Selectional restrictions as phonotactics over sublexicons

Abstract

Affixation and allomorphy are often phonologically predictable: thus, the English indefinite "a" appears before consonants, and "an" before vowels. We propose a theory of phonological selection that separates rules of morpheme realization from phonological knowledge about the bases and the derived words. This phonological knowledge is encoded in miniature phonotactic grammars, which are learned over sublexicons defined by morphological generalizations. Each sublexical phonotactic grammar determines the likelihood that a new word will follow the associated rule. We examine a complex case of suppletive allomorphy in Russian, whose diminutive suffixes define sublexicons differing in constraints on final consonant place and manner, presence and location of consonant clusters, vowel hiatus, and stress. In elicitation, Russians choose allomorphs for words without diminutives based on how these words and the derived diminutives fare in the sublexical phonotactic grammars. In a nonce word study, Russians also chose allomorphs based on sublexical phonotactic well-formedness, even when the phonotactic violations were non-local to the affix itself. These patterns are missed by alternative approaches such as emergence of the unmarked, insertion rules that refer directly to phonological information, and the Minimal Generalization Learner.

Beyond Plain and Extra-Grammatical Morphology: Echo-Pairs in Hungarian

“…We have partly circumvented this issue by reordering the sounds in the input to the MGL, so that both O 1 and O 2 are available for feature-based generalisations, but it appears that this does not fully resolve the issue. Gouskova et al (2015) point out similar issues with the MGL in their paper on Russian diminutive formation, a phenomenon that closely resembles Hungarian echopairs in terms of its (often non-local) phonological conditioning. Therefore, the failure of this approach may stem from limitations that are specific to the MGL, and it is possible that a different model (e.g.…”

“…In such cases, speakers decide between different allomorphs for a given root based on the likelihood of that root under the constraint rankings for each allomorph-specific sub-lexicon. Gouskova et al (2015) show that this model can account for fine-grained and probabilistic patterns of phonological conditioning in Russian speakers' choices of diminutive allomorphs -a phenomenon that is, in many ways, similar to Hungarian echo-pair formation.…”

“…These rules differ from those of traditional rule-based accounts not only in terms of their potential specificity, but also in the fact that they are allowed to apply stochastically. Gouskova et al (2015) present a model based on a constraint-based maximum entropy grammar (Hayes and Wilson, 2008). They assume that (i) roots are indexed in memory for the allomorphs that they typically combine with (e.g.…”

“…We also do not mean to imply that these are the only existing models that could be tested. The models that we focus on suffice for our purposes in that they represent a broad range of variation in terms of lexical specificity -but there are other alternatives such as Gouskova et al's (2015) maxent-based model that are not examined here.…”

“…In fact, many current models of morphology take such patterns of variation as their key explanandum and draw on a rich lexicon coupled with analogical processes or intricate sets of highly specific rules and constraints to account for speakers' variable morphological intuitions (e.g. Albright and Hayes 2003;Dawdy-Hesterberg and Pierrehumbert 2014;Gouskova et al 2015). Second, not all extra-grammatical word-formation processes are created equal.…”