When phonological learning is not statistical: how learning biases have reshaped Malagasy paradigms

One view of phonological learning is that it is driven by a domain-general bias towards frequency-matching, and therefore predictable from statistical distributions of the language (Albright 2002; Ernestus & Baayen 2003; Nosofsky 2011). A competing view is that learning is influenced by various linguistically-motivated biases, such as a bias towards less complex patterns. One way to tease apart these effects is to look at how learners behave when presented with conflicting data patterns––they do not always replicate the language they hear from their parents, and their errors (such as goed instead of went) can be informative, telling us about what biases are present in learning. Such errors can be adopted into speech communities and passed through generations of speakers, resulting in reanalysis of a data pattern.

In this talk, I look at reanalyses over time in Malagasy and show that reanalysis is not entirely predictable from statistical distributions. Instead, I argue that Malagasy reanalysis is sensitive to both frequency-matching and a markedness bias, using an iterated version of Maximum Entropy Harmonic Grammar (Smolensky 1986; Goldwater & Johnson, 2003) that simulates the cumulative effect of reanalyses across generations of speakers. I further propose that the markedness effects active in reanalysis are restricted to those already present in the language as a phonotactic tendency, and show that this is the case for Malagasy. These results demonstrate how findings from language change over time, combined with quantitative modeling, can give us insight into phonological learning.