A Uniied Model of Lexical Acquisition and Lexical Access

This paper presents the rst uniied model of how adults and young children use whatever words they know to segment utterances and isolate novel words. This model is based on a conception of the language processor as a system for reasoning under conditions of uncertainty about the language being heard, the linguistic structure of the current utterance, and the portion of the current utterance that has not yet been heard. The fundamental hypothesis is that the language processor performs both acquisition tasks and sentence analysis tasks simultaneous, by computing the most probable combination of hypotheses about the language and hypotheses about the current utterance, given the available information. Within this framework, a speciic processing is task modeled by (a) making a hypothesis about the information that is most important for performing that task, (b) formalizing the remaining uncertainty in probabilistic terms, 0 Brent, Lexical acquisition and lexical access 1 and (c) deriving a mathematical expression for the most likely analysis of the language and the current utterance, given this information. The current seg-mentation model provides an intuitive explanation for lexical bootstrapping, and computer simulations have shown that it can help explain children's ultimate success at segmentation and word discovery. Further, it yields empirical predictions about oo-line word discovery and on-line lexical activation. Brent, Lexical acquisition and lexical access 2 In the speech signal, there does not appear to be any acoustic analog of the blank spaces that mark word boundaries in many writing systems. Thus, the input to language acquisition is more like a sequence of utterances than a sequence of words. How young learners segment utterances and discover the sounds of words has therefore become a topic of considerable interest. Among language-acquisition researchers, the focus of research has been constraints on the forms of words in particular languages. Both uncertain (\soft") and absolute (\hard") constraints have been discussed. For example, the predominant stress pattern for bisyllabic words in English is strong-weak (e.g., mm ommy); weak-strong words (e.g., alll ow) are substantially rarer (Cutler & Carter, 1987). There is now evidence that English-learning 9-month-olds know this (Jusczyk, Cutler, & Redanz, 1993a) and show a greater tendency to segment out strong-weak words than weak-strong words (Jusczyk, Newsome, & Hous-ton, 1995). Likewise, it is clear that infants could use the phonotactic regularities of their languages, such as the fact that English words cannot begin with gd, to avoid erroneous seg-mentations such as …