Probabilistic pragmatics, or why Bayes’ rule is probably important for pragmatics

Pragmatics is about language use in context. This involves theorizing about speakers’ choices of words and listeners’ ways of interpreting. More often than not, this also involves a certain amount of noise and uncertainty: speakers and listeners may not know exactly what the relevant contextual parameters are, they may make mistakes, believe that their interlocutor is uncertain and possibly prone to err, etc. We believe that taking this picture seriously can, despite its apparent messiness, inspire a stringent formal approach to pragmatics that lends itself to precise empirical testing. We call it probabilistic pragmatics here, to emphasize the role that probabilities play in it. But it contains much more. In the following, we try to sketch its main characterizing features in relation to other approaches and give some example applications. We argue that probability models are the natural and most practicable tool for modeling the richness of pragmatic phenomena which are affected by many unknown contextual factors. Sections 2 and 3 characterize probabilistic pragmatics. Section 2 discusses different levels of analysis in pragmatic theory, so as to contrast probabilistic pragmatics with alternative approaches. Section 3 discusses key properties of probabilistic pragmatics. Sections 4, 5 and 6 sketch examples of applications. Section 4 introduces a baseline model for reasoning about referential expressions to demonstrate how the probabilistic modeling, inspired by classical pragmatic theory, can be fit to experimental data. Section 5 exemplifies further ways in which probabilistic pragmatics can shed light on gradient patterns in empirical data. The leading example for illustration is that of scalar implicature. Section 6 argues that considering (multiple levels of) gradient subjective contextual uncertainty, as captured by a probability distribution, is essential to understanding indirect speech acts. This section demonstrates how explicit representations, inspired from game theory, of interlocutors’ preferences and likely dialogue moves help tackle indirectness of speech in non-cooperative contexts.