Predictive Processing in Story Comprehension.

Abstract

One endeavor within research on reading comprehension is to characterize how sentences are integrated into a coherent structure as the reader progresses through a text. Current approaches to sentence integration focus on backward search and inference processes to model the integration process. This dissertation examines the role of predictive processing in sentence integration. The research has two goals: (1) to test the hypothesis that skilled readers regularly make predictions as they read, and (2) to constrain a model of predictive processing by providing information about the processing consequences of having formed a prediction. The term "prediction" is used to refer to several types of forward inferences, from the minimal prediction of the general topic of the next sentence to a specific content prediction of what will happen next. Three experiments are reported. In all three experiments, stimuli were narrative text fragments which varied in the degree to which they generated a strong prediction at the end of the fragment. In the first experiment subjects were faster to respond that a target sentence was related to the text when it followed a High Expectation (HiE) text than a Low Expectation (LoE) text. Subjects were also faster to respond that a target sentence was unrelated when it followed a HiE text than when it followed a LoE text. In Experiment 2, when subjects read a target sentence which conveyed the next event in a script, they took longer when they had an incorrect expectation than when they had no expectation. Experiment 3 failed to provide evidence that readers were forming specific content predictions as they read. The results suggest that readers do generate predictions as they read. Furthermore, these predictions are generated selectively (not for every sentence), and they have processing consequences. A correct prediction can facilitate comprehension of the sentence where the prediction is fulfilled. An incorrect prediction can interface with the processing of a sentence which violates the prediction. This pattern suggests that predictions are allocated some attention when generated and become involved in the processing of subsequent sentences.