Comparing models of pronoun production and interpretation via observational and experimental evidence

Abstract

Pronouns like she are frequently produced by speakers to refer to entities in discourse. For communication to be successful, comprehenders must be able to interpret these pronouns by identifying the appropriate referent. In the existing literature, three main models of pronoun production and interpretation have been proposed. These models have traditionally been tested through story continuation tasks, using carefully designed stimuli. In our study, we take a different approach by utilizing naturalistic passages from corpora, in two analyses, one observational and one experimental. Our analyses support the Bayesian model. In this model and in our experimental data, the relationship between pronoun production and interpretation can be captured using Bayes’ rule. Specifically, pronoun interpretation is affected both by the probability that the referent will be mentioned next and by the probability that a pronoun will be used to refer to that referent. Moreover, both observational and experimental data provide evidence that pronoun production biases are insensitive to a set of meaning-driven factors — here, discourse relations — which do affect pronoun interpretation, in line with the prediction of the so-called strong form of the Bayesian Model.