Non-deterministic attribute selection in reference production

In producing identifying descriptions, speakers often overspecify and manifest preferences for certain attributes. However, current computational models which incorporate this observation tend not to make precise predictions about when and how much speakers do this. The present paper proposes and evaluates two alternative models, based on the results of a new experiment. Unlike current models, the new ones are nondeterministic and seek to make precise quantitative predictions about the extent to which speakers overspecify.