Learning Repairs for Marked Structures

This paper presents an algorithm which learns phonological grammars. These phonological grammars— like grammars in Optimality Theory (OT, Prince & Smolensky, 2004) or rule-based grammars (Chomsky & Halle, 1968)—describe phonological maps, or the transformations from input (underlying) forms to output (surface) forms (Tesar, 2014). Like OT grammars, the grammars which underlie the algorithm distinguish marked structures from their possible repairs. The advantages of this algorithm include 1) it is provably efficient and correct, and 2) it has significant empirical coverage in that the grammars it acquires can model many phonological maps, including opaque ones which are hard to represent in classical OT. To illustrate the algorithm and these advantages, we focus on repair typologies, by which we mean ‘international conspiracies’ in which different processes in different languages all target a single marked structure. In a classic example from Pater (2004), languages that avoid nasal-voiceless obstruent sequences (NC ̊ ) employ various processes such as fusion, denasalization, and voicing of the obstruent. Examples from Indonesian, Toba Batak, and Quechua, respectively, are presented below in (1), (2), and (3).