Coherence-driven predictability and referential form: Evidence from English corpus data

Abstract

To refer to a discourse entity, speakers need to choose from a variety of expressions, such as a name Donald Trump or a pronoun he. While evidence has accumulated that more predictable words are more likely to be phonologically reduced, there is a long-standing debate regarding whether more reduced referring expressions (e.g., pronouns) are more frequently produced for more predictable referents. This study contributes to this debate with new evidence in two aspects. 1) Rhetorical relation-driven predictability: while the cases that have been studied in previous psycholinguistic studies have been fairly restricted to some particular verbs types, this study is, to the best of our knowledge, the first attempt of broadening the empirical base with expectation primarily driven by rhetorical relations; 2) Naturally-occurring language in corpus data: instead of using constructed language to elicit production, we make use of natural coreference chains that can be automatically retrieved from corpora developed in the field of Computational Linguistics. We found uniform pronominalization rates across rhetorical relations despite the different next-mention rates, supporting a dissociation between likelihood of next mention and pronoun production.