Using E-Z Reader to model the effects of higher level language processing on eye movements during reading

Abstract

Although computational models of eye-movement control during reading have been used to explain how saccadic programming, visual constraints, attention allocation, and lexical processing jointly affect eye movements during reading, these models have largely ignored the issue of how higher level, postlexical language processing affects eye movements. The present article shows how one of these models, E-Z Reader (Pollatsek, Reichle, & Rayner, 2006c), can be augmented to redress this limitation. Simulations show that with a few simple assumptions, the model can account for the fact that effects of higher level language processing are not observed on eye movements when such processing is occurring without difficulty, but can capture the patterns of eye movements that are observed when such processing is slowed or disrupted.

Figure 1

Panel A: Schematic diagram of E-Z Reader 9 (Pollatsek et al., 2006a). Panel B: Schematic diagram of E-Z Reader 10, with its assumption about attention (A) and post-lexical integration (I). The thick arrows indicate how information flows between the model’s components, the thin solid arrows indicate obligatory transitions between components, and the thin dotted arrows indicate probabilistic transitions. See the text for detailed descriptions of both versions of the model.

Figure 2

Schematic diagram showing three possible sequences of events that can occur in E-Z Reader 10. In all three panels, arrows indicate completed processes and solid circles indicate terminated processes. Panel A depicts the most common situation—when integration occurs without difficulty. The completion of the first stage of lexical processing (L₁) of word n results in the continuation of the second stage (L₂) and the initiation of a saccadic program (M₁) to move the eyes to word n+1. The completion of L₂ causes attention (A) to shift to word n+1 and initiates post-lexical integration (I) of word n. As soon as attention finishes shifting to word n+1 (represented by the dotted line labeled “a”), lexical processing (L₁) of that word begins. The non-labile stage of saccadic programming (M₂) then completes and the saccade is executed (S), moving the eyes to word n+1 (represented by the dotted line “b”). (Parafoveal processing of word n+1 thus occurs in the time interval between “a” and “b”.) Finally, the meaning of word n is integrated (indicated by “c”) before lexical processing (L₂) of word n+1 finishes, and the eyes continue to move forward. Panel B shows the situation where the integration (I) of word n fails to complete before the lexical processing (L₂) of word n+1 completes. With this “stalling out” of integration, the labile saccadic program to move the eyes to word n+2 is canceled and both the eyes and attention are drawn back to the location where comprehension difficulty first became apparent (represented by the gray arrows labeled “A” and “M₁”). Finally, Panel C shows what can happen when the early detection of a violation during the integration (I) of word n results in the termination of integrative processing, interrupting lexical processing (A, L₁, or L₂) of word n+1 and causing both the eyes and attention to move back.

Figure 3

Simulation showing how the probability of integration failure (p_F) affects the model’s overall goodness-of-fit (as measured using RMSD’s) to the Schilling et al. (1998) corpus, and the mean probability of one or more inter-word regressions occurring in the sentences. [In the simulation, t(I) = 25 ms and p_N = 1.].

Figure 4

Simulation showing how the time required to complete integration, t(I), and the probability that integration quickly fails, p_F, on word n affect four dependent measures on that word. The four panels show: (A) first-fixation durations, (B) gaze durations, and (C) total viewing times on word n, and (D) the probability of making a regression back to word n. (In the simulation, p_N = 1.) See the text for an explanation of the results.