A Perceptual Basis for Noun Phrase Syntax

Abstract

Human language is the result of both biological and cultural evolution. To have the best chance of understanding language we must seek all the constraints of that evolution. In the first part of my thesis I propose the general hypothesis that the sensorimotor system is one of those constraints and argue that regardless of whether language is the result of biological evolution, cultural evolution, or both, we should expect idiosyncrasies of the sensorimotor system to be reflected in linguistic structure. The bulk of the thesis explores a particular version of this hypothesis – namely that visual attention and classification of objects are reflected in noun phrase syntax. Within the noun phrase the noun stem (e.g. “dog”) and number morphology (e.g. “-s”) are contributed by separate syntactic elements; I argue that this reflects a separation of functionality in the sensorimotor system. To begin an exploration of this hypothesis I draw upon existing models of visual attention by Itti and Koch (2000) and object classification by Mozer and Sitton (1998), adapting and combining them into a new computational model. The key new idea in the model is that object classification is cardinality blind which means its output is the same whether presented with one token of a class or many tokens of the class. This allows groups of similar objects to be handled at once. I implement a model of classification which, like primate object classification, is location invariant. In my model cardinality blindness emerges naturally from location invariance. I argue the same thing happens in primates, reviewing neurophysiological evidence for this. To cater for a cardinality-blind classifier I also implement extensions to a standard model of visual attention. The combined classifier and attentional models elegantly reproduce a number of human results, including Gestalt grouping by similarity, global precedence (Navon, 1977) and the role of stimulus similarity in visual search (Duncan and Humphreys, 1989). These results show that the model does useful work in an account of the visual system. With the visual foundation established I propose a simple model of the interface between visual cognition and noun phrase syntax. Within my model the information corresponding to the noun stem is produced by the classifier and is cardinality blind so carries no number information. The information corresponding to singular or plural number morphology is produced separately by the attentional system. The decomposition of information in my model corresponds to the same decomposition of information in noun phrases. I conclude that cardinality blindness in the visual system can explain this aspect of noun phrase syntax, supporting the general hypothesis that natural language syntax reflects properties of the sensorimotor system and inviting further theories of this nature.