach is a conceptual representation designed to capture the nuances of spatial prepositions. This representation is motivated by three key dimensions of "conventional imagery":

  • Level of Specificity: How much detail is encoded? For example, "near" is less specific than "in front of," which is less specific than "facing."
  • Scale and Scope of Predication: How does the size and context of objects affect the interpretation of a preposition? A "left of" relationship between two houses is different from the same relationship between two apples.
  • Perspective: The viewpoint of the observer (speaker) is crucial. The paper distinguishes between intrinsic (object-centered), deictic (speaker-centered), and extrinsic (environment-centered) reference frames for interpreting spatial relations.

They implement this by:

  1. Object Schemata: Representing objects with dimensional assignment parameters (DAPs) along orthogonal axes.
  2. Domain Coordinate System (DCS): A 3D Cartesian system mirroring the speaker's perceptual axes (r for left-right, y for front-back, z for up-down).
  3. Potential Field Model: Spatial constraints from prepositions are modeled as potential fields. These fields define a "zone of proximity" around a reference object, with directional constraints specifying preferred locations within that zone. Multiple prepositions are combined by linearly adding their respective potential fields.

Datasets

Not discussed in the paper. The paper focuses on a theoretical framework and its implementation for English spatial descriptions, using example sentences and figures for illustration rather than formal dataset evaluation.

Key Insights and Ideas

  • Meaning is Conceptualization: The paper strongly advocates for the cognitive linguistics view that meaning is fundamentally about how we conceptualize the world, rather than just objective truth conditions.
  • Imagery is Key: The way we "imagine" or "construe" a spatial scene (conventional imagery) is as important as the literal information conveyed. Different imagistic choices lead to different interpretations of the same spatial phrase.
  • Fuzzy and Gradual Constraints: Spatial prepositions don't define strict boundaries but rather fuzzy constraints. An object can be "in front of" another to varying degrees, and the system models this by using potential fields.
  • Reference Frames Matter: The interpretation of spatial prepositions heavily depends on the chosen reference frame (intrinsic, deictic, extrinsic), which the authors explicitly model.
  • Compositionality through Field Addition: Combining multiple spatial constraints (e.g., "the chair is in front of the desk and to the left of the window") is achieved by simply adding their corresponding potential fields, allowing for cumulative meaning.

Key Results

The paper demonstrates the utility of its approach by showing how it can automatically construct depictions of scenes described by verbal spatial descriptions. Figure 2 in the paper shows an example of a computer-generated depiction based on a verbal description, illustrating the system's ability to translate linguistic input into visual output. The authors claim their system represents a significant advance in generating possible interpretations and the quality of those interpretations compared to previous approaches.

Limitations and Future Work

  • Informal Approach: The authors acknowledge that their approach is "informal" and that prototype theory, central to cognitive semantics, is difficult to formalize.
  • Experimental Validation: While the constants in their potential field model are discussed in relation to scale effects, the precise values are suggested to require experimental investigation.
  • Limited Scope of Imagery: The system focuses on three specific dimensions of imagery (specificity, scale, perspective) out of six identified by Langacker.
  • Future Work: The paper suggests that their system is a "first attempt" and a "highly specialized implementation." Future work would likely involve expanding the range of spatial expressions handled, further refining the computational model, and potentially incorporating more formal methods for validation.

Summary updated

BibTeX
@inproceedings{olivier1994computational,
  author = {Patrick Olivier and Jun-ichi Tsujii},
  title = {{A Computational View of the Cognitive Semantics of Spatial Prepositions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1994},
  doi = {10.3115/981732.981774},
}
Figure