A Gradient Harmonic Grammar Account of Nasals in Extended Phonological Words

The article aims at contributing to the long-standing research on the prosodic organization of linguistic elements and the criteria used for identifying prosodic structures. Our focus is on final coronal nasals in function words in Greek and the variability in their patterns of realization before lexical words. Certain nasals coalesce before stops and delete before fricatives, whereas others do not. We propose that this split in the behavior of nasals does not pertain to item-specific prosody because the relevant strings are uniformly prosodified into an extended phonological word (Itô & Mester 2007, 2009). It rather stems from the contrastive activity level of nasals in underlying forms in the spirit of Smolensky & Goldrick’s (2016) Gradient Symbolic Representations; nasals with lower activity coalesce and delete in the respective phonological environments, whereas those with higher activity do not. We show that the proposed analysis captures certain gradient effects that alternative analyses cannot account for.

Comparing MaxEnt and Noisy Harmonic Grammar

MaxEnt grammar is a probabilistic version of Harmonic Grammar in which the harmony scores of candidates are mapped onto probabilities. It has become the tool of choice for analyzing phonological phenomena involving probabilistic variation or gradient acceptability, but there is a competing proposal for making Harmonic Grammar probabilistic, Noisy Harmonic Grammar, in which variation is derived by adding random ‘noise’ to constraint weights. In this paper these grammar frameworks, and variants of them, are analyzed by reformulating them all in a format where noise is added to candidate harmonies, and the differences between frameworks lie in the distribution of this noise. This analysis reveals a basic difference between the models: in MaxEnt the relative probabilities of two candidates depend only on the difference in their harmony scores, whereas in Noisy Harmonic Grammar it also depends on the differences in the constraint violations incurred by the two candidates. This difference leads to testable predictions which are evaluated against data on variable realization of schwa in French (Smith & Pater 2020). The results support MaxEnt over Noisy Harmonic Grammar.

Persistence and Opacity in Eastern Andalusian Harmony

This paper proposes a novel account of a derivationally opaque aspect of ATR harmony in Eastern Andalusian. Harmony in the language is driven by Positional Licensing: [-ATR] originating on final vowels must spread to the stressed vowel. Intervening post-tonic vowels optionally also harmonize, as do pretonic vowels. Typically in licensing-driven systems, if harmony is unable to reach the licensor, harmony does not affect non-licensing positions either. Not so in Eastern Andalusian: high vowels do not harmonize, but a stressed high vowel does not prevent unstressed vowels from harmonizing as normal – harmony can overapply on these vowels. The analysis, couched in serial Harmonic Grammar, develops a new mechanism called persistence that accounts for this opacity. Under persistence, once a feature satisfies Positional Licensing by spreading to the licensing position, Positional Licensing remains satisfied for the rest of the derivation, even if the feature vacates the licensing position. This allows a stressed high vowel to harmonize, thereby permitting unstressed vowels to harmonize, too, and then harmony can retract off the high vowel without running afoul of Positional Licensing.

Some Formal Implications of Deletion Saltation

In a phonological saltation alternation, a segment or class “skips” a relatively similar category to surface as something less similar, as when /ɡ/ alternates with [x], skipping [k]. White (2013) and Hayes and White (2015) argue that saltation is unnatural—difficult to learn in the laboratory and diachronically unstable. They propose that the phonological grammar includes a learning bias against such unnatural patterns. White and Hayes further demonstrate that Harmonic Grammar (HG; Legendre, Miyata, and Smolensky 1990) cannot model typical saltation without nondefault mechanisms that would require extra steps in acquisition, making HG consistent with their proposed learning bias. I identify deletion saltation as a distinct saltation subtype and show that HG, with faithfulness formalized in standard Correspondence Theory (CT; McCarthy and Prince 1995), can model this pattern. HG/CT thus predicts that deletion saltation, unlike typical (here called segment-scale) saltation, is natural. Other frameworks fail to distinguish the two saltation types—they can either model both types, or neither. Consequently, if future empirical work finds deletion saltation to be more natural than other saltation patterns, this would support weighted-constraint models such as HG over ranked-constraint models such as Optimality Theory (OT; Prince and Smolensky 1993, 2004); would support CT over the *MAP model of faithfulness (Zuraw 2013); and would support formalizing CT featural-faithfulness constraints in terms of IDENT constraints, binary features, or both.

