Colloquium with Bruce Hayes, UCLA

Title:  A wug-test of Catalan consonant alternations (joint work with Kevin Liang and Victoria Mateu)

Abstract:  We describe a “wug-test” experiment covering four phonological processes affecting consonants in Catalan (Romance, Catalonia):
/n/ Deletion n → ∅ / ___ ]word /san/ → [sa] ‘healthy-m.’ (f. [san-ə])
/r/ Deletion r → ∅ / ___ (s)]word /klar/ → [kla] ‘clear-m.’ (f. [klar-ə])
Cluster Simplification t → ∅ / n ___ ]word /sant/ → [san] ‘holy-m.’ (f. fem. [sant-ə])
/ʒ/ Affrication ʒ → tʃ / ___ ]word /bɔʒ/ → [bɔt͡ʃ] ‘crazy-m.’ (f. fem. [bɔʒ-ə])
For the test, we constructed nonce words that could be used to assess the productivity of these processes. Participants were given a wug feminine word, and asked to supply the corresponding masculine.
We chose these alternations with the goal of testing hypotheses in phonological theory. On the basis of our results we argue, first that speakers are capable of frequency-matching phonological patterns in the lexicon. In particular, the rates of application of /n/ Deletion and /r/ Deletion across phonological subcontexts (such as monosyllables, penult-stressed stems) were matched by our participants. Where frequency-matching was imperfect, we offer hypotheses for why this was so: (a) a bias for generality in forming the grammar; (b) the effect of orthography. Cluster Simplification and /ʒ/ Affrication were included to test the productivity, respectively, of opaque phonology and of “saltatory” alternations (/ʒ/ → [tʒ], skipping over [ʃ]). Both proved to be productive, but quite imperfectly so, with only loose matching of the lexical pattern.
The participant responses included one further type that surprised us: avoidant responses, in which participants employed two rare options of Catalan morphology — [-u] and [-ə] suffixes for masculines — in order to avoid committing themselves regarding the phonological question at hand. An example is masculine [mirbunt-u] in response to feminine [mirbunt-ə], which avoids the need to choose whether to apply /n/ Deletion ([mirbun] vs. [mirbunt]).
We are curious how to analyze avoidance. We first tried simply inserting candidates like [mirbunt-u] into our (otherwise quite accurate) MaxEnt OT tableaux, and found that the model fit was poor. The reason, we suggest, is that avoidant responses are not really candidates in the phonological competition, but rather represent a task-driven response to uncertainty. This uncertainty arises because of a sparsity of relevant examples in the lexicon. For instance, participants were highly avoidant for /ʒ/ stems, which represent a pattern numbering only about 15 in the whole language. If we are correct, participants have rational intuitions about how uncertain they should be. However, to our knowledge, no current grammatical theory or computational learning theory makes predictions about uncertainty.