Exploring statistical learning by 8-month-olds: The role of complexity and variation

Introduction Recent studies suggest that young infants possess powerful statistical learning mechanisms that they can use to find word boundaries in fluent speech. For example, in a landmark study by Saffran, Aslin, and Newport (1996), the Headturn Preference Procedure was used to familiarize English-learning 8-month-olds to an artificial language containing no cues to word boundaries other than transitional probabilities between syllables. The language they used contained four trisyllabic words that were created by combining twelve CV syllables. The four trisyllabic words were concatenated into a continuous two-minute stream of speech containing no pauses between words (e.g.....pabikudaropitibudopabikugolatu....). This speech stream was intended to mimic a real language learning environment, since real speech does not contain reliably placed pauses between words either (see Jusczyk, 1997, for a thorough review of the word segmentation problem). In Saffran et al.’s stimuli no single word repeated itself in immediate succession, and no syllable was used in more than one word. Therefore, the transitional probability between two syllables within a word was equal to 1.0, whereas the transitional probability between two syllables crossing a word boundary was equal to .33 (if pabiku and daropi are words, then the transitional probability between bi and ku is 1.0, whereas the transitional probability between ku and da is .33). Saffran et al. (1996) found that after familiarization with this language, infants listened significantly longer to statistical part words (e.g. kudaro if pabiku and daropi are statistical words) than statistical words. Saffran et al. interpreted this novelty effect as evidence that infants are very good at statistical learning. In addition, they argued that statistical learning might play an important role in early word segmentation. Additional work by the same group of authors demonstrated that infants are also able to track conditional probabilities between syllables (Aslin, Saffran, and Newport, 1998). The finding that infants can segment an artifiical language containing no cues to word boundaries other than the strength of the transitional probabilities between syllables is an impressive finding. However, the language used by Saffran et al. (1996) was highly simplified with respect to natural language input. Thus, one cannot automatically assume that infants can used statistical cues to segment a natural language, especially given past studies demonstrating that statistical learning effects disappear rapidly as the complexity of stimulus patterns increase (Bruner, 1973). Therefore, in the following two experiments, we ask whether infants’ statistical learning abilities are robust enough to deal with a slightly more complex artificial language. The artificial language we use in the following two experiments will be very similar to that used by Saffran et al. (1996). The only difference is that the language used in this study, like real language, will contain words of variable length. Thus, instead of being familiarized to an artificial language containing four trisyllabic words, the infants in our study are familiarized with an