Many recent studies have attempted to model the acceptability of unattested consonant clusters based on their similarity to attested clusters (Hayes & Wilson 2008; Albright 2009; Davidson & Wilson 2009).  Consonant clusters are a rich testing ground for inductive models, because they provide many clear instances of 'overlearning': that is, cases in which native speakers show a preference for one cluster over another, in spite of the fact that the two evidently have equal frequencies.

In this talk, I present new experimental data concerning two cases of overlearning.  The first involves a contrast between word-final nasal-stop vs. fricative-stop clusters.  With voiceless stops, -NC clusters greatly outnumber -SC clusters: e.g., 80 -mp, 99 -ŋk vs. 18 -sp, 29 -sk monosyllabic lemmas in CELEX.  With voiced stops, on the other hand, -mb, -ŋk, -zb, -zg are all unattested.  Surprisingly, English speakers strongly prefer nonce words ending in voiced fricative-stop clusters over their nasal-stop counterparts: e.g., fræzb rated 5.5 (out of 7) vs. fræmb 2.1.  The second involves a contrast between initial fC- vs. θC- clusters. Both fC- and θC- are unattested in English, with sC- being the only attested fricative-obstruent cluster type.  Here, English speakers show a small but systematic preferences for fC- clusters: e.g., fnɛdʒ 1.9, fnoʊ 2.3 vs. θnɛdʒ 1.3, θnuː 2.0.  In both cases, the dispreferred clusters are featurally closer to the better-attested clusters: -mb is featurally similar to well-attested -mp, -md, and θn- is featurally similar to sn-.

I discuss the performance of several inductive models on these comparisons.  The results show that both the Hayes and Wilson (2008) and Albright (2009) models fail to predict the systematic preferences that are observed in the human ratings.  I then compare a modified version of the Albright (2009) model, in which pairs of adjacent natural classes are rewarded or penalized based on the extent to which one class supports acoustic cues for the features of the other.  In the case of nasal+stop combinations, it is hypothesized that the lack of distinct cues to the presence of the homorganic stop strongly disfavors [+nasal][-cont,+voi] combinations. In the case of fC- vs. θC-, it is hypothesized that amplitude of frication noise and coarticulatory cues to labial place favor fC-.  With these biases in place, the model is able to achieve a closer fit to the observed human ratings.

Phonetically Guided Induction of Constraints on Consonant Clusters

Adam Albright

MIT