Constraints on the Learning of Spatial Terms: A Computational Investigation

This chapter explores three possible constraints on the learning of spatial terms. It demonstrates the constraints that are proposed for empirically obtained data concerning the perception of spatial relations, and the learning of spatial terms from English, Dutch, Russian, and Mixtec. It discusses that acceptability judgment for directional spatial terms such as “above” is modeled as a simple linear combination of Gaussian orientational alignment features. The simulations explains that the formulation provides a close fit to empirically collected “above” acceptability judgments, but does not fit other, artificially created, datasets. This suggests that the model's original success is not due to any excessive flexibility of the mechanism, but rather to the appropriateness of the structures proposed. It explains the mutual exclusivity constraint, when adopted as an initially weak bias, can effectuate the learning of spatial terms. Finally, it suggests that the child's categorizations made at end-event may come to inform those made at startevent. This dependency may account for semantic systematicity in terms for motion-into and motion-out in the world's languages. While the eventual fate of this idea remains to be determined by further empirical inquiry, there is some initial empirical support for it. The constrained approach to the learning of spatial terms occupies a position intermediate between two extremes. At one extreme, the hypothesis space is so tightly constrained as to render the learning task trivial, while at the other the space is virtually unconstrained, and the task is consequently much more difficult.