Rosen (2016): Predicting the unpredictable: capturing the apparent semi-regularity of rendaku voicing in Japanese through Gradient Symbolic Computation

Direct link: http://roa.rutgers.edu/content/article/files/1559_rosen_1.pdf

ROA: 1289
Title: Predicting the unpredictable: capturing the apparent semi-regularity of rendaku voicing in Japanese through Gradient Symbolic Computation
Authors: Eric Rosen
Comment: To appear in Clem, Emily, Geoff Bacon, Andrew Cheng, Virginia Dawson, Erik Hans Maier, Alice Shen, & Amalia Horan Skilton (eds.). 2016. Proceedings of the 42nd Annual Meeting of the Berkeley Linguistics Society. Berkeley: Berkeley Linguistics Society.
Length: 15
Abstract: Semi-regular phonological processes occur often in natural language. For example, rendaku voicing in Japanese fails to occur in a seemingly random fashion among roughly 16% of certain classes of compounds. This presents an analytical challenge for generative theories with exceptionless rules or categorical constraints: irregularity of any kind must arise within lexical representations, not the grammar. For example, the compound kuma (‘bear’) + te (‘hand’) -> kuma-de (‘rake’) voices. But yama (‘mountain’) + te (‘hand’) -> yama-te (‘mountainside’) doesn’t. There is no known phonological distinction between kuma and yama that enables a rule or constraint to explain the difference. The necessity of listing exceptional surface forms like yama-te undermines rendaku’s status as a systematic phonological process (Kawahara 2015, Vance 2014).

 

This work proposes a new analysis of rendaku that solves this problem, allowing the correct output forms to be generated without specifying voicing for particular, “exceptional” compounds. I adopt the framework of Gradient Symbolic Computation (Smolensky and Goldrick 2015, henceforth GSC), a type of Harmonic Grammar (Legendre, Miyata and Smolensky 1990) that allows weighted constraints and features with continuous activation levels. In this analysis, rendaku voicing occurs by the coalescence of two stem-specific, partially activated [+voi] features that occur as attributive affixes on compound-forming stems: a variation of the junctural morpheme for rendaku proposed by Ito and Mester 1998. Only when the additive combination of these features exceeds some threshold t does voicing occur. In the above examples, [+voi]_kuma + [+voi]_te > t > [+voi]_yama + [+voi]_te. The contribution of both conjuncts to voicing captures not only the well-known gradient continuum of voicing preference/dispreference among second conjuncts (Irwin 2015) but also a lesser-known gradient effect of first conjuncts on voicing.

 

Adopting the principle of Minimum Description Length (Goldsmith 2011) I show that GSC can provide a better model of this semi-regular phenomenon than other frameworks by reducing the degree of lexicalization with minimum cost to the grammar. Moreover, computer-simulated algorithms show that this proposed grammar is learnable. This analysis holds promise that the GSC model can shed new light on the lexicalization vs. grammaticalization question with respect to other semi-regular processes.

Type: Paper/tech report
Keywords: phonology, gradient symbolic computation, semi-regularity, Japanese rendaku voicing, minimum description length
Facebooktwittergoogle_plusredditpinterestlinkedinmail
Posted in Research (e.g. papers, books)