Carla L. Hudson Kam

Professor, Department of Psychology
University of California, Berkeley

"Learning Probabalistic Languages: Who Learns What, When, and Why"

Monday, April 5, 2004

12:00-1:30 PM

The Beach Room - 3105 Tolman

Statistical learning, an emerging approach to language acquisition, focuses on the mechanisms available to language learners. Initial studies asked whether learners could extract information about language structure present in the input and use it to arrive at correct generalizations. Repeatedly, the answer has been yes, indicating the existence of very powerful learning mechanisms. However, this power must be limited, or learners could end up tracking too much information, much of it irrelevant. This then suggests that the learning mechanisms are constrained. The goal of the work described here is to investigate the constraints.

I present adults and children (ages 5-7) with artificial languages containing inconsistent, probabilistic patterns. These patterns are unlike natural human languages, and therefore might be expected to be difficult to learn. After several exposure sessions, participants perform a variety of tests, including an elicited production task. The question is whether the learner extracts the variation veridically, or fails to learn these abnormal patterns. Moreover, what do her productions look like when she fails? Does she make the language more like natural languages, i.e. consistent?

Findings showed that adults learned the variation veridically when it was simple, but failed when it was more complex. When they failed, adults primarily regularized the language by boosting the frequency of the most frequent input form, for instance, producing the 60% form 80% of the time. In contrast, children typically failed to learn even simple variation. Sometimes they changed the language just like adults, by boosting the frequency of the most common form. But children also imposed other patterns, for example, using no form, or imposing patterns based on what appear to be linguistic categories, e.g. treating subjects differently than objects, even though there was nothing like this pattern in the input.

I end by hypothesizing that the primary cause of regularization is the same in adults and children: Patterns that are difficult to learn lead to regularization, but children and adults find different things hard due to differences in their information processing capacities. Thus, regularization may be due to domain general constraints. However, the nature of the regularization differed between adults and children; some children, but no adults, produced regularizations that were very linguistic, suggesting that the ultimate outcome of learning is dependent on multiple factors, some of which may be specific to language.

Food and Refreshments will be served.