The Manipulation of Images to Handle Indeterminacy in Spatial Reasoning

Abstract

In this article, we propose a novel method for using images to do spatial reasoning. The task we consider is to answer questions about descriptions involving spatial prepositional phrases. Psychological experiments suggest that imagery is an appropriate cognitive model of how people solve such problems when descriptions are determinate. However, natural language discourse often contains indeterminate descriptions, and imagery has been criticized for its apparent inability to represent incomplete information necessary for making inferences in such situations. We propose a new image-based method of reasoning, called ISR for indeterminacy in spatial reasoning, which dynamically constructs and inspects multiple images to reason about spatial prepositional phrases. The consideration of more than one consistent image, as a model of the description, increases the accuracy of reasoning. However, because it is intractable to consider every possible consistent image, ISR uses several heuristics to generate only the most relevant images, admitting some inaccuracy. Thus, ISR exploits the computational trade-off between efficiency and accuracy, and we show empirically in a representative domain that this trade-off is effective. The ISR method demonstrates that, with the help of specialized procedures, imagery can be more accurate for reasoning about spatially indeterminate descriptions. However, as we observe in attempting to extend ISR to a scaled-up domain, the procedural encoding of knowledge hinders the maintenance of an effective computational trade-off. We conclude that imagery is distinct from axiomatic/deductive approaches only in having heuristic knowledge about space for making approximations.