When to Hold and When to Fold: Detecting Structural Changes in Statistical Learning

When to Hold and When to Fold: Detecting Structural Changes in Statistical Learning Benjamin D. Zinszer (bdz107@psu.edu) Center for Language Science Departments of Psychology & Statistics, Penn State University University Park, PA 16802 USA Daniel J. Weiss (djw21@psu.edu) Center for Language Science Department of Psychology and Program in Linguistics, Penn State University University Park, PA 16802 USA Abstract Studies of statistical learning have documented a remarkable sensitivity to structural regularities in both infants and adults. However, most studies of statistical learning have assumed a single underlying causal structure with uniform variance. In previous work in which two structures are presented successively, a primacy effect has been reported in which only the first structure is acquired. The present study explores the conditions under which such primacy effects are observed and learners are capable of acquiring both structures. We argue that learners can detect multiple structures by monitoring the consistency of the input. Keywords: speech segmentation, statistical learning Introduction Over the past twenty years, research on language acquisition has been transformed by the finding that infant and adult learners can use rudimentary statistics to parse artificial speech streams (Saffran, Aslin, & Newport, 1996; Saffran, Newport, & Aslin, 1996). A large number of follow-up studies have replicated and extended the initial findings, determining that statistical learning is neither domain specific (e.g., Fiser & Aslin, 2002a; Kirkham et al. 2002), nor even restricted to humans (Hauser et al., 2001; Toro & Trobalon, 2005). The term statistical learning has consequently come to be associated with a wide range of phenomena that rely on implicit calculations based on distributional regularities in the environmental input. The utility of these statistical learning experiments for simulating the early stages of language acquisition has been widely acknowledged. However, with few exceptions, the input to learners in statistical learning experiments has been characterized by a single, highly invariant statistical structure. This uniform-variance property of the input does not reflect the substantial variability inherent in natural language corpora due to shifts in topic, speaker, accent, and even language (in the case of bilingual acquisition). In some instances, variance in the input may signal to the learner that they are in a new context for which a different statistical structure must be learned (e.g., a language change), but in other cases this variation represents noise and should not trigger a new structural representation (e.g., hearing foreign- accented speech). Thus, the critical challenge confronting language learners is much like Piaget’s description of the processes of assimilation and accommodation (Piaget, 1985). The learner must ultimately determine the number of causal models that best characterizes the input, resolving when a new causal model is required and when the existing model can account for the observed data. There are at least two potential sources of information that may facilitate learners to detect that there has been a change in structure over time, which in turn may facilitate the formation of multiple representations (Gebhart, Aslin, & Newport, 2009). The first source of information is the availability of a contextual cue that is correlated with a particular statistical structure (e.g., Weiss, Gerfen, & Mitchel, 2009; Gebhart, Aslin, & Newport, 2009). The existence of such a cue could result in computations that are performed over a subset of the input and then compared across contexts. If the computations differ by some criterion, it would trigger the learner to form multiple representations to accommodate the inputs associated with each context. A second potential source of information for learners may be derived from monitoring the consistency of the input (Basseville & Nikiforov, 1993; see Gebhart, Aslin, & Newport, 2009). If the surface statistics are entirely consistent, the learner may conclude that the input likely has arisen from a single underlying structure. Conversely, if the variance in the surface statistics exceeds some criterion, then the learner may conclude that the underlying structure has undergone some change (see Gebhart, Aslin, & Newport 2009; Qian, Jaeger, & Aslin, 2012). To date, only a few experiments have tested whether contextual cues facilitate the formation of multiple representations when multiple inputs are presented. In a study by Weiss, Gerfen, & Mitchel (2009), learners were presented with two artificial languages comprised of four words each, in which the words were defined solely by transitional probabilities. The languages were interleaved in two-minute intervals twelve times total. When the languages were presented in a single voice, only congruent language pairs were learned (ones whose statistics, when combined, yielded similar transitional probabilities to the languages