Neural Networks for Semantic and Syntactic Prediction and Visual-Motor Statistical Learning in Adult Readers With and Without Dyslexia

Abstract

Prediction has become a key concept for understanding language comprehension, language production, and more recently reading. Recent studies suggest that predictive mechanisms in reading may be related to domain-general statistical learning (SL) abilities that support the extraction of regularities from sequential input. Both mechanisms have been discussed in relation to developmental dyslexia. Some suggest that SL is impaired in dyslexia with negative effects on the ability to make linguistic predictions. Others suggest that dyslexic readers rely to a greater extent on semantic and syntactic predictions to compensate for lower-level deficits. Here, we followed these two research questions in a single study. We therefore assessed the effects of semantic and syntactic prediction in reading and SL abilities in a population of university students with dyslexia and a group of typical readers using fMRI. The SL task was a serial reaction time (SRT) task that was performed inside and outside the scanner. The predictive reading task was performed in the scanner and used predictive versus nonpredictive semantic and syntactic contexts. Our results revealed distinct neural networks underlying semantic and syntactic predictions in reading, group differences in predictive processing in the left precentral gyrus and right anterior insula, and an association between predictive reading and SL, particularly in dyslexic readers. These findings contribute to our understanding of the interplay between SL, predictive processing, and compensation in dyslexia, providing new insights into the neural mechanisms that support reading.

Keywords: dyslexia; prediction; reading; semantics; sequence learning; syntax.

Figure 1.

Illustration of the experimental procedure and stimuli in the predictive reading task that was performed in the scanner. Abbreviations: ITI = intertrial interval; ISI = interstimulus interval; Stim = stimulation.

Figure 2.

Serial Reaction Time (SRT) task. (A) Experimental design and timing of the SRT task in the fMRI scanning adapted from Schendan et al. (2003). In the training session, participants had to press the S, R, T, and J keyboard buttons with the same four fingers of the right hand as in the subsequent fMRI experiment. (B) Experimental procedure of the SRT task in the training and scanner sessions. Abbreviations: ITI = intertrial interval; ISI = interstimulus interval; Stim = stimulation.

Figure 3.

Illustration of the regions of interest (ROIs). For the predictive reading network, anatomical ROIs were extracted from two recent articles on linguistic predictions (Carter et al., 2019; Schuster et al., 2020). The regions related to word predictability are shown in yellow, those related to semantic predictability in purple, and those related to syntactic predictability in blue. For the single word reading network, anatomical ROIs were extracted from a recent meta-analysis (Murphy et al., 2019) and are shown in pink. ROI spheres are projected respectively on left and right sides of MNI cortical surface. Abbreviations: AG = angular gyrus; aInsula = anterior insula; ATL = anterior temporal lobe; FG = fusiform gyrus; IFG oper = inferior frontal gyrus pars opercularis; IFG orb = IFG pars orbitalis; IFG tri = IFG pars triangularis; IOG = inferior occipital gyrus; MedFG = medial frontal gyrus; MidFG = middle frontal gyrus; MTG = middle temporal gyrus; SPL = superior parietal lobule; STG = superior temporal gyrus.

Figure 4.

Performance in the predictive reading task. Mean reaction times (ms) as a function of group (typical readers [TR] vs. dyslexic readers [DR]) and context (related vs. scrambled) for the (A) semantic and the (B) syntactic condition. Error bars indicate 95% confidence intervals.

Figure 5.

Behavioral results of the SRT task in the scanner. (A) Mean reaction time (ms) as a function of group (TRs vs. DRs) and condition (learned vs. random). (B) Accuracy (raw score on 720 correct responses) as a function of group and condition. (C) Mean distance of saccade landing position to stimulus center in degrees of visual angle (deg VA) as a function of target stimulus (previous, current, next), separated for TRs (circles) and DRs (diamonds) and learned (orange) and random condition (blue). Errors bars indicate 95% confidence intervals.

Figure 6.

Univariate whole-brain results of sentence reading in the localizer task (Sentence Reading − Checkerboards) for all participants. Statistical t-maps are projected respectively on left, right, and below sides of cortical surfaces (from MNI standard human cortex) using an uncorrected voxel-wise threshold of p < 0.001 and a cluster-wise threshold with FWE correction of p < 0.05. Abbreviations: SMA = supplementary motor cortex; STG = superior temporal gyrus; TTG = transverse temporal gyrus.

Figure 7.

Univariate whole-brain results in the predictive reading task for (A) semantic and the (B) syntactic prediction (related minus scrambled) for all participants. Statistical t-maps are projected respectively on left, right, and below sides of cortical surfaces (from MNI standard human cortex) using an uncorrected voxel-wise threshold of p < 0.001 and a cluster-wise threshold with FWE correction of p < 0.05. Abbreviations: AG = angular gyrus; IFG orb = IFG pars orbitalis; IFG tri = IFG pars triangularis; ITG = inferior temporal gyrus; MFG = middle frontal gyrus; SFG = superior frontal gyrus; STG = superior temporal gyrus.

Figure 8.

Univariate whole-brain results for Learned − Random sequences in the SRT task for all participants. Statistical t-maps are projected respectively on left, right, and below sides of cortical surfaces (from MNI standard human cortex) using an uncorrected voxel-wise threshold of p < 0.001 and a cluster-wise threshold with FWE correction of p < 0.05. Abbreviations: AG = angular gyrus; IFG = inferior frontal gyrus; MFG = middle frontal gyrus; MTG = middle temporal gyrus; SFG = superior frontal gyrus.

Figure 9.

Effects of semantic and syntactic predictions (PSem vs. PSyn) and their interactions for typical (TR) and dyslexic readers (DR). Only significant effects per ROI are illustrated. ROIs with significant (A) prediction effects (semantic compared to syntactic in dark blue and syntactic compared to semantic in light blue) and (B) group effects (TRs compared to DRs in dark green and DRs compared to TRs in light green) are projected on the posterior view of the brain. ROIs with significant (C) interaction between the effects of prediction and group are projected on the left view of the brain. Abbreviations: AG = angular gyrus; aInsula = anterior insula; IOG = inferior occipital gyrus.

Figure 10.

Percent signal change (PSC) in specific ROIs in the predictive reading task as predicted by the behavioral SRT measures (Saccade, SRT score, and SRT learning) and their interaction with group. For the two regression lines of each scatterplot, only the significant correlation coefficients are provided. The scatter plots show (A) ROIs that exhibit a significant interaction between group and saccades, (B) interaction between group and SRT score projected on the left view of human brain, and (C) ROI with significant interactions between group and SRT learning projected on the posterior and right views of human brain. TRs compared to DRs are represented in dark green; DRs compared to TRs are represented in light green. Abbreviations: FG = fusiform gyrus; IFG oper = inferior frontal gyrus pars opercularis; IFG orb = IFG pars orbitalis; MedFG = medial frontal gyrus.