|Title:||Differential Difficulty in the Acquisition of Second Language Phonology|
|Authors:||Ellen Broselow, Zheng Xu|
|Abstract:||Differential Difficulty in the Acquisition of Second Language Phonology
This paper reports on Mandarin speakers' acquisition of English final voiced and voiceless obstruents and final labial nasals, none of which occurs in Mandarin codas. The learners' production patterns are compared with a simulation using the Gradual Learning Algorithm (Boersma & Hayes 2001). We demonstrate that by assuming the Mandarin Chinese rankings as the initial state and providing this system with representative English input, the GLA correctly models the order of acquisition of obstruent codas (voiceless before voiced), as well as the variation seen in the data between deletion of coda obstruents, insertion of a vowel after coda obstruents, devoicing of voiced final obstruents, and correct production of obstruent codas.
However, the GLA fails to correctly model the relative order of acquisition of obstruents and labial nasals. Because coda labials are relatively infrequent in the English input, the GLA predicts they should be acquired after coda obstruents, which are more frequent in the input. Yet speakers made many fewer errors with final labial nasals than with final obstruents. We outline several possible explanations of the differential acquisition of coda labial nasals and coda obstruents: (i) an articulatory account (e.g., Ussishkin and Wedel 2003): learners have a gestural program for producing nasal codas in their repertoire, so that production of [m] in coda position requires only superimposition of a labial closure on this gestural score, but their repertoire lacks a program for producing either voiced obstruents or obstruent closures not followed by a vowel; (ii) a perceptual account (e.g., Silverman 1992): nasal codas are more salient than obstruent codas; (iii) an account based on constraint induction (e.g., Hayes 1999): if we assume that the constraints banning obstruent codas and voiced obstruent codas are universal and innate, while the constraint banning [m] in coda is language-specific and learned, then we can build into the model a preference for demoting language specific constraints over universal ones.