Bayes nets and the dynamics of probabilistic language

Abstract

This paper is about a shared concern of linguistic semantics and pragmatics, epistemology, and many other areas of cognitive science: the formal representation of information and uncertainty. It is common in many of these areas, and increasingly in linguistics, to represent agents’ information using probability—an enrichment of classical semantics. However, each application of probability must answer difficult questions about whether a probabilistic representation is rich enough to capture the nuances of our information states. Two such problems—the epistemological distinction between uncertainty and ignorance, and the dynamic effects of probabilistic language in formal pragmatics—seem to suggest a negative answer, and to support a more complicated model that repesents uncertainty in terms of sets of probability measures. Following insights due to Judea Pearl (1988), I argue that the simplest probabilistic approach may be sufficient to handle these problems if we pay close attention to the hierarchical structure of information states, as encoded explicitly in the graphical representation of relationships among questions known as “Bayesian networks” or “Bayes nets”.