UMass Amherst
UMass Amherst | Library | Umail | Spire | People Finder 

Search

Match case Regex

Current issue

Categories

WHISC Back Issues

Other Ling Newsletters

Subscribe to this blog's feed
[What is this?]
Creative Commons License
This weblog is licensed under a Creative Commons License.

« Michael Becker Accepts Position at Reed College | Main | Maria Polinsky Colloquium»

Kie Zuraw Lecture

Kie Zuraw
UCLA

Natural and unnatural generalizations: early results from a Hungarian wug test

(Joint work with Bruce Hayes, Zsuzsa Londe, and Péter Siptár.)

Friday, May 2, 1:30 pm, in the Partee Room

Abstract

Hayes & Londe (2006) argue that Hungarian speakers have implicit knowledge of certain statistical patterns in vowel harmony. For example, stems with a back vowel and a neutral vowel (e.g., farmer 'blue jeans') may take either the back allomorph -nak or the front allomorph -nek of the dative suffix, but higher neutral vowels behave as more transparent, allowing -nak more often. Hayes & Londe found evidence for these patterns in the rates at which real words take each allomorph in a written corpus (the Web) and in subjects' choices in a wug test (Berko 1958).

The patterns studied by Hayes & Londe were phonologically natural ones (with one possible exception), well known in the literature. It was therefore unknown whether the patterns seen reflect emergent universal tendencies or rather lexical learning. And, if the patterns reflect lexical learning, are they learnable only because of their phonological naturalness? Would unnatural patterns that happen to be present in the data be equally learnable?

To address the question of whether natural and unnatural patterns are equally learnable, we designed a larger-scale wug test. We first identified four unnatural generalizations--albeit stated over phonological categories--in the real-word data. For example, words ending in a bilabial consonant tend to take -nek. Participants, recruited and surveyed over the web, were presented with novel items balanced for the four unnatural generalizations (to the extent that the generalizations are orthogonal) and otherwise reflective of the statistical phonotactics of Hungarian nouns. To avoid item-specific effects, a fresh set of stimuli was generated for every subject. For each item, subjects chose the -nek or -nak form, and also rated each choice.

Our results show clear evidence for two of the four unnatural constraints (and five of the six Hayes/Londe constraints; the sixth-- the possibly-unnatural constraint identified above--is applicable to few test items). Thus--assuming we are right to classify these constraints as unnatural--we reject the hypothesis that only natural constraints can be learned and used.

More interestingly, our preliminary analysis also argues against the other extreme hypothesis, which is that any constraint stated over phonological categories is equally learnable or usable. When fitting constraint weights not to the experimental data but to the real-word data, we obtain the best match to the experimental data not when all constraints are treated equally, nor when the unnatural constraints are excluded, but rather somewhere in between: unnatural constraints are available, but weighting them strongly is somewhat penalized and their weights are thus lower than they would otherwise be. This echoes the "substantively biased" model proposed in Wilson 2006, where a constraint's phonetic consequences determine learners' a- priori willingness to weight it strongly. In Wilson's case, markedness constraints are biased to remain low-weighted to the extent that they motivate perceptually small changes. In the absence of the detailed phonetic information needed to calculate such detailed bias factors, we merely make a binary distinction between natural and unnatural markedness constraints.

Our results thus seem to be a graded version of Becker, Ketrez & Nevins's (2007), in which Turkish speakers extended natural but not unnatural regularities to new items. If at least some unnatural patterns are learnable but penalized (see Pertsova 2004 for another example), studies of differential learnability for real and artificial languages could reach conflicting results depending on whether the strength of the pattern and the difficulty of the task cause the unnatural pattern to fall below an observable-learnability threshold or to hit a ceiling.

May 1, 2008 12:09 AM | 06:15 (2008-05-01) , colloquia