INCIDENTAL LANGUAGE LEARNING : Listening ( and Learning ) out of the Corner of Your Ear

Two experiments investigated the performance of firstgrade children and adults on an incidental language-learning task. Learning entailed word segmentation from continuous speech, an initial and crucial component of language acquisition. Subjects were briefly exposed to an unsegmented artificial language, presented auditorily, in which the only cues to word boundaries were the transitional probabilities between syllables. Subjects were not told that they were listening to a language, or even to listen at all; rather, they were engaged in a cover task of creating computer illustrations. Both adults and children learned the words of the language. Moreover, the children performed as well as the adults. These data suggest that a statistical learning mechanism (transitional probability computation) is able to operate incidentally and, surprisingly, as well in children as in adults. Language acquisition by children is an instinctive, and apparently effortless, process, which typically occurs amidst a vast array of other sensory and intellectual experiences. The process of acquiring language is thus incidental in the sense that the child's primary task is presumably understanding, rather than acquiring, language (e.g., Chomsky, 1975; Krashen, 1985). Moreover, language is acquired by infants and young children, who are unlikely to be engaged in explicit, conscious learning. However, little research has examined the characteristics of incidental learning pertinent to the process of children's language acquisition. Outside the realm of natural language acquisition, there are several experimental literatures that have investigated the incidental learning of complex patterns. One potentially relevant phenomenon is implicit learning. Introduced by Reber (1967), implicit learning is claimed to involve the unconscious and unintentional acquisition of abstract information (see Reber, 1993, for an extensive review). Although a number of controversies have emerged in discussions of implicit learning (e.g., Dulany, Carlson, & Dewey, 1984; Shanks & St. John, 1993), it is clear throughout this literature that subjects are, at least to some degree, able to induce certain aspects of the structure of patterned stimuli incidentally. A second set of phenomena that may bear on the process of incidental language learning is found in the literature on frequency estimation. It has been suggested that the frequency of events present in the environment is a fundamental type of information that is encoded in memory incidentally (see Hasher & Zacks, 1984, for an overview). Information about event frequency is acquired by humans across a broad range of natural and experimental situations, and is maintained even when there is no reason to remember the events in question (Hasher, Zacks, Rose, & Sanft, 1987). Among the many features of implicit learning and frequency estimation, two characteristics in particular suggest that the phenomena, and their underlying mechanisms, may be related to one another. First, both types of learning appear to be ageinvariant, with young grade-school-aged children and adults demonstrating equivalent performance on these tasks (e.g., N.R. Ellis, Palmer, & Reeves, 1988; Hasher & Chromiak, 1977; Roter, 1985, cited in Reber, 1993). Such findings of age invariance stand in sharp contrast to most other phenomena in developmental psychology, for which the most obvious and gross generalization is that performance improves with age. A second shared characteristic appears in the types of mechanisms hypothesized to underlie these two types of learning. Knowledge of event frequency involves statistical computations, performed either on the input or across memory representations. Similarly, it has been argued that implicit-learning phenomena are based on learning mechanisms that capitalize on the statistical structure of the input (e.g., Cleeremans, 1993; Perruchet & Pacteau, 1990; Reber, 1993; ServanSchreiber & Anderson, 1990; Stadler, 1992). To what extent can an understanding of these two learning processes shed light on natural language acquisition? Traditionally, the process of language acquisition has been viewed as distinct and qualitatively different from learning of other types (e.g., Chomsky, 1965; Osherson & Wasow, 1976). Most models of acquisition are formulated in specifically linguistic terms, with little attempt to relate even the earliest stages of this process to mechanisms capable of acquiring other types of patterned information. However, recent findings in language acquisition focus on some of the same characteristics as do the literatures on implicit learning and frequency estimation. Unlike most other aspects of cognitive development, language acquisition does not favor adults over children. Rather, the initial stages of language acquisition are characterized by age invariance. In later stages, learners who began as children surpass those who began as adults (e.g., Johnson & Newport, 1989; Krashen, Long, & Scarcella, 1982; Newport, 1990; Slavoff & Johnson, 1995). Moreover, recent computational models of language acquisition suggest that at least some pertinent language-learning mechanisms may induce structure from the input using statistical methods similar to those suggested by incidental-learning research (e.g., Brent & Cartwright, 1996; Christiansen, 1994; Cleeremans, Servan-Schreiber, & McClelland, 1995; Elman, 1990; Mintz, Newport, & Bever, 