van Gompel, R. P. , Snoeren, R. , Goudbeek, M. & Krahmer, E.
University of Dundee, bol.com, University of Tilburg, University of Tilburg
Speakers often include more information in noun-phrase descriptions than required for identifying the referent amongst its visual distractors (Engelhardt et al., 2006; Deutsch & Pechmann, 1982). We tested PRO (Gatt et al., 2013), a computational algorithm that makes quantitative predictions about the likelihood speakers use particular concepts in such overspecifications. It claims that speakers first select a concept that rules out all distractors. If there is more than one, then the likelihood of selecting each is determined by its preference. Next, the likelihood of adding further concepts is determined by the same preference.
Whereas our previous experiments tested visual environments where only two concepts/properties provided discriminating information, the current study examined whether PRO generalises to environments with three such concepts/properties (colour, size, border), one being postnominal (border). This makes the number of possible descriptions (e.g., blue bottle, large blue bottle in circle) larger, resulting in more complex decisions for PRO. We used conditions in which either one, two or three properties each ruled out all distractors (the other properties ruled out one distractor). The PRO predictions closely matched the proportions that participants produced particular descriptions, suggesting that PRO provides a good model of the mechanisms underlying concept selection and overspecification.