Indexing expectations of informativity

Abstract

Predictions about upcoming content are important to language processing and, among other things, serve to facilitate communicative interactions. In studying how guesses about upcoming content are generated, comprehenders have been found to rely on real world knowledge, such that, for example, real-world typical content is associated with processing ease. This preference for typical content has been found in both online and offline tasks indexing people’s expectations about upcoming content. However, this does not align with the production choices that speakers tend to make; speakers favour the inclusion of non-typical, informative content and will often omit typical and thereby easily inferable content where optional to do so.

This thesis attempts to reconcile this asymmetry between what comprehenders prefer and what speakers do. Specifically, it posits that in addition to a comprehender’s knowledge of the world, such as what situations and events are typical, comprehenders also take into account speakers’ production preferences when they make guesses of what someone will say next. To investigate this, I have developed a paradigm testing how variations in the much-used sentence completion or Cloze task (Taylor 1953) can induce variations in the kinds of responses participants provide. In this way, this thesis indexes comprehenders’ expectations of the informativity of upcoming content and further examines how fine-grained these informativity expectations are. Additionally, I posit a method for measuring informativity that relies on five distinct measures, each of which is intended to capture a different sense in which someone can be informative.

An expectation for upcoming content to be about real-world typical situations would reflect an expectation in comprehenders for language use to be transparent; i.e. that speakers tend to communicate about the world as it typically is. However, if comprehenders are sensitive to the production preferences of speakers, they should expect content that cooperative speakers are likely to mention, rather than simply content that is likely to be the case in the real world; that is, they should expect filtered language use.

Experiment I tests this prediction by manipulating the salience of the speaker across four Cloze task conditions to see whether an increased emphasis on the speaker prompts participants to estimate more informative content. Results show that the most speaker-salient condition yielded the most informative sentence completions. I argue that this reflects an increased awareness of speaker intentions and thereby speakers’ production preference to be informative. Experiment I thus establishes that, despite the well-demonstrated finding that comprehenders rely on real-world knowledge when anticipating content, they also have expectations of informativity.

The following two experiments ask how malleable comprehenders’ expectations for informativity are by testing aspects of the context that may affect comprehenders’ content estimates. Experiment II tests properties of the speaker: Participants are familiarised with two different speakers who vary in the informativity of their utterances. When completing utterances from each of the two speakers, comprehenders provide more informative completions for the HIGH-INFORMATIVITY speaker compared to the LOWINFORMATIVITY speaker. This shows that comprehenders are able to adapt their expectations of informativity to individual speakers’ communicative styles. Experiment III tests the role of the addressee: As in Experiment I, participants complete utterances from several different speakers, with the manipulation that the utterances are addressed either to an adult or to a child. Although less clear-cut than in the preceding experiments, results indicate that the identity of the addressee affects comprehenders’ guesses such that they provide less informative, more typical completions in child-addressed utterances.

Experiments I-III begin to model the informativity bias observed in adult comprehenders by positing that (at least) two main factors underlie the process of generating guesses for upcoming content: (Real-)world knowledge and sensitivity to speakers’ production preferences. Experiment IV tests the robustness of this model by extending the paradigm to children, a population in which one or both of these underlying factors may vary - children’s knowledge of the world is different to adults’, and how attuned they are to speakers’ productions may also be different. As such, Experiment I tests if speaker saliency influences children’s guesses of what someone will say next by using a simplified version of Experiment I. Overall, results indicate that children rely more on their world knowledge when completing sentences than adults do, although there is some indication that speaker salience plays a role in older children’s guesses. I discuss how this finding informs our understanding of the interplay of world knowledge and perspective-taking in (typical) development, as well as future directions of study which could further illuminate this issue.

Overall, this thesis demonstrates that comprehenders are sensitive to speakers’ production preferences and have an expectation for informativity when they generate guesses of what someone will say next. In other words, comprehenders seem to have a bias towards filtered language when content is conveyed by a speaker. These findings contribute to the wider field in three ways. Firstly, it demonstrates that sentence completion tasks are not necessarily static measures of predictability and need precise fine-tuning to answer the research questions one is interested in. Secondly, it proposes a method for measuring informativity that attempts to capture several strategies that may be used by a speaker intending to be informative and which the comprehender might take into account when generating their guesses about upcoming content. And finally, this thesis shows that our models of language processing need to include a role for informativity-driven reasoning about the speaker to more accurately capture the processes at play.