Statistical Cross-Situational Learning in Adults and Infants

We learn words in ambiguous contexts, with multiple word candidates for any referent and multiple referent candidates for any word. For example, a child may see a boy, a bat, a ball, and a dog and hear “Look at the boy. The dog wants his ball.” This is the word-to-world mapping problem (e.g. Gleitman, 1990; Bloom, 2000; Smith, 2000). How could a learner who knows no words associate object names with the right referents? Developmentalists have studied a number of solutions to this problem, including ways in which the mature partner limits words and referents and directs attention to the relevant referent (Baldwin, 1993; Tomassalo, 2000), and internal perceptual and conceptual constraints (Genter, 1982). This paper is concerned with an alternative solution, cross-situational statistical learning, a process in which statistics are calculated across different learning instances to determine across multiple experiences, the most likely word-referent mappings. We are also interested in how internal constraints, such as whole object assumption or mutual exclusivity, may be realized or embedded in these mechanisms.