[lingtalks] Reminder: Colloquium Today

[Katie McGee]

Today's colloquium is presented by

Bruce Hayes, UCLA

"A Model of Phonotactics and Phonotactic Learning"

Time: 2-3:30
Place: 2148 McGill (TV Studio)
Date: May 1

Abstract:

Phonotactics is the study of well-formedness in phonological
representations.  Although this topic has always been important to
phonologists, no complete phonotactic analysis of a language has ever been
produced.  This is due to (a) the size of the problem, and (b) difficulty
in modeling the gradience always observed in phonotactic well-formedness
intuitions.

Our approach is based on several recent theoretical ideas.

•	Following Boersma, Goldwater and Johnson, and others, we assume that
phonological constraints should be assigned numerical strengths rather
than ordinally ranked; the result is a grammar that can model gradient
intuitions.
•	Following Frisch, Pierrehumbert, and others, we assume that much of the
gradience of intuitions matches the quantitative pattern of the lexicon: 
that which is underrepresented is felt by native speakers to be less well
formed.
•	To deal with the large size of the problem, we abandon traditional
yellow-pad analysis and instead use a learning algorithm.   Given a
training set of well-formed words, our algorithm selects from the
constraints made available under a particular version of phonological
theory, assigns weights to the selected constraints, and uses them to make
predictions about the well-formedness of novel words (e.g.
[blk] vs. *[bnk]).

We have applied our model several phonotactic problems:  English syllable
onsets, Shona vowel harmony, quantity-insensitive stress systems (Gordon
2002), and the complete phonotactics of Wargamay (Dixon 1981).   Our
modeling can be taken as support for various proposals from the
phonological literature concerning representations (autosegmental,
metrical, etc.):  we often find that a phonotactic pattern is learnable by
our system only when the training data are represented using an
appropriate theory.