The Probabilistic Turn in Semantics and Pragmatics

This article provides an overview of graded and probabilistic approaches in semantics and pragmatics. These approaches share a common set of core research goals: (a) a concern with phenomena that are best described as graded, including a vast lexicon of words whose meanings adapt flexibly to the contexts in which they are used, as well as reasoning under uncertainty about interlocutors, their goals, and their strategies; (b) the need to show that representations are learnable, i.e., that a listener can learn semantic representations and pragmatic reasoning from data; (c) an emphasis on empirical evaluation against experimental data or corpus data at scale; and (d) scaling up to the full size of the lexicon. The methods used are sometimes explicitly probabilistic and sometimes not. Previously, there were assumed to be clear boundaries among probabilistic frameworks, classifiers in machine learning, and distributional approaches, but these boundaries have been blurred. Frameworks in semantics and pragmatics use all three of these, sometimes in combination, to address the four core research questions above.