Extending Phone Prediction Models of Word Segmentation to a More Realistic Representation of Prosody
Loading...
Date
2009-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Ohio State University
Abstract
Laboratory results (e.g. Saffran et al. (1996)) have shown that infants can use statistical cues for word segmentation, and Christiansen et al. (1998) propose Simple Recurrent Networks (SRNs) as a model for this phenomenon that can integrate prosodic and segmental cues. I describe an augmentation of the corpus in Rytting (2007) with measurements of acoustic prosodic correlates, and perform a series of SRN phone prediction experiments on this enriched corpus. Although certain information-theoretic properties of this enhanced corpus suggest that prosodic correlates are predictive of word boundaries, two sets of experiments suggest that an SRN phone prediction task is an unsuitable basis for finding the strongest prosodic predictors of word boundaries. The first set explores manipulations on the inputs presented to the model, and the second set explores modifications of the model itself. I close by describing peculiarities of SRNs and the phone prediction task, and presenting desiderata of models for integrating acoustic correlates to prosody in word segmentation.
Description
Keywords
Language Acquisition, Computational Linguistics, Simple Recurrent Networks, Prosody, Word Segmentation