Severe processing capacity limits for sub-lexical features of letter strings

Abstract

When reading, the visual system is confronted with many words simultaneously. How much of that information can a reader process at once? Previous studies demonstrated that low-level visual features of multiple words are processed in parallel, but lexical attributes are processed serially, for one word at a time. This implies that an internal bottleneck lies somewhere between early visual and lexical analysis. We used a dual-task behavioral paradigm to investigate whether this bottleneck lies at the stage of letter recognition or phonological decoding. On each trial, two letter strings were flashed briefly, one above and one below fixation, and then masked. In the letter identification experiment, participants indicated whether a vowel was present in a particular letter string. In the phonological decoding experiment, participants indicated whether the letter string was pronounceable. We compared accuracy in a focused attention condition, in which participants judged only one of the two strings, with accuracy in a divided attention condition, in which participants judged both strings independently. In both experiments, the cost of dividing attention was so large that it supported a serial model: participants were able to process only one letter string per trial. Furthermore, we found a stimulus processing trade-off that is characteristic of serial processing: When participants judged one string correctly, they were less likely to judge the other string correctly. Therefore, the bottleneck that constrains word recognition under these conditions arises at a sub-lexical level, perhaps due to a limit on the efficiency of letter recognition.

Keywords: Attention: divided attention and inattention; Attention: theoretical and computational models; Reading; Visual word recognition.

Fig. 1

Stimuli and trial sequence. A Example dual-task trial sequence. ISI = interstimulus interval. Not shown is the 1,000-ms intertrial interval after the last response on each trial. Single-task trials were identical, except the precue (green line) pointed to just one side (top or bottom), and there was only one postcue at the end of the trial that prompted the participant to judge just the precued side. In Experiment 1, the task was to report whether the letter string on the postcued side contained a vowel or not. In Experiment 2, the task was to report whether it was pronounceable or not (and all letter strings contained one vowel). B Stimulus alternatives that the participant had to distinguish between, with examples for the two tasks. (Color figure online)

Fig. 2

Attention operating characteristics (AOCs) and stimulus processing trade-offs. A–B AOCs constructed from mean accuracy (area under the ROC curve, A_g) in Experiments 1 (left) and 2 (right). We compare the dual-task accuracy (open symbol) with the predictions of three models of processing capacity limits: independent parallel, fixed-capacity parallel, and all-or-none serial. In both experiments, dual-task accuracy is significantly worse than the fixed-capacity parallel model and indistinguishable from the serial model. Error bars indicating ±1 SEM are smaller than most of the data points. C–D Stimulus processing trade-offs in Experiments 1 (left) and 2 (right). We compare accuracy in two sets of dual-task responses—those in which the response to the other side on the same trial was incorrect, and those in which the response to the other side on the same trial was correct. Thin gray lines connect data points from the same individual participant. The horizontal positions of data points are jittered to avoid total overlap, but points from the same participant have the same relative jitter. The thick horizontal lines represent the means, with ±1 SEM error bars

Fig. 3

Stimulus processing trade-offs compared with model predictions and other tasks. Mean accuracy in the dual-task condition according to whether the response to the stimulus on the other side was incorrect (x-axis) versus correct (y-axis). The dotted diagonal line is the prediction of the fixed-capacity parallel model (no trade-off). The curved solid line is the prediction of the all-or-none serial model, generated by varying single-task discriminability. The data points labeled “2020 study” are from White et al. (2020)