This dissertation focuses on replacive grammatical tone in the Dogon languages of Mali, where a word's lexical tone is replaced with a tonal overlay in specific morphosyntactic contexts. Unlike more typologically common systems of replacive tone, where overlays are triggered by morphemes or morphological features and are confined to a single word, Dogon overlays in the DP may span multiple words and are triggered by other words in the phrase. DP elements are divided into two categories: controllers (those elements that trigger tonal overlays) and non-controllers (those elements that impose no tonal demands on surrounding words). I show that controller status and the phonological content of the associated tonal overlay is dependent on syntactic category. Further, I show that a controller can only impose its overlay on words that it c-commands, or itself.

I argue that the sensitivity to specific details of syntactic category and structure indicate that Dogon replacive tone is not synchronically a phonological system, though its origins almost certainly lie in regular phrasal phonology. Drawing on inspiration from Construction Morphology, I develop a morphological framework in which morphology is defined as the idiosyncratic mapping of phonological, syntactic, and semantic information, explicitly learned by speakers in the form of a construction. The same constructional format can be used for both word-level and phrase-level phenomena. Under this view, Dogon replacive tone is idiosyncratic phonology (tonal overlays) associated with syntactic structure (category and c-command).

A key feature of the Dogon system of grammatical tone is that when two or more controllers target the same word(s), conflicts can arise, with each constructional schema demanding different output forms. I argue that morphological constructions act as constraints on output form, and cases of competition are resolved through constraint interaction. I further show that the cyclic spell-out of syntactic structure can block the application of overlays, but that this blocking is a language-specific parameter, suggesting violable constraints protecting the morphophonological form of phases. In most cases, strict ranking (as in Optimality Theory) is sufficient to account for the data, but in two languages, Tommo So and Nanga, even nonstochastic patterns require constraint ganging, indicating the need for weighted constraints. I couch the analyses of Tommo So and nine other Dogon languages in Maximum Entropy Grammar, a stochastic version of Harmonic Grammar that can account for both constraint ganging and for surface variation. I show that the same constraint set, suitably weighted, is able to capture eight of the ten languages. The other two languages, Tiranige and Togo Kan, have followed different diachronic paths and require different constraints but still fit into the constructional framework.

The framework developed in this dissertation seeks to explain how restructured phrasal phonological systems are learned and implemented. I show how predictions of the model regarding the nature of the trigger, target domain, and resolution of conflicts are upheld by other phenomena outside of the Dogon language family.