Modeling Taiwanese speakers’ knowledge of tone sandhi in reduplication

Recent studies on productivity have shown that native speakers’ phonological knowledge not only includes statistical patterns in the lexicon, but also patterns that cannot be gleaned from the lexicon. This is demonstrated in speakers’ analytical bias towards abstract phonological representations (Davidson, 2005), single-feature dependency (Moreton, 2008), perceptually motivated phonological scales (Zuraw, 2007), as well as their difficulties with exceptionless opaque patterns in wug tests (Zhang and Lai, 2008). Based on the results of a wug-test, we show in this paper that the opaque tone sandhi pattern in Taiwanese reduplication is a case in which the speakers’ knowledge is a combination of more than, less than, and exactly what their lexicon informs them of: the phonetic effects of the sandhis are overlearned, the opaque tone sandhis are underlearned, and the lexical statistics are properly learned. We further argue that a substantively biased Maximum Entropy grammar that encodes learning biases against lexical listing and phonetically unmotivated patterns can model the simultaneous underlearning, overlearning, and proper learning of the lexical patterns by Taiwanese speakers.