Abstract:
One of the central problems in language learning is ambiguity in the overt
data available to the learner. An overt form (the audible portion of an
utterance) may be ambiguous between more than one full linguistic analysis,
and different languages may choose different analyses. In metrical stress,
a three syllable word with main stress on the middle syllable is ambiguous
between an iambic analysis (in which the first two syllables are grouped
together into a foot) and a trochaic analysis (in which the final two
syllables are grouped together into a foot). Because linguistic principles
evaluate entire linguistic analyses (and not just overt forms), the learner
must overcome this ambiguity in order to use the overt forms to learn the
grammar of their native language.
This talk will present results from a learning algorithm capable of learning
from ambiguous data. This research uses the linguistic framework of
Optimality Theory, so that the primary goal of the learner is to learn the
language-specific ranking of the universal constraints. The learner
overcomes ambiguity by temporarily considering more than one analysis of the
ambiguous form, and then eliminating those analyses which are inconsistent
with other data from the language. The simulation results demonstrate that
this strategy can succeed without explosive growth in the number of
grammatical hypotheses the learner has to maintain during the course of
learning.
Back to Talks Page