|
Abstract:
Children‟s language learning is remarkable not only for the speed and apparent ease with which it happens, but also for the complexity of the input on which learners must operate. Consider the case of learning concrete nouns, which make up the preponderance of early vocabularies (Macnamara, 1972). Children must solve at least two significant problems from scratch: speech segmentation – identifying the individual words in continuous speech, and word-object mapping – determining which of the segmented words should be mapped to which candidate referent in the environment. In both problems, learners must contend with significant ambiguity of the information available in natural learning environments. For instance, the majority of everyday spoken language, even child-directed speech, is comprised of multi-word utterances (Brent & Siskind, 2001). Because spoken utterances lack an acoustic analog of white spaces to mark word boundaries, the segmentation task is nontrivial. Similarly, a single word-learning context often contains multiple candidate words and multiple object referents. In order to learn the meanings of these words, young word learners must determine which of the word-object pairs are correct mappings and which are spurious. One way that human learners could contend with the ambiguity of a single instance is through accumulation of statistical evidence across instances. Saffran, Aslin, and Newport (1996) showed evidence for such a mechanism in 8-month-old infants, who are sensitive to transitional probabilities between syllables in continual speech. This sensitivity supports statistical speech segmentation. Recently, Smith and Yu (2008) demonstrated the availability of a statistical solution to the word-object mapping problem in 12 and 14-month-old infants (see also Gleitman, 1990). Infants at this age show sensitivity to cooccurrence frequencies between words and objects, a sensitivity which may underlie early word learning. Although such mechanisms are known to be available to young infants, the exact nature of their interaction remains an open question. Graf Estes, Evans, Alibali, and Saffran (2007) demonstrated a potential serial link by exposing 17-month-old infants to a speech segmentation task followed by word-
|