ACL 2010 SIGMORPHON 2010 Eleventh Meeting of the ACL Special Interest Group on Computational Morphology and Phonology

I present LIBPHON, a nonparametric regression-based model of phonological acquisition that induces a generalised and productive pattern of vowel harmony—including opaque and transparent neutrality—on the basis of simplified formant data. The model quickly learns to generate harmonically correct morphologically complex forms to which it has not been exposed. 1 Explaining phonological patterns How do infants learn the phonetic categories and phonotactic patterns of their native languages? How strong are the biases that learners bring to the task of phonological acquistion? Phonologists from the rationalist tradition that dominated the past half-century of linguistic research typically posit strong biases in acquisition, with language learners using innatelygiven, domain-specific representations (Chomsky and Halle, 1968), constraints (Prince and Smolensky, 2004) and learning algorithms (Tesar and Smolensky, 2000; Dresher, 1999) to learn abstract rules or constraint rankings from which they can classify or produce novel instances. In the last decade, however, there has been a shift toward empiricist approaches to phonological acquisition, use and knowledge. In this literature, eager learning algorithms (Aha, 1997), in which training data are used to update intensional representations of functions or categories then discarded, have been the norm.1 However, research in related fields—particularly speech perception— indicates that speakers’ knowledge and use of language, both in production and comprehension, is at least partly episodic, or instance-based (Goldinger, 1996; Johnson, 1997). Additionally, Daelemans et al. (1994) is a notable exception. motivation for instance-based models of categorisation has a lengthy history in cognitive psychology (Medin and Schaffer, 1978), and these methods are well-known in the statistical and machine learning literature, having been studied for over half a century (Fix and Hodges, 1951; Cover and Hart, 1967; Hastie et al., 2009). Consequently, it seems a worthy endeavour applying an instancebased method to a problem that is of interest to traditional phonologists, the acquisition and use of vowel harmony, while simultaneously effecting a rapprochement with adjacent disicplines in the cognitive sciences. In sections 2 and 3 I give some brief background on vowel harmony and instancebased models, respectively. Section 4 introduces my model, LIBPHON, and section 5 the languages it learns. I discuss some simulations and results in section 6, and conclude in section 7.