Paper in preparation. Some of the material will be presented at GALA 1997: Language Acquisition: Knowledge Representation and Processing, University of Edinburgh.


Language Acquisition: Learning and Applying Probabilistic Constraints

Mark S. Seidenberg, Joseph Allen & Morten H. Christiansen

Program in Neural, Informational and Behavioral Sciences
University of Southern California
University Park MC-2520
Los Angeles, CA 90089-2520


Abstract

We will describe a new framework for thinking about language acquisition that is emerging from renewed interest in statistical and probabilistic aspects of language, insights from connectionism, and recent behavioral studies of the learning capacities of infants. Standard approaches to child language equate knowledge of language with a competence grammar. The child's task is grammar identification and standard poverty of the stimulus arguments suggest that this task is intractable unless there is significant innate grammatical knowledge. The newer approach has three main components. First, there are analyses of the actual utterances to which children are exposed suggesting that they contain a rich set of probabilistic cues to different aspects of linguistic structure. This kind of statistical information has been excluded from competence grammar ever since "colorless green ideas sleep furiously." Second, there is theoretical work on how such cues can be extracted, represented, and combined. Much of this research draws on connectionist concepts of knowledge representation, learning, and processing. The constraint satisfaction process that such networks implement shows how probabilistic cues can be combined in powerful ways. Even though particular cues may not be highly reliable in isolation, combinations of such cues yield non-linear increases in their informativeness. Such networks provide a computationally explicit interpretation of the concept of "bootstrapping." Finally, behavioral studies of infants are beginning to show that they rapidly and effortlessly pick up this kind of statistical information from a very young age (e.g., Saffran, Newport & Aslin, in press). Recent applications of this framework include the Christiansen, Allen & Seidenberg's (in press) work on the word segmentation problem and Allen's (1996) work on the acquisition of verbs and their argument structures.

This framework raises a number of important issues. First, it calls into question many deeply-held assumptions about the nature of language acquisition. In particular, many of the standard poverty of the stimulus arguments no longer apply. For example, the approach explains how children could converge on essentially the same knowledge of language despite variability in experience; and, there is no negative evidence problem because acquisition is driven by analyses of statistical regularities in the input, not feedback about the grammaticality of utterances. Second, the approach provides closer ties between acquisition and skilled performance. We ask how the child acquires the capacity to comprehend and produce utterances and draw upon theories of adult performance (e.g., MacDonald, Pearlmutter & Seidenberg, 1994) in framing questions about language learning. Finally, there are questions about the adequacy of the framework: can it explain a broad range of facts about the structure of language and how children acquire it; is it compatible with other things that are known about the brain bases of language and about the relationship between language and other aspects of cognition?

Click to request a copy of this paper.

rule.gif (155 bytes)

Home | People | Research | Links | Contact | Publications | Presentations
Cognitive Neuroscience Laboratory

Please email suggestions/errors to mhc27@cornell.edu