Mende tonotactics in surface optimizing multilevel grammar

This talk presents a quantitative study of lexical tonotactics in Mende (Mande, Sierra Leone). Surface tone patterns in Mende—known also as 'tone melodies'—were once an empirical centerpiece in the development of Autosegmental Phonology (AP; e.g., Leben 1978). AP explained the commonality of certain surface tone patterns by positing a limited inventory of underlying melodies, which procedurally map to surface tone bearing units via universal association conventions in the representation. However, since the introduction of surface optimizing grammars (e.g., Optimality Theory: Prince & Smolensky 1993; Harmonic Grammar: Legendre et al. 1990), there has been a marked move away from representational theories like AP, and optimization over correspondence relationships has supplanted AP in much of its original empirical domains. Moreover, there are noted problems with AP’s inventory-and-mapping approach for Mende, and data show that the assumptions of limited melodic inventories and universal tone mapping cannot coexist without significantly weakening one or another fundamental element of AP (e.g., Dwyer 1978; Conteh et al. 1983; Zhang 2002).

            I argue here that Mende tonotactic patterns can be modeled using an optimizing grammar, based not on representations but on surface phonological correspondences. Surface correspondences are rooted in independently-motivated principles of similarity and proximity interaction, which have proven effects for phonotactics in the segmental domain (e.g., Hansson 2001; Rose & Walker 2004; Wayment 2009; Rhodes 2012; Bennett 2013; Inkelas & Shih 2014). The results reveal that Mende’s tone patterns and their lexical frequencies emerge naturally from a set of cross-linguistic principles that optimize phonotactics, some of which are not predicted under AP approaches: e.g., alignment of transitions with syllable boundaries (cf. Hyman 2007’s “Principle of Ups and Downs”). The analysis is couched in a multilevel implementation of probabilistic grammar (e.g., an extension of Maximum Entropy Harmonic Grammar: MaxEnt; Goldwater & Johnson 2003), which captures lexically-driven variation in Mende tonotactics beyond what was reported in the original Leben 1978 study. This multilevel approach allows for the quantification of lexically-conditioned phonological variation in natural language, addressing the current stalemate between single grammar (e.g., Ito & Mester 1994; Alderete 2001) versus multiple grammar (e.g., Anttila 2002; Inkelas & Zoll 2005; Kiparsky 2008) approaches to morphophonology. Taken altogether, the reexamination of Mende provides a united and more fine-grained treatment of the Mende tone system, without the representational mechanisms required in AP. The conclusions illustrate that generalisations about tonotactics that original drove the development of AP in fact offer unique arguments in support of probabilistic surface optimizing grammars.

(based on joint work with Sharon Inkelas, University of California, Berkeley)