Evaluation of a refinement algorithm for the generation of referring expressions

Abstract

In this paper we describe and evaluate an algorithm for generating referring expressions that uses linear regression for learning the probability of using certain properties to describe an object in a given scene. The algorithm we present is an extension of a refinement algorithm modified to take probabilities learnt from corpora into account. As a result, the algorithm is able not only to generate correct referring expressions that uniquely identify the referents but it also generates referring expressions that are considered equal or better than those generated by humans in 92% of the cases by a human judge. We classify and give examples of the referring expressions that humans prefer, and indicate the potential impact of our work for theories of the egocentric use of language.