Modeling gradient processes in Polabian vowel chain shifting and blocking

This paper presents an analysis of two interacting sound changes in the extinct West Slavic language Polabian. Polabian is known to have two types of vowel innovations: (i) the incorporation of acoustic properties from consonant secondary co-articulations (either palatalization or velarization) and (ii) a systematic rotation of vowels (Timberlake 1995). This paper argues that the innovation in (ii) is a vowel chain shift similar to those analyzed in Labov (1994). Unlike the other languages surveyed in Labov (1994), Polabian has phonologically predictable exceptions to the general direction of vowel movement through the acoustic space. Unlike previous work on Polabian, this paper proposes that the vowel chain shift operated simultaneously with the innovation in (i) resulting in phonologically predictable exceptions. This paper tests Timberlake’s (1995) proposal and the current proposal in a Harmonic Grammar (Flemming 2001) which uses Purcell’s (1979) acoustic data from Russian as a proxy. The model only captures the correct distribution of vowel reflexes under the assumption that co-articulatory based innovations and vowel chain shifting were active at the same time.

Constraint summation in phonological theory

The classical constraints used in phonological theory apply to a single candidate at a time. Yet, some proposals in the phonological literature have enriched the classical constraint toolkit with constraints that instead apply to multiple candidates simultaneously. For instance, Dispersion Theory (Flemming 2002, 2004, 2008) adopts distinctiveness constraints that penalize pairs of surface forms which are not sufficiently dispersed. Also, some approaches to paradigm uniformity effects (Kenstowicz 1997; McCarthy 2005) adopt Optimal Paradigm faithfulness constraints that penalize pairs of stems in a paradigm which are not sufficiently similar. As a consequence, these approaches need to “lift” the classical constraints from a single candidate to multiple candidates by summing constraint violations across multiple candidates.Is this assumption of constraint summation typologically innocuous? Or do the classical constraints make different typological predictions when they are summed, independently of the presence of distinctiveness or optimal paradigm faithfulness constraints? The answer depends on the underlying model of constraint optimization, namely on how the profiles of constraint violations are ordered to determine the smallest one. Extending an independent result by Prince (2015), this paper characterizes those orderings for which the assumption of constraint summation is typologically innocuous. As a corollary, the typological innocuousness of constraint summation is established within both Optimality Theory and Harmonic Grammar.

Cumulative constraint interaction and the equalizer of OT and HG

We show that, in general, Optimality Theory (OT) grammars containing a restricted family of locally-conjoined constraints (Smolensky 2006) make the same typological predictions as corresponding Harmonic Grammar (HG) grammars. We provide an intuition for the generalization using a simple constrast and neutralization typology, as well as a formal proof. This demonstration adds structure to claims about the (non)equivalence of HG and OT with local conjunction (Legendre et al. 2006, Pater 2016) and provides a tool for understanding how different sets of constraints lead to the same typological predictions in HG and OT.

Gradient categories in lexically-conditioned phonology: An example from sound symbolism

There are many approaches to modeling lexically-conditioned phonology in current formal theories, including lexically-indexed constraints and cophonologies. Nearly all of these existing approaches assume categorical membership in the lexical classes that condition differential phonotactics or phonological behaviours: for example, a lexical item is either a noun or a verb, or of one gender class or another. This paper proposes an implementation of Maximum Entropy Harmonic Grammar with lexically-indexed constraints and gradient symbolic activations over classes that allows us to model differences in phonological patterns over both discrete and gradient class membership. This theoretical implementation is a natural extension of the scales and gradient activations that have been shown to be necessary in recent phonological theory: sound symbolic evidence highlights the necessity for such increased explanatory power in our phonological models. Crucially, we find gradient lexically-conditioned patterns not only in sound symbolism—where they are often most obvious—but also in what is considered “core” language (e.g., morphosyntactic classes), and allowing gradient class structures in our phonological models may ultimately make for cleaner interfaces with other parts of grammar such as morphosyntax.

Overshoot in licensing-driven harmony

Kaplan (2018a) argues for a positive and gradient version of positional licensing in Harmonic Grammar. A chief difference between this formalism and standard positional licensing is that it predicts that harmony whose goal is to place a feature in a licensing position may overshoot its target by extending beyond the licensing position. Centralisation harmony in Tudanca Montañés bears out this prediction: though harmony triggered by a final vowel typically stops at the stressed syllable, under particular circumstances it extends into the pretonic domain. Positive gradient positional licensing is indispensable in an account of this. It plays a central role in a gang effect that drives overshoot, an interaction that cannot be replicated with standard versions of positional licensing.