Title Pending

Most phonologists assume that phonological processes are encoded in terms of natural classes defined by sets of features. Thus, children acquiring a phonological grammar need some way to determine the natural classes identifying target and trigger segments. Exactly how children solve this feature specification problem (FSP) is unknown. Many phonologists endorse the view that the feature specifications should be minimal in the sense of containing the minimal number of feature specifications necessary to uniquely identify the class of segments, but no algorithm has been proposed to achieve this objective. We show that there is no guarantee that there is a unique minimal definition of every natural class. Then, drawing on recent work by Chen & Hulden (2018), we show that any algorithm which minimizes the number of features is intractable in the sense that it can only be solved by sifting through trillions of candidate natural classes. Recognizing that feature minimization is intractable, we propose an alternative objective, maximization, and a tractable algorithm that yields a unique result to implement it. Finally, we argue that the bias for minimal specification reflects a misapplication of Occam’s Razor to acquisition.