The need for abstraction in phonology: A commentary on Ambridge (2020)

In his article in this issue of First Language, Ambridge (2020) summarizes the evidence in favor of exemplar models of phonology. Exemplar models are well suited for handling gradient representations, including talker (e.g., Johnson, 1997), lexical (e.g., Bybee, 2002; Pierrehumbert, 2001), and sociolinguistic (e.g., Hay et al., 2006) variation. The vast majority of phonologists accept that phonological processes are grounded in phonetics (Archangeli & Pulleyblank, 1994), and that the phonetic variation that arises in everyday speech must be accounted for at some level (though see Hale and Reiss, 2000, for an alternative ‘substance free’ approach). However, Ambridge (2020) is misguided in the abrupt dismissal of generative models such as Optimality Theory (OT) (Prince & Smolensky, 2004), as it is not clear that exemplar models can (or should) supplant abstraction in phonology altogether.

Exemplar models are well suited to handle gradient representations in pronunciation, but phonology is more than pronunciation of words. Phonology is the interaction of algebraic rules and constraints on sound patterns that are bound by structural constraints, and can be generalized to novel contexts (Berent, 2013; Heinz, 2011). While gradient effects, such as talker specificity, have their place at the word level, it is not clear how well these constraints hold for phonological patterns (such as vowel harmony) (Finley, 2013). More likely, gradient effects discussed in exemplar models may sit at a different level of abstraction from the majority of phonological processes. Exemplar models of phonology have often advocated for various levels of abstraction, or a ‘ladder of abstraction’ (Beckman et al., 2007; Munson et al., 2011; Pierrehumbert, 2003). Learning phonological representations is not just about learning the phonetic details of a language, but how those phonetic details fit into larger representations. While models of phonology and language acquisition must account for the gradient, categorical phenomena are still present at increasingly greater levels of abstraction. These levels can include phonological context (e.g., stress, syllable structure), but could also include social context (e.g., sexual orientation, gender). In order to know to account for sociolinguistic variation in pronunciation, the listener must make an index of the social context, an abstract concept.

Allowing for multiple levels of abstraction in language has many benefits. First, it provides an avenue to account for cases where the statistical tendencies of the input do not match speaker representations (Becker et al., 2011). For example, there are no words in English that start with [bw], but speakers find this cluster to be more acceptable than *[dl]. This difference in acceptability cannot be explained in terms of frequency or similarity. Rather, structural constraints on phonological representations (sonority) predict [bw] to be more acceptable than *[dl] (Moreton, 2002). Other studies have demonstrated that speakers have knowledge of constraints on consonant clusters (the sonority sequencing principle) (Berent & Lennertz, 2010; Berent et al., 2009), even when the speakers of the language have no experience with clusters (Berent et al., 2008). Input statistics alone often make incorrect predictions about speaker representations, but when combined with abstract representations, make the appropriate predictions about learnability (Jarosz et al., 2017).

Second, allowing for multiple levels of abstraction allows for a bridge between the descriptive power of an exemplar model, and the explanatory power found in generative formalisms. For example, an exemplar model can adequately model the gradient phonetic effects in transparent vowels in Hungarian vowel harmony, but such a model fails to explain why the phonetics behave the way they do, in the environments that they do. It also fails to capture the typological tendencies across languages (for example, that high vowels are more likely to be transparent to back vowel harmony than mid vowels) (Benus, 2010). Further, artificial grammar learning experiments in adults have shown that learners may be sensitive to typological constraints on phonological processes (Finley, 2012; Finley & Badecker, 2008). Theories that use abstract representation (such as OT) allow for predictions about the typological tendencies across languages of the world, which can help provide explanatory power to models of phonological learning.

Third, allowing for a ‘ladder of abstraction’ is compatible with an integrative account of phonological knowledge (Smolensky & Legendre, 2006). While generative formalisms often focus on the categorical, it is not clear that OT cannot handle forms at the exemplar level. Just as a specific morpheme can trigger ‘morpheme-specific’ constraints (Pater, 2010), so too could a specific lexical item trigger a specific ranking or set of constraints related to a social index. Ambridge (2020) argues that generalization in language is about analogy among outputs, and OT can account for this type of comparison, in the form of output-output constraints (Benua, 2000; Burzio, 1998). Gradient versions of OT (Boersma & Hayes, 2001), including Harmonic Grammar (Legendre et al., 2006), have been implemented to account for a variety of gradient effects in phonology (Coetzee & Pater, 2011; Hayes & Londe, 2006; Hayes & Wilson, 2008). Thus, it is a shortsighted notion to outright dismiss generative models like OT because they are not designed to handle forms at the exemplar level. It is also not clear whether generative formalisms should account for representations at the individual exemplar level. Rather, it may be necessary to combine the insights of exemplar theory to account for gradient representations at the individual exemplar level at one level of abstraction, with the computational power of OT at higher levels of abstraction. An important question in phonology is how to integrate the gradient with the categorical, and whether a single formalism like OT (or exemplar theory) should handle everything, or whether different formalisms can work at various levels of abstraction. Optimality Theory, as part of an Integrative Connectionist/Symbolic architecture (Smolensky & Legendre, 2006), may provide a possible framework for integrating these elements.