Posted - 03/05/2001 : 14:32:56
Department of Chinese, Translation & Linguistics
&
Institute of Chinese Linguistics
Joint Seminar
Modeling Tone Acquisition
By Professor Jane TSAY
National Chung Cheng University
Time: 4:30 - 6:00 pm
Date: Friday, 11 May 2001
Venue: Room P4909, Academic Builiding, CityU
Abstract
Tone theories based on binary articulator-based features (e.g. Yip
1980, Bao 1990, Duanmu 1990) divide tone levels into universal categories
that should be respected by children learning a tone system. For example,
Yip (1980) would predict that the adjacent middle tones in a four-level tone
system cannot form a natural class, since they are represented with distinct
features (i. e. [+upper, -raised] vs. [-upper, +raised]). Moreover, in a
three-level tone system, the representation of the mid tone is ambiguous.
Theories like Yip's also predict that four-level tone systems should be as
natural and easy to learn as three-level tone systems, since both require
the same number of binary tone features to represent. By contrast, we
hypothesize that lexical tone should be represented by a multi-valued tone
feature based on the psychoacoustic properties of pitch (following Tsay
1994).
This hypothesis makes at least two predictions regarding tone acquisition.
First, the more tone levels a language has, the more difficult it is to learn,
due to constraints of perception and memory. Second, acoustically adjacent
tones should act as natural classes in language acquisition. To make our
predictions more concrete, we built two connectionist models, using the tlearn
software by Plunkett and Elman (1997). The first network was to test the
learning rate for two-, three-, and four-tone (level) systems, respectively,
using artificial data. The results showed that learning a two-tone system
was easier than learning a three-tone system, both being much easier than
learning a four-tone system. This is consistent with Maddieson (1978) who
found in a random sample of tone systems, as the number of tone levels
increases,
the number of languages decreases. The second network was trained using data
based on pitch extractions of real adult productions of Taiwanese syllables.
As expected, the model was more likely to confuse acoustically adjacent
tones:
H was often confused with M but very rarely with L, L showed the opposite
pattern, and M was often confused with both. These results are consistent
with longitudinal data (2;1-2;3) of children acquiring Taiwanese as their
first language (TAICORP, in preparation).
About the Speaker: Professor Jane Tsay obtained her doctoral degree in
Linguistics from the University of Arizona specializing in the phonology of
tone. Her recent research has been concerned with the syntax-phonology
interface, as well as phonological and morphological development in children,
with special reference to classifiers and tone sandhi in Taiwanese. She is
currently Associate Professor at the Graduate Institute of Linguistics,
National Chung Cheng University.
Enquiries: 2788-8705
___________________ All are welcome! ____________________
|