Pronominalization and Expectations for Re-Mention: Modeling Coreference in Contexts With Three Referents

Jet Hoek,Hannah Rohde,Andrew Kehler

doi:10.3389/fcomm.2021.674126

Abstract

The relationship between pronoun production and pronoun interpretation has been proposed to follow Bayesian principles, combining a comprehender’s expectation about which referent will be mentioned next and their estimate of how likely it is that a potential referent will be re-mentioned using a pronoun. The Bayesian Model has received support from studies in several languages (English, Mandarin Chinese, Catalan, German), but tested contexts have been limited to two event participants, whereas natural language discourse often involves contexts with more than two event participants. In this study, we conducted three story continuation experiments to assess how the Bayesian Model performs in more complex contexts. Our results show that even in contexts with three event participants, comprehenders can behave rationally when interpreting pronouns, but that they appear to require sufficient context to build up a coherent representation of the situation to do so. In addition to testing the basic claim of the Bayesian Model (Weak Bayes), we test the central prediction of the Strong form of the hypothesis: that the two components of the model (next-mention expectations and choice of referring expression) are influenced by dissociated sets of factors. In a model comparison, Experiments 2 and 3 confirm the closest fit from the Bayesian Model, which supports Weak Bayes, and none of our experiments find evidence that the predictability of a referent affects pronominalization rates, which corroborates Strong Bayes. Finally, we test whether the rate of pronominalization is sensitive to factors related to ambiguity and argument/adjunct status of referents; we find that participants vary their production of pronouns most strongly based on the grammatical role of the antecedent (subject or not), with a smaller effect from the presence/absence of a gender-matched competitor and no effect from the syntactic position of this competing referent.

Highlights

Reduced reference to previously mentioned entities–such as that achieved via pronominalization–is a hallmark of coherent discourse
These results suggest that the Bayesian Model’s poor fit for the observed pronoun interpretation data in Experiment 1 was likely not due to properties intrinsic to the benefactive construction, but rather to the number of human event participants in the prompts
Whereas we opted not to adopt the type of continued-story task they used in their Experiment 1, we agree that contexts that support richer discourse representations might better approximate natural language understanding scenarios, when constructions as syntactically and semantically complex as benefactives are involved

Summary

Introduction

Reduced reference to previously mentioned entities–such as that achieved via pronominalization–is a hallmark of coherent discourse. A commonly held view is that speakers and hearers coordinate on the reference problem through a notion of entity salience: Speakers consult a set of factors that contribute to salience in deciding to use a pronoun, and hearers consult those same factors when interpreting it. In a context like (1), hearers are more likely to interpret a subsequent pronoun she as in (1-a) as referring to Jill than speakers are to produce a pronoun when referring to Jill in a subsequent sentence as in (1-b); likewise, speakers are more likely to use a pronoun to refer to Sue in (1-b) even though Sue will not be the preferred referent for the hearer in (1-a) (e.g., Stevenson et al, 1994; Kehler et al, 2008; Kehler and Rohde 2013)

Objectives

Methods

Results

Conclusion