A number of puzzles about speech and language are solved, or at least simplified, if we adopt the view that some kinds of linguistic knowledge, normally modeled as symbolic strings or structures, are instead discrete random variables: that is, probability distributions over a space of possible strings or structures. This perspective also allows a radically simple piece of antique learning theory to be brought to bear, with results that may be surprising to some. Illustrations of this claim will include the development of shared vocabulary within a speech community; the case of phonological "near mergers"; many of the phenomena generally cited in support of "exemplar theory", e.g. the phonological effects of common contexts of use; and some key aspects of the puzzling problem of intersubjective (dis)agreement in linguistic annotation.

Mark Liberman is Trustee Professor of Phonetics; Professor in the Departments of Linguistics and Computer Science; and Director of the Linguistic Data Consortium -- all at the University of Pennsylvania.