Word learning as the interaction of statistical association and dynamic competition
McMurray, B., Apfelbaum, K. S., Toscano, J. C., & Samuelson, L. K. (2010, March). Paper presented at the 2010 International Conference on Infant Studies, Baltimore, MD.
Abstract:
Research on word learning is riddled with meta-theoretical claims about the difficulty of the problem–the amount of ambiguity; the rate of acquisition; non-obvious or conceptual information sources—that point toward specialized mechanisms attuned for lexical development. These are based on intuitions about what a reasonable learning system can do. Yet, we still do not know what a reasonable learning system is.
We present a reasonable model of learning based on simple, universal principles. It starts from an old idea: associative learning. However, to be reasonable, we need to consider a second factor—behavior. Infants don’t just learn, they act (sometimes intelligently). Thus, something new—dynamic competition—was added to model children’s referent selection given constraints from their knowledge and the situation.
Association was insufficient for learning in this model. However, buttressed by competition, the network learned (via cross-situational statistics: Siskind, 1996) all the words it was trained on, even when 90% of them were present on any trial. Thus, referential ambiguity is not a problem. It shows many classic findings: acceleration; disparities between comprehension and production; changes in familiar word recognition, and fast-mapping by mutual exclusivity. All can be described as the interaction of statistical associations and online competition. Competition also overcame a second challenge for associative learning: learning multiple names for one object. This model can learn both basic-level and super-ordinate names simultaneously, although it has a basic-level advantage. Finally, we demonstrate the generality of these principles by accounting for findings in early speech perception.
This suggests that slow and relatively weak associative learning could be buttressed by competition to yield intelligent lexical behavior—a sort of reversal of the classic performance/competence debate. It also suggests that until we know the power of such simple mechanisms, we shouldn’t assume the need for more complex developmental processes.