Does referent predictability affect the choice of referential form? A computational approach using masked coreference resolution

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 454–469 November 10–11, 2021. ©2021 Association for Computational Linguistics 454 Does referent predictability affect the choice of referential form? A computational approach using masked coreference resolution Laura Aina1∗, Xixian Liao1∗, Gemma Boleda1,2 and Matthijs Westera3 1Department of Translation and Language Sciences, Universitat Pompeu Fabra 2Catalan Institution for Research and Advanced Studies - ICREA 3Leiden University Centre for Linguistics, Universiteit Leiden {laura.aina,xixian.liao,gemma.boleda}@upf.edu matthijs.westera@gmail.com Abstract It is often posited that more predictable parts of a speaker’s meaning tend to be made less ex- plicit, for instance using shorter, less informa- tive words. Studying these dynamics in the do- main of referring expressions has proven diffi- cult, with existing studies, both psycholinguis- tic and corpus-based, providing contradictory results. We test the hypothesis that speakers produce less informative referring expressions (e.g., pronouns vs. full noun phrases) when the context is more informative about the referent, using novel computational estimates of refer- ent predictability. We obtain these estimates training an existing coreference resolution sys- tem for English on a new task, masked corefer- ence resolution, giving us a probability distri- bution over referents that is conditioned on the context but not the referring expression. The resulting system retains standard coreference resolution performance while yielding a better estimate of human-derived referent predictabil- ity than previous attempts. A statistical analy- sis of the relationship between model output and mention form supports the hypothesis that predictability affects the form of a mention, both its morphosyntactic type and its length. 1 Introduction A long-standing hypothesis in linguistics relates predictability to linguistic form: the more pre- dictable parts of a speaker’s meaning tend to be communicated with shorter, less informative words and be pronounced more quickly and flatly (Fer- rer i Cancho and Solé, 2003; Aylett and Turk, 2004; Levy and Jaeger, 2007; Piantadosi et al., 2011). This is posited to result from a trade-off between clarity and cost: when the context already clearly conveys certain information, making it explicit would be inefficient. In the domain of referring expressions, or mentions, the prediction is that pro- nouns such as “they” and “her” are used for more * First two authors contributed equally. predictable referents while proper names such as “Kamala Harris” or full noun phrases such as “my child’s teacher” are required for less predictable referents (Tily and Piantadosi, 2009). The aim of this paper is to test this prediction using a novel, computational estimate of referent predictability. Existing work on the relation between pre- dictability and mention form paints a mixed pic- ture (Section 2). A set of studies in psycholinguis- tics used controlled stimuli designed to manipulate predictability (Arnold, 2001; Rohde and Kehler, 2014). An alternative approach uses naturally oc- curring corpus data to elicit judgments from human subjects (Tily and Piantadosi, 2009; Modi et al., 2017b). Neither approach, however, scales well. We instead obtain predictability estimates from a coreference system (trained on English data); if successful, this strategy would allow to test the hypothesis in …

Content truncated to 500 words