A Bayesian Model of Linguistic Prominence

This paper presents a Bayesian model of linguistic prominence. Prominences of two kinds arise as natural components of the interpretation of overt forms to infer covert communicative intentions. One form of prominence reflects the distinctiveness of the overt representation. This we call {form prominence}. The other kind of prominence reflects the likelihood of particular covert linguistic intentions being the message intended. This is {predictive prominence}. Some implications of the model are explored. First, the model accounts for the uniform information density findings arising at many levels of linguistic structure. Second, it also explains the previously noted complementarity of form and predictive prominence. Finally, the model offers reasons for why ambiguous forms are interpreted as communicating the messages with the highest predictive prominence.