Evaluating an Expectation-Driven Question-Under-Discussion Model of Discourse Interpretation

Abstract

According to Question-Under-Discussion (QUD) models of discourse interpretation, clauses cohere with the preceding context by virtue of providing answers to (usually implicit) questions that are situated within a speaker's goal-driven strategy of inquiry. In this article we present four experiments that examine the predictions of a QUD model of interpretation when cast in terms of an integrated, expectation-driven model of discourse processing. The results of these studies together support the predictions of the model, demonstrating that contextual cues affect comprehenders' expectations about ensuing QUDs (Experiment 1), QUD expectations in turn influence the interpretation of discourse-dependent linguistic expressions (Experiment 2), and the biases associated with those expressions in turn influence the anticipation of QUDs (Experiments 3a and b).

Funding

This research was supported by an Andrew W. Mellon postdoctoral fellowship to Hannah Rohde. We thank research assistants Erica Gold, Brittany Young, and Meredith Larson for their work in annotating and conducting the studies.

Notes

¹ A remaining question, not addressed by the current work, is whether such evidence strongly favors QUD analyses over other analyses of discourse coherence on offer. We return to this question in the General Discussion.

² Von Kuppevelt (1995) in fact proposed that a set of possible questions arise in a listener's mind during discourse comprehension and that these questions are entertained as ones that may be answered in subsequent discourse. He leaves open the question of which factors in a context contribute to the anticipation of these upcoming questions, however.

³ Note that although we are interested in investigating the effect of IC verbs on the number of Why? questions, we are not addressing the question of what properties of these verbs (or of the eventualities they denote) are responsible for the effect. For discussion of this issue, see e.g. Hartshorne and Snedeker (2013), Bott and Solstad (2014), and Hartshorne et al. (2015), inter alia.

⁴ We also performed a correlation analysis between the distributions of coherence relations in the free-prompt monologue condition (2a-b) and of QUDs posed in the dialog condition. Because the mapping between coherence relations and QUDs is not one-to-one (e.g., Violated-Expectations are somewhat common in monologue, but rarely posed as a question), we expect only a partial alignment between the percentages. If we include the 1,514 full-stop monologue responses from Kehler et al. (2008) in our analysis, we can calculate the proportions for each participant or each item across verb types. A linear model of the proportion of Why?-type questions as a function of the proportion of Explanation relations across verb types, by participants and by items, show that the proportions (of Why?-type questions and Explanation coherence relations) are reliably correlated across the IC and non-IC conditions (adjusted R² = 0.37, F(1,148) = 88.39, p < .001; adjusted R² = 0.44, F(1,77) = 63.23, p < .001). This demonstrates that, at least to some extent, participants not only use contextual clues to anticipate ensuing QUDs but also that they continue passages by answering those same QUDs.

⁵ The use of 21 stimuli was a result of carrying over the verb set from Rohde et al. (2006), who had included a three-way verb classification in their analysis.

⁶ An anonymous reviewer points out that stimuli such as (7), in which the pronoun is Goal-referring, violate Rule 2 of Centering theory (Grosz, Joshi, & Weinstein, 1995), since a repeated name is used to refer to the previous subject (the presumptive backward-looking center, or C_b), whereas a pronoun is used to refer to a non-C_b. This is not true of stimuli like (6), in which the pronoun is Source-referential. This may explain why a slowdown was found in the spillover region of the Goal-referential stimuli as compared to the Source-referential ones. It does not explain our interactive effects, however, since the Goal-referential stimuli in both instruction conditions were the same in this respect. Indeed, as indicated above, the Goal-referential stimuli in the What happened next? condition were read the fastest of all four conditions.

⁷ Data regarding reference production and interpretation from this experiment was previously reported in Rohde and Kehler (2014). Here we analyze the data with respect to our hypotheses regarding the distribution of coherence relations in participant continuations, which were not addressed in that work.

⁸ To be clear, this does not mean we expect that the overall pronoun interpretation bias will be toward the subject, and hence yield an overall bias toward subject-biased coherence relations, particularly for IC-2 contexts. The prediction is that including a pronoun in the prompt will lead to more first-mentions of the subject when compared with the full-stop condition and that this will lead to comparatively more subject-biased relations.

⁹ As reported in Rohde and Kehler (2014), the analysis of coreference confirms the predicted higher rates of (1) pronominalizations of rementions of subject referents as compared with nonsubject referents in the no-pronoun condition and (2) continuations about the previous subject in the pronoun prompt condition as compared with the no-pronoun condition.