Phonological naturalness and phonotactic learning
by Bruce Hayes and James White, UCLA
Published in Linguistic Inquiry 44:45-75 (2013)
Abstract
-
We investigate whether the patterns of phonotactic well-formedness
internalized by language learners are direct reflections of the
phonological patterns they encounter, or reflect in addition principles
of phonological naturalness. As a research tool we employ the
phonotactic learning system of Hayes and Wilson (2008), which carries
out an unbiased search of the lexicon for valid phonotactic
generalizations. Applying this system to English data, we find that it
learns many constraints that seem to be unnatural—they have no evident
typological or phonetic basis, yet hold true of the English lexicon. We
tested the status of ten of these constraints in a nonce-probe study,
obtaining native-speaker ratings of novel words that violated them. We
used 40 such words: 10 violating our unnatural constraints, 10 violating
natural constraints assigned comparable weights by the Hayes/Wilson
learner, and 20 violation-free forms, each similar to a test form and
employed as a control. In our experiment, we found that violations of
the natural constraints had a powerful effect on native speaker judgment
and violations of the unnatural constraints had at best a weak one. We
conclude by assessing a variety of hypotheses intended to explain this
disparity, opting ultimately for a learning bias account.
Would you like to replicate?
- It should be possible just unzip this
file, put the folder on a Windows computer, and click on
Experiment.exe. Data will be stored in Results.txt.
Feature set
- Features, used for constructing
all constraints
Grammars
- Grammar 1, 160 constraints, used for
constructing the experimental stimuli
- Grammar 2, 110 natural and 10 unnatural
constraints, addressing the question, "could the unnatural constraints
have been excluded on statistical grounds?" (section xxx)
Sound files
"Pure" model
If, in setting up our grammar, we had used no a priori natural constraints
(section 4), would the result have been different? Probably not;
see text fn. 7.
- "Pure"
grammar, 160 constraints, all selected by the Hayes/Wilson model
- Predictions
of the pure grammar for our experimental stimuli