Knowledge of Statistics or Statistical Learning? Readers Prioritize the Statistics of their Native Language Over the Learning of Local Regularities

A large body of evidence suggests that people spontaneously and implicitly learn about regularities present in the visual input. Although theorized as critical for reading, this ability has been demonstrated mostly with pseudo-fonts or highly atypical artificial words. We tested whether local statistical regularities are extracted from materials that more closely resemble one’s native language. In two experiments, Italian speakers saw a set of letter strings modelled on the Italian lexicon and guessed which of these strings were words in a fictitious language and which were foils. Unknown to participants, words could be distinguished from foils based on their average bigram frequency. Surprisingly, in both experiments, we found no evidence that participants relied on this regularity. Instead, lexical decisions were guided by minimal bigram frequency, a cue rooted in participants’ native language. We discuss the implications of these findings for accounts of statistical learning and visual word processing.


Investigating the Effect of Type in Seen vs. Unseen Strings
As reported in the main text, we followed-up on the Type * Seen interaction observed in Experiments 1-2: We wanted to determine if participants" ability to distinguish between fictitious words and foils differed for strings that had been previously seen vs. those that had not. To this end, we re-fit our main model (i.e., the one including ABF and minBF) separately for seen and unseen strings (as a consequence, we no longer included Seen as a predictor in the model). We found that, for either experiment, the effect of Type was significant only in seen strings (see Tables S1 and S2 below). Figure S1 below illustrates how the ability to distinguish between fictitious words and foils changed between the first, second, and third presentation of each item. Although the numerical pattern suggests that participants produced somewhat more "word" responses to fictitious words than foils already upon the first presentation, there is considerable variabilitya likely explanation of why our statistical analyses did not confirm this pattern.

Investigating the Effect of Procedural Learning
To test whether the use of statistical cues in our task was related to the ability to learn information procedurally, we re-fit our main GLMM including also the measure of procedural learning. This analysis, discussed in the main text of our paper and reported in details in Table S7 below, revealed a three-way interaction between P (procedural learning), Type and Seen. To further explore this interaction, we ran the follow-up models reported in Tables S8 and S9. All in all, we found that procedural learning impacted the processing of previously seen word strings. See the main text for an interpretation of these findings.
In addition, we re-ran the GLMM investigating procedural learning this time including data only from participants who reported noticing a pattern in the procedural learning task. The results, reported in Table S10, do not differ substantially from the analysis on the full sample: The model replicates the effect of minBF and the Type * Seen interaction, and the three-way interaction P * Type * Seen is still noticeable (although here it does not reach the conventional significance threshold, possibly because using a subset of data resulted in a lower statistical power).

Model Controlling for String Length
Given that word length is known to affect visual word processing (e.g., New, Ferrand, Pallier, & Brysbaert, 2006), we wanted to ensure that the findings from our main analyses were not confounded by the fact that the strings varied in length. To test this, we re-fit our main model but this time we also included Length (i.e., string length counted in letters) as an additional predictor. For either Experiment 1 and 2, these analyses replicated our main findings, thus confirming that length was not a confounding factor (see Tables S11 and S12 below).

Model Controlling for Orthographic Similarity (Coltheart´s N)
To ensure that our findings were not confounded by orthographic similarity (between our stimuli and the Italian lexicon), we refit our main statistical model controlling for Coltheart´s N. We found no effect of this measure on lexical judgments and the main pattern of results remained unchanged (i.e., the effect of minBF was once again statistically significant; the effect of ABF was not). See Tables S13 and S14 below.

Model Including ABF and Model Including ABF and maxBF
As discussed in the main text, prior to establishing that our data are best represented by the GLMM including both ABF and minBF (i.e., our main model reported in this paper provided the best fit), we ran two other models: A GLMM including ABF, and a GLMM including ABF and maxBF. The results from these models are reported in Tables S15 and S16 below.

Participants, Methods and Results
To validate our procedure and materials, we conducted two miniature pilot studies.
First, we recruited 4 further participants who took part in a version of our lexical judgment task where foil strings were never repeated, and word strings were repeated 10 times each. A descriptive analysis of data suggested that this design primed participants to follow a simple strategy: They responded "word" for any string they had seen, and responded "foil" to any string they had not previously seen. We thus decided to modify the procedure and tested further 3 participants on a task which was identical to Experiment 1 (i.e., both words and foils were repeated 3 times each). A descriptive analysis of the performance did not suggest any obvious flaws of this design, and so it was used in Experiments 1-2.