English Infants Learn Italian: a Statistical Word Segmentation Study

Previous research has shown that infants are equipped with powerful statistical learning mechanisms that they can use during word segmentation tasks (e .g ., Saffran et al ., 1996; Aslin et al ., 1998) . Evidence comes from artificial language studies, where ‘words’ are defined solely by their transitional probabilities (TP) . Artificial languages lack the auditory complexity and acoustic informa- tion found in natural languages . To what extent do these studies scale up to real-world language acquisition? To address this question, we tested whether infants use TP for segmenting a naturally-produced foreign-language. Methods: English-learning 8-month-old infants were exposed to 2 minutes of infant-directed Italian . In addition to over 100 words with variable length, the narrative contained 2 disyllabic words with high internal TP (melo & fuga) and 2 words with low TP (casa & bici) . Both word types were presented with equal frequency, and were equated acoustically . Using a headturn procedure, infants in Experiment 1 were tested on high TP words vs . novel words (tema & pane) . In Experiment 2 infants were tested on high TP vs. low TP words. Half of the infants were randomly assigned to counterbalanced versions of the language. Results: In both experiments infants showed a familiarity preference, listening longer to words with high TP that to novel (Exp . 1, p<.01) or low TP words (Exp . 2, p<.01). Conclusion: These results demonstrate that infants are able to use transition probability even in the presence of complex linguistic input, such as an unknown language . Disyllabic words in both Italian and English tend to have a trochaic stress pattern, which may have aided infants’ segmentation . However, given that all test words conformed to the same stress pattern, TP remains the most informative cue to differentiate the various word types . Results support the hypothesis that infants use TP during real-world language learning