Roger van Gompel
University of Dundee
PRO: A Probabilistic model of Referential Overspecification
Most psycholinguistic research has focused on the formulation and phonological encoding stages of language production rather than on conceptualisation, that is, how speakers decide which concepts to express. In this talk, we will discuss PRO, a computational model that aims to explain concept selection during reference production and why speakers frequently overspecify (e.g., say large red candle when red candle is sufficient for identifying the object). PRO assumes that both the discriminatory power of a property (how many alternative referential candidates does it rule out?) and the relative preference for the properties of a referent play a role in concept selection. First, speakers choose a property that rules out all distractors; if there is more than one such property, then the probability with which speakers choose one is determined by the relative preference for that property. Next, speakers may add additional properties, with the chance that this happens being determined (1) by the preference for a property and (2) by the overspecification eagerness of a speaker. We will demonstrate that PRO accounts very well for concept selection in several experiments varying different properties (e.g., size, colour, category) and discuss future avenues of research for testing the model.
Roger van Gompel is a Senior Lecturer at the School of Psychology at the University of Dundee, where he has been since 1999 following his PhD in Psychology at the University of Glasgow. His research interests are the psychological processes underlying language production and comprehension, with a focus on sentence and reference processing. He has published experimental research articles on syntactic ambiguity resolution, the comprehension of anaphors, structural priming and the production of referring expressions.