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.

Reinventing the redundant target paradigm to distinguish serial and parallel processing of written words

“…When the task was to judge the colors of the letters, rather than the meaning of the words, accuracy was consistent with an unlimited-capacity or modestly limitedcapacity, parallel model (White et al, 2018(White et al, , 2020. When the task was to detect vowels within letter strings, or the judge the pronounceability of letter strings, accuracy was again consistent with the standard serial model (Campbell et al, 2024). These data lead to a similar conclusion as the redundant target studies reported in the present article: low-level features are processed in parallel, but linguistic features of words are processed serially.…”