Caleb Belth
Department of Linguistics
University of Utah

Phonological Grammar as Interplay Between Learner and Input: An Algorithmic Account of Developmental Differences among West Germanic Noun Voicing Alternations

The relationship between phonological theory and acquisition is usually viewed as one in which phonological theory delineates the learner’s hypothesis space. From this perspective, acquisition is about discovering a target grammar from that hypothesis space, and abstract structure like underlying representations is treated as structure hidden in the grammar. Consider a canonical example of a morphophonological process: voicing alternations arising from a phonotactic restriction against final voiced obstruents. The many cross-linguistic realizations of this phenomena are often swept into a single analysis in which underlyingly voiced forms are posited for alternating obstruents, which are devoiced in final position. Since various cross-linguistic instances are treated in the same theoretical terms, we might expect the learning trajectory and outcome to be comparable across instances. Yet a closer look reveals dramatic cross-linguistic differences in the environments that exhibit the alternation and in its developmental trajectories and outcomes, even in closely genetically related languages. Dutch-learning children appear to acquire the alternation in noun paradigms over a protracted period, with little evidence of productivity by 3.5yrs old or even 6-7. In contrast, German-learning children have more advanced knowledge of the alternation at 3.5yrs, and are extending it to nonce words by at least age 5, likely earlier. Several Yiddish and Bavarian dialects lost the voicing alternation, most likely through acquisition directly following the loss of final reduced -e. This talk argues for a conceptualization of the relationship between phonological theory and acquisition in which phonological representations and generalizations are dynamic products of an interplay between learner and input. This is made concrete for the case of West Germanic voicing alternations, by proposing that learners construct non-surface-faithful underlying forms in response to being unable to morphologically generalize without them. When morphological generalization over surface-faithful representations is sufficient, alternations can slip by as exceptions to the generalizations that do not account for them. This proposal is implemented as a computational learning model for West Germanic noun morphophonology, which provides a learning-based explanation for these developmental differences. Applied to child directed speech, the model correctly predicts diverging developmental trajectories for Dutch- and German-learning children. When applied to a corpus of Middle High German nouns, the model leads to a substantial drop in the productivity of the voicing alternation following simulated -e apocope, consistent with a leading theory for its loss in some Yiddish and Bavarian dialects.