1995; Schutze, Address correspondence to Jenny R. Saffran, Department of Brain and Cognitive Sciences, University of Rochester, Rochester, NY 14627; e-mail: [email protected]. VOL. 8, NO. 2, MARCH 1997 Copyright © 1997 American Psychological Society 101 This content downloaded from 141.161.13.195 on Wed, 26 Aug 2015 17:57:37 UTC All use subject to JSTOR Terms and Conditions PSYCHOLOGICAL SCIENCE Incidental Language Learning 1994). These similarities suggest a potentially substantial relationship between mechanisms of incidental learning and those of natural language acquisition. We therefore decided to ask more directly whether at least the first stages of language acquisition could be shown to involve all of these characteristics: to be accomplished incidentally, to show age invariance, and to employ statistical computations. Those aspects of language acquired by infants are particularly good candidates for investigations of incidental learning, as infants are unlikely to be engaged in explicit, directed learning. Moreover, the linguistic knowledge acquired during the 1st year is largely the result of distributional analyses of the input, including the vowel space (Kuhl, Williams, Lacerda, Stevens, & Lindblom, 1992), consonant categories (Werker & Tees, 1984), phonotactic rules (Jusczyk, Friederici, Wessels, Svenkerud, & Jusczyk, 1993), phonological regularities (Jusczyk, Cutler, & Redanz, 1993), and frequent biphones (Jusczyk, Luce, & Charles-Luce, 1994) of the native language. These early abilities suggest that infants are adept at performing incidental computations of statistical information about their native language. This, in turn, suggests that probing the incidental nature of language learning may render insights about the role of learning in language acquisition. The particular aspect of acquisition that we investigated concerned word segmentation. Before children can begin to acquire syntax, they must discover the words of their language, a process complicated by the fact that the speech stream is mostly continuous, without consistent pauses or other acoustic cues marking word boundaries (Cole & Jakimik, 1980). Although adults are faced with this problem whenever they are confronted with a novel word embedded in fluent speech, they can use the surrounding familiar words as markers of where the novel word begins and ends. The prelinguistic infant does not have this luxury, at least not initially. Despite the difficulty of segmenting words from continuous speech, however, experimental evidence indicates that infants can succeed at word segmentation tasks by 8 months of age, well before the onset of word production (Jusczyk & Aslin, 1995). One proposed solution to the word segmentation problem is that infants may be able to exploit statistical cues to word boundaries (Aslin, Woodward, LaMendola, & Bever, 1996; Brent & Cartwright, 1996; Hayes & Clark, 1970; Saffran, Newport, & Aslin, 1996). Across a language sample, sounds that co-occur within words tend to be more highly correlated with one another than sound pairs spanning word boundaries. We (Saffran et al., 1996) have suggested a computational mechanism for using these contrasts to discover word boundaries. To take a simple example, consider the word sequence prettybaby. Pre is followed by ty with some probability. Ty, however, is followed by ba rarely, in particular, only when a word ending in ty happens to be followed by a word beginning with ba. Thus, the transitional probability from one sound to the next will generally be highest when the two sounds follow one another word-internally; transitional probabilities spanning word boundaries will tend to be relatively low. We (Saffran et al., 1996) demonstrated, in an explicit-learning task, that adult subjects were able to use transitional probabilities to learn the multisyllabic words of an artificial language presented as a synthesized speech stream containing no other cues to word boundaries. In the present study, we asked whether this purely statistical word segmentation task can be accomplished incidentally, and, in addition, whether children might perform comparably to adults. The experiments tested a fairly extreme version of the incidental-learning question, by investigating whether this aspect of early language learning can be achieved while subjects are focused on an entirely unrelated task. Incidental-learning studies typically require the subject to perform some task involving the relevant information, such as memorizing the stimulus strings or predicting the next item in a string. Even studies in which learning occurs solely by observation nonetheless require subjects to attend to some aspect of the relevant stimuli (e.g., Hasher et al., 1987; Reber & Allen, 1978). The present study, however, utilized a cover task in which subjects created computer illustrations, and were told nothing about the language stimuli except that an audiotape that would be playing in the background might affect their artistic creativity. Subjects were not told that the tape consisted of a language, nor that they would be tested in any way during the course of the experiment. Any learning that occurred was thus doubly incidental, in that attention was directed neither to the word segmentation task nor to the acoustic stimuli forming the words. The question of interest was whether word units could be discovered under these conditions, and whether children could approach the performance level of adults on such a complex task.