Comprehension priming as rational expectation for repetition: Evidence from syntactic processing

Abstract

Why do comprehenders process repeated stimuli more rapidly than novel stimuli? We consider an adaptive explanation for why such facilitation may be beneficial: priming is a consequence of expectation for repetition due to rational adaptation to the environment. If occurrences of a stimulus cluster in time, given one occurrence it is rational to expect a second occurrence closely following. Leveraging such knowledge may be particularly useful in online processing of language, where pervasive clustering may help comprehenders negotiate the considerable challenge of continual expectation update at multiple levels of linguistic structure and environmental variability. We test this account in the domain of structural priming in syntax, making use of the sentential complement-direct object (SC-DO) ambiguity. We first show that sentences containing SC continuations cluster in natural language, motivating an expectation for repetition of this structure. Second, we show that comprehenders are indeed sensitive to the syntactic clustering properties of their current environment. In a series of between-groups self-paced reading studies, we find that participants who are exposed to clusters of SC sentences subsequently process repetitions of SC structure more rapidly than participants who are exposed to the same number of SCs spaced in time, and attribute the difference to the learned degree of expectation for repetition. We model this behavior through Bayesian belief update, showing that (the optimal degree of) sensitivity to clustering properties of syntactic structures is indeed learnable through experience. Comprehension priming effects are thus consistent with rational expectation for repetition based on adaptation to the linguistic environment.

Keywords: Expectation; Priming; Probabilistic models of cognition; Psycholinguistics; Rational analysis; Syntax.

Figure 1

(a) Distances between sentences containing SCs in the Brown Corpus, computed as number of sentences intervening between SCs. Baseline (geometric distribution) reflecting hypothetical random distribution of SCs throughout the corpus is plotted in red. (b) For each corpus document, observed number of immediate SC repetitions (black or gray stars) versus expected number of repetitions given random ordering of sentences in the document (red squares).

Figure 2

Structure of the self-paced reading experiments.

Figure 3

(a) Region-by-region residual reading times on the second sentence of critical test vignettes in Experiment 1, conditioned on training condition and sentence continuation (all training lengths). (b) Residual reading times aggregated over Disambiguation, Spillover, and Conclusion regions. (c) Residual reading times aggregated over Disambiguation and Spillover regions.

Figure 4

(a) Region-by-region residual reading times on the second sentence of critical test vignettes of structure SC.SC in Experiment 1, conditioned on training condition and presence of disambiguating complementizer that. (b) Residual reading times aggregated over Disambiguation, Spillover, and Conclusion regions (all training lengths). (c) Residual reading times aggregated over Disambiguation and Spillover regions (all training lengths).

Figure 5

(a) Region-by-region residual reading times on the second sentence of critical test vignettes in Experiment 2, conditioned on training condition and sentence continuation (all training lengths). (b) Residual reading times aggregated over Disambiguation, Spillover, and Conclusion regions. (c) Residual reading times aggregated over Disambiguation and Spillover regions.

Figure 6

(a) Region-by-region residual reading times on the second sentence of critical test vignettes of structure SC.SC in Experiment 2, conditioned on training condition and presence of disambiguating complementizer that. (b) Residual reading times aggregated over Disambiguation, Spillover, and Conclusion regions (all training lengths). (c) Residual reading times aggregated over Disambiguation and Spillover regions (all training lengths).

Figure 7

Graphical representation of the Bayesian belief-update model (variable explanations in Table 7.1; full specification in Section 7.1) (a) Complete model fit to Experiment 1 data. (b) Fragment of analogous model for Experiment 2: n plate in the case of a three-sentence vignette. Nodes are added analogously in the cases of four- and five-sentence vignettes. Following the structure of the model in (a), the following dependences always hold: ρ₀ and ρ_m to all V; θ₀ and θ_m to all C; and α to all V and C.

Figure 8

Model predictions for test vignettes.

Figure 9

Processing advantage of clustered training (versus anti-clustered training) for final three regions of second sentence of test vignettes, given an SC continuation in the first sentence and an SC-DO verb in the second sentence. (a–b) Reading times at test, given anti-clustered experience, minus reading times at test, given clustered experience. Error bars reflect 95% confidence intervals of by-subject means. (c) Model-predicted surprisal at test, given anti-clustered experience, minus surprisal at test, given clustered experience. See Sections 7.2–7.3 for model prediction details.