For readers of a language written with an alphabetic script, fluent reading behavior essentially involves extracting information about letter identities and their positions to identify words and word order, and from there to construct a sentence-level representation for comprehension. Although this deliberate over-simplification ignores the well-established roles played by phonology1,2 and morphology3 in skilled reading, and also the higher-level processes involved in text comprehension4, we believe that it accurately highlights three key component processes involved in transforming visual features into meaning during reading. In the present study we investigate, for the first time, the processing interactions between these three levels. Prior research has either focused on a single level of processing, or the interactivity between two levels (letter-word or word-sentence: these interactions are respectively illustrated by path (a) and path (b) in Fig. 1). A further advantage of focusing on letter, word, and sentence processing is that it allows us to employ three very comparable tasks when measuring the processing at each of these levels. These are the alphabetic decision task5, the lexical decision task6, and the grammatical decision task7. All three tasks are speeded binary decision tasks with a clearly defined target category and well-defined criteria for constructing non-target stimuli (see examples in Fig. 1).
In order to investigate interactions between these three key component processes, in the present study participants performed alphabetic, lexical, and grammatical decision tasks, and we examined correlations between performance in each of the three tasks with the aim to evaluate the interdependence of processing across each of the three putative levels being examined (letter, word, sentence). We then used the obtained correlations to examine possible differences in the interdependencies between processing at the letter, word, and sentence levels. Thus, it is possible that word recognition is highly constrained by letter-level processing, whereas a similar contingency might not be so strong for word and sentence-level processing. It is also theoretically interesting to ask whether letter-level processing can directly constrain sentence-level processing. For example, in the OB1-reader model of sentence reading8, word length in number of letters has a direct impact on how different word identities are assigned to a specific position in a line of text.
The over-arching theoretical framework guiding this research is that of interactive-activation9–12. Within this framework, processing proceeds hierarchically, from one level to the next in a cascaded fashion. That is, information that accumulates at level 1 is transferred to level 2 before processing at level 1 is complete. Processing is also interactive, such that information at level 2 is fed-back to level 1 and constrains processing at that level. This theoretical framework is illustrated in Fig. 1.
The specific combination of architecture (hierarchical) and processing style (cascaded-interactive) adopted in this framework makes clear predictions about how processing at a given level should influence processing at the other levels. If letter-level processing is a key component of word recognition, and if the alphabetic decision task accurately reflects letter-level processing and the lexical decision task accurately reflects word-level processing, then the correlation between performance in these two tasks should be very high. The same reasoning holds for word-level and sentence-level processing, assuming that the grammatical decision task accurately reflects processing at the sentence level. Examining the cross-task correlations between alphabetic decision and lexical decision on the one hand, and lexical decision and grammatical decision on the other, will allow us to estimate the relative contributions of the different component processes to the overall task of reading.
What is the evidence in favor of the architecture described in Fig. 1 and the cascaded-interactive nature of processing between the three levels in that architecture? With respect to interactions between letter-level and word-level processing (path (a) in Fig. 1), the most relevant research here concerns the so-called “word superiority effect”13–15. This effect refers to the higher accuracy in single letter identification when the target letter is presented in a word (e.g., the letter B in TABLE) compared with a pseudoword (e.g., the letter B in PABLE). Although alternative interpretations have been proposed16, the interpretation that has best stood the test of further empirical investigation is that of cascaded-interactivity, such that upon presentation of a written word, letter-level activation immediately feeds-forward to the word-level which can then influence on-going letter-level processing via feedback. This explains why the letter B is easier to identify in the context of TABLE compared with a pseudoword context (PABLE), and very much more than in a nonword context (PFBGH). The very large effect of pseudoword superiority (pseudoword vs. nonword) is explained by pseudowords providing much greater partial activation of word representations compared with nonwords.
More recently, similar support for interactions between word-level and sentence-level processing (path (b) in Fig. 1) has been obtained in the form of a “sentence superiority effect”17. That is, identification of a single word target is better when that word is presented in the context of a correct sentence (e.g., target BOY in the sentence: “the boy runs fast”) compared with identification of the same word at the same position in an ungrammatical sequence (e.g., “runs boy fast the”). Snell and Grainger17 (see also Wen et al.18) interpreted these findings as supporting a cascaded-interactive processing account of sentence comprehension. Initial processing of multiple word identities leads to the activation of a primitive sentence level representation, possibly just a tentative ordering of the parts-of-speech associated with partially identified words, that then provides feedback to on-going word identification. The important point, with respect to the present study, is the fact that similar effects are seen across the letter-word and the word-sentence interface, and this clearly suggests a common underlying mechanism. According to the theoretical framework shown in Fig. 1, this common mechanism involves cascaded-interactive processing across adjacent levels.
In the present study we used three tasks that have been previously applied to study letter, word, and sentence-level processing. Crucially, all three tasks require a speeded binary decision as to whether or not the target stimulus belongs to a well-defined category (letters, words, sentences) relative to a background of stimuli that are designed make the discrimination difficult. The present study was motivated by the hypothesis that these three tasks could provide comparable insights into letter, word, and sentence-level processing. The alphabetic decision task involves speeded letter vs. non-letter discrimination. In the present study we opted to use the pseudo-letters provided by Vidal et al.19 as representing the best comparison relative to the pseudowords that are typically used in the lexical decision task. Direct proof that this task does reflect letter-level processing was provided by New and Grainger20, where robust effects of letter frequency were reported. The lexical decision task is quite simply the most widely used task to study single word recognition. The speeded version of the grammatical decision task is a more recent invention. Traditionally, grammaticality judgements, or well-formedness judgments, have been used by linguists in paper-and-pencil investigations of the nature of syntactic knowledge. Mirault et al.7 used a speeded binary decision version of grammaticality judgments (termed the “grammatical decision task” by Mirault & Grainger21) where they manipulated the nature of the ungrammatical sequences. Here we used the grammatical decision task as the sentence-level equivalent of lexical decisions to words and alphabetic decisions to letters. Thus, the ungrammatical sequences were chosen to be sentence-like in the same way that the pseudo-words were word-like, and the pseudo-letters were letter-like.
In the present study we set-out to examine cross-task correlations with performance in the three tasks described above with the same group of participants. This is the first time that such cross-task correlations have been examined across different levels of processing. Because the amount of shared processing is expected to be greater between two adjacent levels (letter-word; word-sentence) than between two non-adjacent levels (letter-sentence), we predicted that adjacent levels of processing (letter-word; word-sentence) should reveal stronger correlations than the correlation for non-adjacent levels of processing (letter-sentence). Participants were also tested in a speeded animal / non-animal decision task with drawings of familiar animals and inanimate objects. The aim here was to use performance on this non-reading task to partial out the contribution of common binary-decision making mechanisms in driving correlations across the three reading tasks. This specific task, compared with a simple stimulus detection task for example, has the advantage of involving a similar depth of processing as the three reading tasks while using non-linguistic stimuli. That is, the animal decision task involves making speeded binary decisions based on semantic information (i.e., “animalness”) extracted from visual information, and we considered this to be the best average approximation to the amount of processing involved in the three reading tasks, although there are obviously clear differences across these tasks.