Modelling unnatural classes of harmonic vowels in substance-free phonology

Vowel harmony (VH) refers to the crosslinguistically widespread process by which multiple vowels in a domain share the same value for a certain feature or set of features (van der Hulst, 2016). However, it is rarely the case that harmony applies exceptionlessly in all contexts. Languages with vowel harmony often contain neutral vowels, which fail to participate in harmony in some way or other, e.g. being transparent or antiharmonic (van der Hulst, 2016). This dissertation identifies a problem in formulating the structural description of harmony rules in certain languages: although the set of neutral vowels can be described as a featurally natural class, the set of harmonic vowels cannot; there is no way of identifying the set of harmonic vowels without also including non-harmonic vowels. The primary goal of this dissertation is thus to provide an account of this problem: how can we formulate rules that target unnatural classes of features? I assume a derivational model of phonology, in particular the Search-and-Copy theory (SCT) (Samuels, 2009); this approach is also substance-free, in that ‘markedness’ and other phonetically-based notions have no intra-grammatical status (Hale and Reiss, 2008). SCT models exceptions to harmony by allowing conditions on phonological operations. For example, Search may look only for [-high] vowels; as a result, high vowels are ignored by the Search, rendering them transparent in harmony. Neutral vowels are thus treated as exceptions by individual rules, rather than being exceptional throughout the grammar. However, if the set of harmonic vowels constitutes an unnatural class, it is unclear how this can be modelled in SCT. Current rule-based accounts to the problem are also evaluated. Rule ordering has been cited as an explanation for unnatural-class behaviour in vowel harmony (Mailhot and Reiss, 2007; Leduc et al., 2020). Similarly, explanations based on undeRspecification (Dresher, 2009) and visibility (Nevins, 2010) are considered. I argue that these accounts are empirically insufficient, in addition to raising conceptual issues.

I propose that the simplest possible account of the issue is to allow the phonological rule component to target unnatural classes of segments. I suggest two possible ways of implementing this. One is to introduce union operators into the phonological component, allowing rules to target featurally unnatural classes of segments through set union: A ∪ B, even if A ∩ B = {Ø} (i.e. A and B do not form a natural class). The alternative is to allow simultaneous (in addition to ordered) rule application; together with the assumption that Search is distance-sensitive (Nevins, 2010), this proposal predicts that rules can consist of multiple competing Search operations, allowing us to derive unnatural-class behaviour. The typological and computational consequences of both proposals are also considered. Allowing phonological rules to target unnatural classes extends the power of the phonological component by a non-trivial amount. It is thus worth asking why most rules can be expressed in terms of natural classes. I suggest that learning biases may account for the relative scarcity of featurally unnatural processes, even though they are computationally possible from the perspective of the phonology. For example, economy or simplicity principles in grammar construction (i.e. L1 acquisition) may render unnatural rules less preferable.