UCLA

This dissertation examines the question of how phonological alternations are learnt. In constraint-based models of phonological learning, it is hypothesized that prior learning of phonotactics from the lexicon facilitates the learning of alternations. While this is an influential assumption, the empirical evidence for it is equivocal. In this dissertation, I investigate this link by examining the learning outcomes in cases where phonotactics and alternations mismatch, particularly in cases of derived environment effects. For example, in Korean, /t/ palatalizes to [c] before [i] across a morpheme boundary, yet [ti] sequences are attested within stems. Derived-environment effects have proven theoretically challenging to account for precisely because of the mismatch in generalizations within stems and across morpheme boundaries.

Using an artificial grammar learning paradigm, I first show that alternation learning is facilitated when the phonotactics in the lexicon match the alternation. Participants who were trained on a language with vowel harmony within stems and across morpheme boundaries

successfully learnt the vowel harmony alternation. Conversely, those who were trained on a language without vowel harmony in stems (derived-environment effect language) failed to learn the alternation across the morpheme boundary, despite being trained on this pattern. This supports the hypothesis of current models of phonological learning. Yet, when participants were only trained on a static phonotactic generalization, they did not readily extend this to unseen alternations. This suggests that learners are conservative, and need experience with alternations before extending a learnt phonotactic generalization.

What does this mean for phonological patterns with a mismatch between phonotactics and alternations? I present corpus analyses as well as computational learning simulations of two cases which show these derived-environment patterns - Korean palatalization and Turkish velar deletion. I show that in both cases the reported mismatches between phonotactics and alternations are superficial. Korean palatalization is an active alternation supported by a gradient phonotactic constraint in the lexicon. Turkish velar deletion, however, is a morphologically circumscribed alternation with no accompanying phonotactic generalization in the lexicon. I also briefly discuss how this statistical pattern in Turkish is similarly the case with the well-known Finnish assibilation pattern. This undermines assumptions in previous analyses of these patterns. Instead, I argue that there is a bias to maintain similar generalizations, captured by more general constraints, in phonotactics as well as alternations. This further supports the claim that learning phonotactics aids in learning alternations, and further suggests that derived-environment patterns are typologically dispreferred.

I conclude by exploring how a bias for more general constraints might be implemented using a Maximum Entropy learner, making specific use of the model’s prior. Specifically, I propose a bias for the learner to prefer generalizations to be explained by more general constraints which are blind to morphological structure, over constraints which reference morphological structure. I show that, under such a model, a canonical derived-environment pattern with a mismatch in phonotactics and alternations is never accurately learnt.