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

Search

Match case Regex

Categories

Creative Commons License
This weblog is licensed under a Creative Commons License.

« UMass Amherst Faculty Collaboration Graph | Main | UMass Amherst Linguistics Sentiment Corpora»

Elliott Moreton Colloquium

Elliott Moreton*
UNC Chapel Hill

Syntagmatic simplicity bias in human and artificial learners

In collaboration with Joe Pater and Michael Becker** (Reed College)

Friday, February 20, 3:30 pm, Machmer W-23

Refreshments to follow in the department lounge

*2002 UMass Amherst Linguistics PhD
**2009 UMass Amherst Linguistics PhD

Abstract

Phonological dependencies in natural language tend to be assimilatory or dissimilatory, i.e., they relate two tokens of the same phonetic feature in an utterance. Six minimally-different experiments in phonotactic pattern learning by English speakers support the hypothesis that single-feature dependencies, even when they are not phonetically grounded, are detected more readily than two-feature dependences, even when they are phonetically grounded.

The single-feature learning bias is shown to emerge in a constraint- inducing and -weighting phonotactic learner when "Greek-letter" variables are restricted to instances of the same feature. The result of this restriction is that single-feature dependencies are supported by multiple overlapping constraints, leading to faster learning in Maximum Entropy and Gradual Learning Algorithm learners, whereas two- feature dependencies must be learned piecemeal. This treatment unifies the bias towards syntagmatically-simple patterns with that towards paradigmatically-simple ones, and points towards applications of constraint multiplicity and constraint generality in explaining learning and typology.