A Learning Based Account of Turkish Laryngeal Alternations

A common conception of the relationship between phonological theory and learning is that theory delineates a space of possible grammars and learning involves induction from overt linguistic data to a target grammar in that space. From this perspective, phonologists have taken interest when learners fail to form a generalization that appears tenable given analysis of language data, because such cases may indicate something about (im)possible generalizations. A prominent example is Turkish voicing alternations, where some but not all noun-stem-final stops alternate in voicing based on whether they are in final position, and devoiced, or followed by a vowel. While alternation or non-alternation it is a property of particular nouns, a number of features are predictive of whether a noun is likely to be in the alternating or non-alternating group. One such feature—the quality of the vowel preceding the final stop—shows no evidence of being internalized by Turkish speakers. One suggestion has been that UG analytically filters out this dependency. In this paper, I argue for a learning-based explanation in which the lack of sensitivity is a consequence of the algorithms by which generalizations are constructed, rather than a reflection of an analytical bias in UG.