Anticipatory prediction during online language processing

Abstract

Most investigations of linguistic prediction focus on evidence of predictability benefits when comprehenders encounter expected input during reading. However, there remain several unresolved empirical issues that are important for the broader question of whether prediction plays a fundamental role during real-time language comprehension. These include whether there are processing costs for misprediction, what the contents of predictions are, and whether readers differ in the extent to which they engage in prediction. In six experiments, these issues were systematically investigated by presenting different groups of readers with predictable words and unpredictable alternatives that were either semantically related or unrelated in constraining or non-constraining context conditions. The primary methodology was the recording of eye movements during natural reading for comprehension. Self-paced reading was also used to assess the contribution of stimuli presentation format on predictive processing. Across most experiments, there was evidence of early and late processing benefits for predictable completions in constraining contexts, which also extended to unpredictable completions that were semantically related. However, evidence of immediate processing costs for unexpected input that replaced readers’ predictions was more mixed and appeared to depend on a variety of linguistic and non-linguistic factors. Overall, these results provide some support for the idea that the language processor is a “prediction machine” in line with general predictive accounts of cognitive functioning. The results also provide insight into the mechanisms underpinning prediction and provide opportunities for future research to refine theories of prediction.