Seminar Speaker, Kevin McMullin, University of Ottawa
Linguistics Colloquium Series

Many languages exhibit patterns of consonant harmony that are seemingly enforced across any distance. In Samala, for example, a process of sibilant harmony is seen even in words such as /k-su-k’ili-mekeken-ʃ/ ➝ [k*ʃ*uk’ilimekeket*ʃ*] ‘he had a stroke of good luck’ (Applegate 1972). As it stands, we do not have an explanation as to where these unbounded processes come from. Although there is evidence that interaction of consonants over relatively short distances is functionally grounded in speech planning, this does not seem to extend to arbitrary distances. One possibility is that unbounded harmony emerges when a new generation of learners encodes the pattern in their grammar. However, synchronic patterns of harmony restricted to *C*V.*C* contexts are widely attested, and
artificial grammar learning studies show that learners do not readily misinterpret these as unbounded patterns. In this talk, I highlight two ways in which analytic bias may instead provide an *indirect* route to unbounded consonant harmony.

The first conjectural account relies on the inherent ambiguity of the *C*V. *C* window, which has been analyzed both as “syllable-adjacent” and “transvocalic”. In particular, it so happens that all languages with this variant of harmony lack the appropriate syllable structures or specific morphemes that would resolve the ambiguity. The idea, then, is that syllable-adjacent harmony would be enforced at an intermediate *C*V.CV <http://v.cv/>*C* distance, thereby making it more likely that learners (or at least a sufficient number of them) might infer an unbounded pattern. Secondly, I suggest that iterative application to multiple targets could result in overgeneralization to unbounded distances, and present preliminary results of an ongoing artificial grammar learning study as tentative support for this hypothesis. Specifically, learners exposed to regressive liquid harmony only in CVCV*L*V-*L*V items tend to apply alternations to both of the targets in CV*L*V*L*V-*L*V items. Furthermore, learners trained only on contexts with multiple targets (i.e., CV*L*V*L*V- *L*V) are more likely than the former group to enforce harmony across longer distances, even with no intervening liquid (i.e., in CV*L*VCV-*L*V contexts). This points toward a two-step process whereby unbounded harmony could emerge due to the “long-range” dependencies exhibited when the process is applied iteratively.