The acquisition of lexical semantics for spatial terms: A connectionist model of perceptual categorization

This thesis describes a connectionist model which learns to perceive spatial events and relations in simple movies of 2-dimensional objects, so as to name the events and relations as a speaker of a particular natural language would. Thus, the model learns perceptually grounded semantics for natural language spatial terms.; Natural languages differ--sometimes dramatically--in the ways in which they structure space. The aim here has been to have the model be able to perform this learning task for terms from any natural language, and to have learning take place in the absence of explicit negative evidence, in order to rule out ad hoc solutions and to approximate the conditions under which children learn.; The central focus of this thesis is a connectionist system which has succeeded in learning spatial terms from a number of different languages. The design and construction of this system have resulted in several technical contributions. The first is a very simple but effective means of learning without explicit negative evidence, taking positive instances of other concepts as implicit negative instances of the concept being learned, and deliberately weakening the evidence from these implicit negatives so as to reduce the effect of false implicit negatives in cases where concepts overlap. This method is shown to be effective even in situations involving a high degree of overlap among concepts. In addition, this thesis presents the notion of partially-structured connectionism, a marriage of structured and unstructured network design techniques capturing the flexibility afforded by unstructured networks, and the tractability in learning and improved generalization ability that result from highly structured network design. Finally, the idea of learning within highly specialized structural devices is introduced, with the purpose of restricting the search during learning to a set of options known in advance to be of possible relevance.; Scientifically, the primary result of the work described here is a computational model of the acquisition of visually grounded semantics. This model is shown to successfully learn terms for spatial events and relations from a range of languages with widely differing spatial systems, including English, Mixtec (a Mexican Indian language), German, Bengali, and Russian. The model exhibits prototype effects which roughly match human intuitions, and extensions to the core system model the linguistic phenomena of polysemy and deixis. While no claims are made regarding structure-to-structure correspondences between architectural elements of the model and structures in the brain, the model is informed by a number of neuroscientific results. And perhaps most importantly, the model does more than just recapitulate the data; it also generates a number of falsifiable linguistic predictions regarding the sorts of semantic features, and combinations of features, one might expect to find in lexemes for spatial events and relations in the world's natural languages.